Timeline of AI in writing

This is a Timeline of AI in writing, covering mechanical and computational methods for generating, editing, analyzing, and detecting written text. It traces their development from 17th-century verse-making devices and early computer poetry through spell and grammar checkers, automated journalism, and content farms, to the large language models that reshape writing today, along with the legal, institutional, and cultural responses they provoke among writers, readers, publishers, and courts.

Sample questions

The following are some interesting questions that can be answered by reading this timeline:

• How did people imagine and build writing machines before computers?
  • Sort the full timeline by "Event type" and look for the group of rows with value "Backdrop".
  • You will see paper devices for generating rhymes and Latin verse, a novel's satire of machine-made literature, and the typewriters and word processors that preceded the first computer writing tools.
• What tools have researchers and hobbyists built to generate, check, or analyze text?
  • Sort the full timeline by "Event type" and look for the group of rows with value "Tool development".
  • You will see a clockwork verse machine, computer love letters, poems, and story generators, an early chatbot, the first spell checkers and style analyzers, and an AI text detector that was built but never released.
• How did AI writing assistance reach ordinary writers?
  • Sort the full timeline by "Event type" and look for the group of rows with value "Product launch" or "Product integration".
  • You will see grammar checking move from standalone programs into word processors, followed by an animated office assistant, suggested email replies, browser-based writing assistants, neural machine translation, fiction-writing apps, the launch of ChatGPT, AI text detectors, and generative features built into office software, email, and messaging.
• What technical breakthroughs made fluent machine writing possible?
  • Sort the full timeline by "Event type" and look for the group of rows with value "Model release" or "Research publication".
  • You will see word embeddings, the Transformer architecture, BERT, GPT-2, GPT-3, instruction tuning, and GPT-4, along with models released for science and business writing.
• What has research found about how AI changes writing and reading?
  • Sort the full timeline by "Event type" and look for the group of rows with value "Research publication".
  • You will see studies finding that AI detectors misjudge non-native English writers, that readers prefer AI poems they cannot tell apart from human ones, that writing is the most common work use of ChatGPT, and that words favored by language models have spread through scientific abstracts and peer reviews, along with a method for watermarking AI-generated text.
• How have organizations adopted AI to produce or handle writing?
  • Sort the full timeline by "Event type" and look for the group of rows with value "Institutional adoption" or "Company founded".
  • You will see machine translation at the European Commission, automated essay scoring on a graduate admission test, automated sports, weather, and earnings stories at news organizations, publishers licensing their archives to AI companies, and companies founded to produce text at scale with algorithms.
• How have writers, scholars, and public figures responded to AI writing?
  • Sort the full timeline by "Event type" and look for the group of rows with value "Commentary".
  • You will see predictions that computers would write most news, open letters signed by thousands of authors, novelists testing chatbots, a writing nonprofit's claim that condemning AI is classist and ableist, and a dictionary naming "AI slop" its word of the year.
• How have courts, legislatures, and institutions tried to govern AI-generated text?
  • Sort the full timeline by "Event type" and look for the group of rows with value "Legal and policy".
  • You will see copyright rulings on human authorship and fair use, sanctions against lawyers who filed fake AI citations, bans on AI text at a programming Q&A site, schools, journals, and Wikipedia, labeling rules in China and the European Union, a screenwriters' union contract, and a certification for human-written books.
• Where has AI-written text been used, passed off, or exposed in practice?
  • Sort the full timeline by "Event type" and look for the group of rows with value "Notable use".
  • You will see bots writing millions of encyclopedia articles, computer-generated papers accepted by conferences, AI fiction entering literary contests and winning a major literary prize, news outlets caught publishing AI articles under invented bylines, AI-written foraging guides with dangerous advice, and authors producing hundreds of AI-assisted novels.

Big picture

Time period	Development summary	More details
1651–1979	Prelude and early experiments	Long before computers, poets and inventors build devices that generate text by combining fixed parts: Georg Philipp Harsdörffer's rhyme wheel (1651), John Peter's Artificial Versifying (1677) for assembling Latin verse, and John Clark's clockwork Eureka (1845). George Orwell's Nineteen Eighty-Four (1949) imagines machine-made literature as a tool of political distraction. The first computer texts are experiments: Christopher Strachey's love letters (1952), Theo Lutz's Stochastische Texte (1959), whose undisclosed "electronus" poem (1960) readers take for human work, Joseph Weizenbaum's ELIZA (1966), and the story generator Tale-Spin (1976). Practical tools analyze rather than generate, from the first interactive spell checker, SPELL (1971), to Bell Labs' Writer's Workbench and the European Commission's adoption of SYSTRAN machine translation (1976).
1980–2016	Writing aids go mainstream and algorithms write at scale	Spell and grammar checkers reach personal computers (1980–1981), IBM's EPISTLE (1982) introduces full syntactic parsing, and Microsoft Word and WordPerfect build in grammar checking (1992), with Word 97 (1997) adding a parser-based checker and the much-mocked Clippy. Racter is credited as a book's author (1984), and the GMAT adopts automated essay scoring (1999). Bots write Wikipedia articles en masse, from the rambot (2002) to Lsjbot (2014), SCIgen (2005) gets fake papers accepted, content farms such as Demand Media (2006) use algorithms to decide what freelancers write, and the Associated Press automates earnings stories (2014). Grammarly (2009) moves writing help to the cloud, while word2vec (2013), Smart Reply (2015), and Google Neural Machine Translation (2016) bring neural methods to language.
2017–2022	Large language models	The Transformer, introduced in "Attention Is All You Need" (2017), leads to BERT (2018) and the GPT series. OpenAI withholds the full GPT-2 (2019) over misuse concerns, and GPT-3 (2020) writes news articles that evaluators struggle to tell from human ones; The Guardian runs an op-ed credited to it. AI Dungeon (2019), NovelAI, and Sudowrite (2021) bring generative fiction to the public, and Meta Platforms's Galactica (2022) is withdrawn within days for producing confident falsehoods. Instruction-tuned models lead to ChatGPT (November 2022), whose answers Stack Overflow bans within days.
2023–present	AI writing becomes routine, contested, and regulated	Schools, scientific journals, and later Wikipedia (2026) restrict AI text, while detectors such as GPTZero and Turnitin prove biased against non-native English writers. Generative writing is built into Grammarly, Google Workspace, and Microsoft 365, and GPT-4 (2023) follows. Fabricated AI content surfaces at CNET and Sports Illustrated (2023), the Chicago Sun-Times (2025), and in courtrooms, from Mata v. Avianca, Inc. (2023) to a Mississippi case where both sides are sanctioned (2026). Writers push back through an Authors Guild letter, screenwriters' contract protections, lawsuits including the New York Times suit (2023), and "Human Authored" certification (2026), while publishers either license their books for AI training or forbid it (2024). Courts rule that training on books can be fair use but pirating them is not, leading to a $1.5 billion settlement in Bartz v. Anthropic (2026), and that AI alone cannot be an author (2025). China (2025) and the European Union (2026) require labeling of AI-generated content. Studies find writing to be ChatGPT's most common work use and LLM-style vocabulary in a large share of scientific abstracts (2025), and "AI slop" becomes a word of the year (2025).

Full timeline

Inclusion criteria

We include:

• Pre-computer devices, methods, and works of fiction that generate text by rule or imagine machine-made writing, when later scholarship links them to computational text generation.
• The development or release of programs, models, and architectures that materially advance the ability to generate, correct, analyze, translate, or detect written text, including spell and grammar checkers, language models, AI detectors, and text watermarking.
• Launches of AI writing tools to a broad public, and the integration of writing features into widely used software such as word processors, email, office suites, and messaging apps.
• Cases of news organizations, publishers, governments, testing bodies, and companies adopting AI to produce or process writing, or licensing written work for AI training, when they have broad sectoral significance.
• Court rulings, lawsuits, legislation, regulations, union contracts, professional ethics rules, and institutional policies that concern AI-generated text, its authorship, its disclosure, or the use of written works to train AI.
• Rules on AI-generated content in general, such as copyright rulings or labeling laws, when they apply to written works as well as other media.
• Peer-reviewed or widely cited studies and investigations on how AI writing works, how widely it is used, how readers respond to it, or how reliably it can be detected.
• Essays, open letters, lectures, and public statements that mark significant turning points in how writers, educators, and institutions view AI writing, especially those that drew wide discussion or represent a major position in an ongoing debate.
• Cases of AI-written text being published, entered in competitions, passed off as human work, or exposed, when they drew wide attention or illustrate a broader pattern.

We do not include:

• Events concerning AI-generated speech, audio, images, or video that do not involve written language or set a rule applying to written works.
• AI applications that handle text only incidentally, such as search ranking, recommendation, or content moderation, unless they directly shape what is written or published.
• Routine product updates, minor feature releases, and pricing or availability changes for AI writing tools.
• Funding rounds, valuations, and corporate deals, unless they mark a significant shift in how AI writing is produced, sold, or detected.
• Individual blog posts, opinion pieces, product announcements, and course launches that address the topic without marking a significant moment in public or professional discourse.
• Isolated cases of AI misuse by individuals, such as single students or single writers, unless they drew wide attention or led to a policy response.
• Proposed bills that do not advance beyond introduction, unless they illustrate a broader legislative pattern.

Year	Writing domain	Event type	Event description
1651	Creative and literary writing	Backdrop	German Baroque poet Georg Philipp Harsdörffer publishes the Fünffacher Denckring der Teutschen Sprache ("Five-fold Thought-ring of the German Language"), a volvelle of five nested paper discs. Each disc is inscribed with word parts: 48 prefixes, 50 initial letters or diphthongs, 12 medial letters, 120 final letters or diphthongs, and 24 suffixes. Spinning the rings combines these parts into German words, with up to 97,209,600 possible combinations; printed as a dictionary, the output would fill hundreds of volumes. Harsdörffer presents the device as an aid for inventing rhyme words, in his poetry manual Poetischer Trichter and in Deliciae physico-mathematicae, a study of mathematical and linguistic games written with Daniel Schwenter. Rather than looking words up in a list, the poet generates them by hand in the moment. The Denckring would influence Athanasius Kircher's combinatorial work and Gottfried Wilhelm Leibniz's dissertation on the art of combinations. Electronic literature scholars would later regard it as a little-recognized precursor of generative and combinatory electronic poetry.
1677	Academic and educational writing	Backdrop	Mathematician John Peter publishes Artificial Versifying, or the School-boy's Recreation: A New Way to Make Latin Verses in London, a system he developed during the Great Plague of London a decade earlier. Its cover promises that anyone who knows the alphabet and can count to nine, even without understanding Latin, can produce true verse. The booklet contains six grids of letters. Choosing any six-digit number and counting through the grids reveals one hidden word per grid. Because the words follow a fixed grammatical order and fit the meter, the result is always a line of dactylic hexameter, one of 531,441 possible combinations. Aimed at schoolboys required to compose Latin verse, the booklet proves popular. By 1679 it reaches a third edition that adds pentameters and promises thousands of verses, and its procedure is imitated and reprinted in books and periodicals for some two centuries. Its fill-a-template method would be mechanized in John Clark's Eureka machine in 1845, and scholars would compare it to early computer poetry such as Theo Lutz's Stochastische Texte.
1845 (April)	Creative and literary writing	Tool development	Handbills advertise daily exhibitions at London's Egyptian Hall of the Eureka, a clockwork machine for composing Latin verse. It was built over 13 years by John Clark, a self-taught printer, poet, and inventor from Bridgwater, Somerset. At each pull of a lever, six word drums spin to random positions and wooden staves fall onto wires of different lengths, spelling out a new line of hexameter verse letter by letter, from over 26 million possible combinations. Clark presents the machine as an illustration of his theory of "kaleidoscopic evolution", in which each verse is first conceived in the mind of the machine and then mechanically displayed. Reception is skeptical. A letter to The Athenaeum dismisses it as a puzzle any schoolboy could solve with slips of paper, and Charles Babbage calls Clark as great a curiosity as his machine. Punch jokes about public schools ordering Eurekas to do their Latin exercises, a satire that educational technologist Mike Sharples would later compare to concerns about students using AI to write their essays. Restored in 2015, the machine survives in working order.
1949	Creative and literary writing	Backdrop	George Orwell publishes Nineteen Eighty-Four. In the novel, the totalitarian Ministry of Truth runs departments that mass-produce entertainment for the working class: newspapers filled with sport, crime, and astrology, sensational novelettes, films, and sentimental songs composed entirely by mechanical means on a kaleidoscope-like machine called a versificator. The heroine, Julia, works in the fiction department, where the work is done by machines that she services with a wrench. Orwell imagines machine-made text not as a technical achievement but as a tool of mass distraction and political control. Educational technologist Mike Sharples would later wonder whether Orwell's choice of the words "kaleidoscope" and "versificator" was a nod to John Clark's Latin verse machine and its theory of "kaleidoscopic evolution". In 2026, amid reports of an author producing more than 200 AI-assisted romance novels in a single year and of publications caught running AI-written stories and inventing AI reporters, James T. Keane, a senior editor at America magazine, calls the novel's fiction department a grim prophecy of the role AI has come to play in everyday life.
1952	Creative and literary writing	Tool development	British computer scientist Christopher Strachey, later a pioneer of programming language design and denotational semantics, writes a love letter generator in his spare time for the Manchester Mark 1 computer. The program prints two salutations. It then builds five sentences by randomly choosing between two fixed templates ("You are my [adjective] [noun]" or "My [adjective] [noun] [adverb] [verb] your [adjective] [noun]") and filling them from thesaurus-based word lists, before signing off as "M.U.C." (Manchester University Computer). By August 1953, its letters are appearing as anonymous notes on the noticeboard of the Manchester University computer department. Colleagues doing scientific work on the machine find it silly, but it amuses Strachey and computing pioneer Alan Turing. Written more than a decade before ELIZA, it is often described as the first work of computational or electronic literature. Scholar Noah Wardrip-Fruin would later read it as a parody of the process of writing love letters, and media historian Jacob Gaboury would frame it as a queer critique of normative expressions of love.
1959	Creative and literary writing	Tool development	Theo Lutz, a student of German philosopher Max Bense, uses the Zuse Z22 mainframe at the University of Stuttgart to generate Stochastische Texte ("Stochastic Texts"), considered the first experiment with digital literature in the German-speaking world. Using vocabulary drawn from Franz Kafka, the program joins words through weighted random operations into short logical statements such as "NOT EVERY CASTLE IS OLD." Lutz prints a selection in Bense's literary journal augenblick with an essay explaining how the texts were produced, so readers know their origin, and neither he nor Bense regards the computer as their author. In a 1962 essay, Bense would distinguish "natural poetry," grounded in a human consciousness that links words to the world, from "artificial poetry" produced by executing a rule, which has no ego behind it and can be described only through mathematical properties such as frequency and entropy. Scholars would later note that Lutz's combinatorial method resembles John Peter's 1677 Artificial Versifying.
1960 (December)	Creative and literary writing	Notable use	Theo Lutz publishes a second computer-generated poem, und kein engel ist schön ("and no angel is beautiful"), in the December issue of the German youth magazine Ja und Nein. It uses the same method as his Stochastische Texte but with Christmas vocabulary. Unlike the earlier texts, it appears without explanation among the magazine's other poems, credited only to "electronus." The next issue reveals that a computer wrote it, printing a photo of the Zuse Z22, a second poem in the form of a teletype printout, and letters from readers who had not known its origin. The letters are divided: one reader objects to giving space to such modern poets, another welcomes the avant-garde, a third admits to not understanding the poem but liking it, and only one recognizes it as computer poetry. Literary scholar Hannes Bajohr would later cite the episode as evidence of a "standard expectation" that unknown texts are written by humans who mean to communicate. He argues that large language models would erode this expectation, leaving readers in doubt about any text's origin.
Pre-1966	General-purpose writing	Backdrop	Throughout the first half of the twentieth century, professional writing is produced on manual and electric typewriters, and revising a page means cutting, pasting, and retyping it. The IBM MT/ST (1964), the first machine IBM markets as a word processor, stores typed text on magnetic tape so it can be edited before printing. By the mid-1970s, dedicated systems such as the Wang WPS (1976) are in wide corporate use. These machines process text mechanically but offer no analysis, feedback, or generation of language, leaving grammar, style, and correctness entirely to the writer. The first computer tools that work on language itself, such as ELIZA (1966) and the Bell Labs style programs of Lorinda Cherry (~1976), emerge against this background.
1966	Personal communication	Tool development	Joseph Weizenbaum develops ELIZA, an early rule-based chatbot that mimics human-like conversation using pattern-matching techniques. While simple, it marks the first instance of AI simulating natural written interaction.
1971 (February)	Editing and proofreading	Tool development	Ralph Gorin, a graduate student at Stanford University's Artificial Intelligence Laboratory working under computer scientist Les Earnest, creates SPELL for the DEC PDP-10, the first interactive spell-checking program written as a general-purpose application for English text rather than as a research prototype. Written in assembly language for speed, SPELL searches a word list for plausible corrections differing by a single letter or adjacent transposition and presents suggestions to the user interactively, making it the first spelling corrector as well as the first spell checker. Gorin makes SPELL publicly accessible following standard practice at the Stanford AI Lab, and it spreads rapidly through the nascent ARPANET, reaching universities and research institutions worldwide. By the late 1970s, spell checkers based on SPELL's approach become standard tools on mainframe computers at universities and large corporations, though their use requires expensive mainframe access unavailable to most writers.
1976	Creative and literary writing	Tool development	Yale graduate student James Meehan completes his dissertation The Metanovel: Writing Stories by Computer, describing Tale-Spin, a program that generates fables about talking animals such as Joe Bear. His advisor Alan Perlis had suggested a storytelling program but not how to build one. Meehan finds his approach after AI researcher Roger Schank arrives at Yale University in 1974 and, with psychologist Robert Abelson, develops representations of plans, goals, and scripts for understanding stories. Meehan realizes the same structures can generate stories as well as understand them. Instead of assembling stories from prewritten pieces, Tale-Spin simulates characters pursuing their goals in a small virtual world, so changing a single rule or fact can produce very different tales, which its own language generator renders as English prose. Unexpected rule interactions yield absurd "mis-spun tales," which Meehan posts on the department's corridor walls. A 1977 conference paper on the program reaches the front page of The Boston Globe. Scholar Noah Wardrip-Fruin would call it one of the most influential works in the history of digital literature, though its original code appears to be lost.
1976	Business and professional writing	Institutional adoption	The European Commission licenses SYSTRAN, a rule-based machine translation system developed by Hungarian-American linguist Peter Toma, founded in San Diego in 1968, and originally built for the United States Air Force to translate Russian scientific and technical documents during the Cold War. Work begins on an English-to-French system in 1976, followed by French-to-English the next year and a third language pair in 1979. The adoption makes SYSTRAN one of the first systems used at institutional scale to produce written output for professional use, and it establishes post-editing as a writing workflow: human translators review and correct machine-generated text rather than translating from scratch, anticipating the human-AI collaboration that large language models later make widespread. By 1982, SYSTRAN handles 50 percent of the Commission's English-to-Italian translation workload.
~1976	Editing and proofreading	Tool development	American computer scientist Lorinda Cherry at Bell Labs in Murray Hill, NJ begins developing programs to analyze English texts for weaknesses in diction and style.
Late 1970s	Editing and proofreading	Tool development	Bell Labs' Documentation Technologies Group in Piscataway, New Jersey, creates complementary programs forming the UNIX Writer's Workbench software suite.
1980	Editing and proofreading	Product launch	WordCheck, developed for Commodore personal computer systems, becomes the first commercial spell-checking software available for personal computers, marking the transition of spell checking from an exclusively mainframe-based research tool to a product accessible to ordinary writers. Where SPELL and its successors had required access to expensive institutional computing infrastructure, WordCheck brings automated spelling correction to desktop machines costing a fraction of a mainframe, prefiguring the integration of spell checking into mass-market word processors that follows in the early 1980s.
1981	Editing and proofreading	Product launch	Aspen Software of Tijeras, New Mexico, releases Grammatik, the first grammar-checking program for personal computers. Developed by computer scientist Bruce Wampler, who had found Writer's Workbench invaluable during his doctoral work but was frustrated by having to move between a personal computer and a Unix mainframe to use it, Grammatik brings automated writing analysis to individual desktop users for the first time. Initially available for the TRS-80, it soon ports to CP/M and the IBM PC. Reference Software International acquires Grammatik in 1985 and develops it into a full grammar checker capable of detecting errors beyond diction and style. WordPerfect Corporation later integrates it directly into its word processor, completing the arc from Unix research tool to mass-market writing aid.
1982	Editing and proofreading	Research publication	IBM researchers George Heidorn, Karen Jensen, Lance Miller, Roy Byrd, and Martin Chodorow publish a description of EPISTLE, an experimental text-critiquing system designed for business correspondence, in the IBM Systems Journal. Unlike earlier pattern-matching tools such as Writer's Workbench and Grammatik, EPISTLE applies full syntactic parsing to sentences using a computerized English grammar, operating at three levels: word lookup, grammar processing, and style analysis. The system achieves a 64% parse success rate on business letters tested as of December 1981. It marks what one later analysis calls the first "really visible foray" into second-generation grammar checking based on real syntactic processing. Heidorn and Jensen later move to Microsoft, where they lead development of the grammar checker built into Microsoft Word.
1984 (October)	Creative and literary writing	Notable use	Warner Books publishes The Policeman's Beard Is Half Constructed, a collection of prose and poetry credited to Racter (short for "raconteur"), a program written by William Chamberlain and Thomas Etter, with collage illustrations by artist Joan Hall. Chamberlain, a writer of short stories, television scripts, and mass-market fiction, has spent five years developing the program and says his background as a writer led him to teach a computer to write prose. The book's jacket promotes Racter as an artificial writer that produces original work without prompting from a human operator, quoting Chamberlain that once running, "It's just cooking by itself." Racter's writing has already appeared in Omni magazine and been the subject of a 1983 exhibit at the Whitney Museum in New York. Although the text is attributed to the program, the copyright is registered to Chamberlain and Hall, leaving the question of who the author of machine-generated writing is unresolved in the book itself.
1992	Editing and proofreading	Product integration	Microsoft integrates grammar checking into Microsoft Word by licensing CorrecText, a pattern-based checker developed by Houghton Mifflin, making grammar assistance a built-in feature of a mass-market word processor for the first time. WordPerfect Corporation responds by acquiring Reference Software International and bundling Grammatik directly into WordPerfect. The simultaneous move by both dominant word processors effectively ends the market for standalone grammar-checker products. The CorrecText-based checker is a first-generation pattern-matching tool; a more sophisticated full-parsing grammar engine, developed by former IBM EPISTLE researchers George Heidorn and Karen Jensen, would not reach Word users until 1997, when Microsoft replaces CorrecText with their parser-based system.
1997	Editing and proofreading	Product integration	Microsoft releases Word 97, introducing a grammar checker built on a full natural language processing system developed by George Heidorn and Karen Jensen, the same researchers who had built IBM's EPISTLE system fifteen years earlier. Unlike the pattern-matching CorrecText engine it replaces, the new checker parses sentences syntactically, enabling it to detect a far wider range of grammatical errors. One later analysis describes the upgrade as "a major breakthrough" in automated writing assistance, and notes that the full-parsing Word grammar checker remains for years the only publicly well-documented commercial-grade syntax checker at scale. The Heidorn–Jensen system would remain the core of Word's grammar checking for over two decades, until Microsoft launches a neural grammar checker in 2020.
1997	Business and professional writing	Product integration	Microsoft introduces the Office Assistant in Microsoft Office 97, an animated character that watches what users do in Word and other Office programs and offers help. The default character, a paperclip named Clippit (popularly "Clippy") drawn by illustrator Kevan Atteberry, pops up with prompts such as "It looks like you're writing a letter. Would you like help?" The concept grows out of the failed Microsoft Bob interface and the research of Stanford's Clifford Nass and Byron Reeves, who argue that people treat computers as social actors. Even inside Microsoft the project meets skepticism: Bill Gates mocks the idea of a character constantly interrupting, and developers embed the code name "TFC" in its source, privately meaning "the fucking clown". Frequent users soon find its presumptuous suggestions intrusive, and Reeves would later say the worst thing about Clippy was that it interrupted. Microsoft turns the assistant off by default a few years later, mocking it in its own marketing, and removes it entirely in Office 2007. Time would list it among the 50 worst inventions, though some former developers argue the idea was simply ahead of the technology.
1999	Academic and educational writing	Institutional adoption	The Graduate Management Admission Council adopts e-rater, an automated essay scoring system from the Educational Testing Service, for the essay section of the GMAT. This makes the business school admission test the first large-scale assessment to incorporate machine scoring of writing. Test takers write two 30-minute essays, analyzing an issue and critiquing an argument, scored on a 0 to 6 scale. Once human raters have scored enough responses to a prompt, a model trained on them replaces one of the two human raters, and a third human adjudicates if the scores differ by more than a point. The change is controversial, drawing student newspaper editorials against computer grading. In 2006 the council switches to Vantage Learning's IntelliMetric. In a 2005 evaluation, its scores agree with human raters about as closely as humans agree with each other, and it reliably flags essays that copy the prompt, though it struggles with off-topic responses and its methods remain a trade secret. Automated scoring would come to be widely accepted as a check on, rather than a replacement for, human raters.
2002 (October)	General-purpose writing	Notable use	Over the course of a week, Wikipedia user Ram-Man uses an automated script, the rambot, to create more than 30,000 articles on cities and towns in the United States. Each follows a consistent format, filled with location, geography, and population data drawn from the 2000 U.S. census and the CIA's World Factbook. With the encyclopedia at roughly 50,000 articles, the rambot expands it by over 60 percent overnight, flooding the site's lists of recent changes and slowing its single server. Critics object to the site having a page for every one-horse town across the U.S. while core topics remain uncovered or poorly written; defenders call the stubs a foundation for others to improve. Ram-Man argues that users are biased against bots, noting that nobody complained about nearly identical county articles he had added by hand. The dispute leads developers to create the "bot flag", which hides automated edits, and Ram-Man to draft Wikipedia's first bot policy, requiring separate accounts, approval, and a limited editing rate. By 2016, bots make more than 20 percent of all edits to Wikipedia.
2005	Academic and educational writing	Notable use	Three computer science graduate students at the MIT set out to expose dubious conferences that solicit papers from researchers and accept almost anything in order to collect registration fees. In a few days they write SCIgen, a program that strings together randomly selected words into bogus computer science papers, complete with realistic-looking graphs. They put their own names on a generated paper, Rooter: A Methodology for the Typical Unification of Access Points and Redundancy, and submit it to the World Multiconference on Systemics, Cybernetics and Informatics in Orlando, Florida, which promptly accepts it. Its references include fictitious papers co-authored by long-dead physicist Erwin Schrödinger. The generator outlives the prank, as some researchers later use it to pad their publication records. Computer scientists Cyril Labbé and Guillaume Cabanac would build software to detect such papers, finding 243 papers written wholly or partly by SCIgen across 19 reputable publishers. In 2010, Labbé uses SCIgen papers that cite one another to make an invented scientist, Ike Antkare, one of the most cited researchers on Google Scholar.
2006	General-purpose writing	Company founded	Richard Rosenblatt and Shawn Colo found Demand Media, acquiring eHow and building what becomes the defining example of the content farm. Proprietary algorithms identify search queries likely to attract traffic, and freelance writers are then paid around $15 to $20 per article to answer them; at its peak the company adds about 5,000 articles and videos to eHow a day. Algorithms decide what to write while humans produce the text, an early form of algorithmically directed mass writing. Similar operations include Associated Content, Suite101, and Examiner.com. Critics accuse content farms of degrading the quality of information online and crowding substantive writing out of search results. In February 2011, Google's Panda algorithm update, informally called the "farmer update", targets thin, low-quality content and cuts traffic to Demand Media sites by up to 40 percent, ending the first era of algorithmically driven mass writing. The episode anticipates later debates over AI-generated content, pitting scale against quality, with search engines acting as enforcers.
2009 (July)	Editing and proofreading	Product launch	Ukrainian software developers Alex Shevchenko, Max Lytvyn, and Dmytro Lider launch Grammarly, initially as a grammar and plagiarism checker aimed at university students, building on an earlier plagiarism-detection product called MyDropbox that the founders had developed and sold to Blackboard Inc. Where grammar checking had previously required either a standalone desktop application or a built-in word processor feature, Grammarly operates as a cloud-based service accessible through browser extensions, making AI writing assistance available across any text field on any website. The company later evolves into a broader AI-powered writing platform offering style, tone, and clarity suggestions, reaching a reported valuation of $13 billion by 2021. Its real-time, cross-platform, cloud-based architecture becomes the model for a subsequent generation of AI writing assistants.
2010	Journalism	Company founded	Northwestern University computer science professors Kristian Hammond and Larry Birnbaum co-found Narrative Science in Evanston, Illinois, commercializing a student research project called StatsMonkey that automatically generated college baseball game stories from box scores and play-by-play data. The company develops Quill, a natural language generation platform that extracts the most significant patterns from structured datasets and converts them into readable narrative prose. Quill's early clients include financial institutions and media organizations seeking to cover data-rich domains, such as earnings reports, sports results, and market summaries, at a scale impossible with human writers alone. The company is later acquired by Salesforce in 2021. Narrative Science and its contemporary Automated Insights establish natural language generation as a commercially viable writing technology in the years before large language models render the underlying rule-based approach largely obsolete.
2012 (April 24)	Journalism	Commentary	Technology journalist Steven Levy profiles Narrative Science in Wired. The Chicago startup now produces a story roughly every 30 seconds, from corporate earnings previews published by Forbes to nearly 400,000 Little League game recaps the previous year, built from data that parents enter in a phone app. Trained journalists working as "meta-writers" build the templates and vocabulary that the software fills with angles it finds in the data. The output is also shaped for clients: at the request of the Big Ten conference, stories praise its players even in defeat, and Little League recaps skip children's errors. Cofounder Kristian Hammond, who came under the influence of AI researcher Roger Schank as a Yale undergraduate, predicts that within 15 years "more than 90 percent" of news will be written by computers, and that a computer will win a Pulitzer Prize within five years. He argues the technology will expand coverage to events no journalist covers rather than replace reporters, telling a critic that nobody has lost a job because of it.
2013 (January)	General-purpose writing	Research publication	Google researchers Tomas Mikolov, Kai Chen, Greg Corrado, and Jeffrey Dean publish Efficient Estimation of Word Representations in Vector Space, introducing two simple architectures, continuous bag-of-words and skip-gram, for learning word vectors from very large text collections. Most language systems of the time treat words as unrelated symbols in a vocabulary, with no notion of similarity between them. By dropping the costly hidden layer of earlier neural language models, the new architectures can learn high-quality vectors from 1.6 billion words in less than a day. The resulting vectors capture relationships through simple arithmetic: subtracting the vector for "France" from "Paris" and adding "Italy" yields a vector closest to "Rome", and similar operations recover capitals, comparatives, and plurals, though imperfectly. The team trains on a Google News corpus of about six billion words, then releases its code, distributed as word2vec, along with vectors for more than 1.4 million named entities. The authors expect word vectors to become an important building block for future natural language applications, including machine translation.
2014 (July)	Journalism	Institutional adoption	The Associated Press begins using Wordsmith, a natural language generation platform developed by Automated Insights of Durham, North Carolina, to automatically produce corporate earnings reports from structured financial data supplied by Zacks Investment Research. Where AP reporters previously produced around 300 such stories per quarter, Wordsmith generates over 3,000 in the same period, a tenfold increase, freeing journalists to focus on investigative and enterprise reporting. The project demonstrates that natural language generation can produce publishable prose at scale for data-rich domains, and prompts widespread discussion in newsrooms about the future division of labor between human writers and automated systems. AP later expands the program to cover minor league baseball and announces plans to extend coverage to Canadian and European company earnings.
2014 (July 13)	General-purpose writing	Notable use	The Wall Street Journal profiles Sverker Johansson, a 53-year-old Swede from Falun who is credited with 2.7 million Wikipedia articles after seven years of volunteering, about 8.5 percent of the entire collection by Wikimedia's count and far more than any other user. Most of his articles are produced by an automated program, Lsjbot, which lets him turn out as many as 10,000 articles in a day. Some see Johansson as a hero, while purists object to articles mass-produced by software. The bot's output raises the Swedish Wikipedia to the second-largest edition after English. It continues, on a far larger scale, the practice of bot-generated articles begun by the rambot on the English Wikipedia in 2002.
2015 (November 5)	Personal communication	Product integration	Google launches Smart Reply in its Inbox by Gmail app for Android and iOS, offering three short suggested responses to emails it judges can be answered with a tap. In 2009, an April Fools' Day joke had announced a fake "Gmail Autopilot" that would read and answer email automatically. Google takes the idea seriously in early 2015, after rapid progress in machine learning for language and as more than half of Americans' email comes to be read on phones, where replying is awkward. A neural network trained on Gmail messages, which humans may not read for privacy reasons, groups messages that say the same thing in different ways and suggests the most typical replies, tuned to offer distinct options. Early versions suggest "I love you" far too often, a habit traced to affectionate internal testers and then filtered out. Testing shows that users prefer to write intimate messages themselves, so suggestions are capped at five or six words, since longer replies feel as if they need the user's own tone. Writer Nicola Twilley, testing it for The New Yorker, finds most suggestions usable but is unsettled by how much of her correspondence could be handled with stock phrases.
2016 (March)	Creative and literary writing	Notable use	A short novel produced by a team led by professor Hitoshi Matsubara of Future University Hakodate with an AI program, titled The Day a Computer Writes a Novel, passes the first round of screening for Japan's third Nikkei Hoshi Shinichi Literary Award. The award, named after science fiction writer Shinichi Hoshi, accepts entries from non-human applicants, and this is the first year it receives AI-assisted submissions, 11 in all. Judges are not told which entries involve AI. The humans decide the plot and characters and supply words and sentences, which the program assembles into the text, so commentators such as Slate's Jacob Brogan argue that the result shows how heavily humans were involved. The novel is eliminated after the first round, with the team acknowledging weaknesses such as character descriptions. Literary scholar Hannes Bajohr would later cite the episode among early cases of machine-assisted fiction entering literary competition.
2016 (September 27)	Business and professional writing	Product integration	Google Brain researchers Quoc V. Le and Mike Schuster announce the Google Neural Machine Translation system (GNMT), and Google Translate switches all of its Chinese-to-English translation to it, about 18 million translations a day. Since its launch a decade earlier, Google Translate has used phrase-based machine translation, which breaks sentences into words and phrases translated largely independently. GNMT instead treats the whole sentence as a unit: an encoder turns the source words into vectors, and a decoder generates the translation one word at a time, using an attention mechanism to focus on the most relevant source words. In side-by-side ratings by bilingual human evaluators on sentences from Wikipedia and news sites, it reduces translation errors by 55 to 85 percent on several major language pairs. Deployment at this scale relies on Google's TensorFlow toolkit and its Tensor Processing Units, and Google plans to extend GNMT to many more of Translate's more than 10,000 language pairs. The researchers acknowledge that it still drops words, mistranslates names and rare terms, and translates sentences in isolation from their context, errors a human translator would not make.
2017 (June 12)	General-purpose writing	Research publication	Eight researchers at Google Brain and Google Research, Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Łukasz Kaiser, and Illia Polosukhin, publish "Attention Is All You Need", introducing the Transformer architecture for sequence-to-sequence tasks. Where dominant language models at the time relied on recurrent or convolutional neural networks that processed text sequentially, the Transformer dispenses with recurrence entirely in favor of a self-attention mechanism that relates all positions in a sequence simultaneously, making training substantially more parallelizable. Originally developed for machine translation, the architecture proves broadly applicable to tasks involving sequential text. It becomes the foundation of BERT (2018), GPT-2 (2019), and the subsequent lineage of large language models that reshape AI-assisted writing. "Attention Is All You Need" becomes one of the most cited papers in machine learning.
2017 (August)	Business and professional writing	Product launch	DeepL GmbH, a startup from the developers of the online dictionary Linguee, launches DeepL Translator, a neural machine translation service. It later supports translation between nine European languages, English, German, French, Spanish, Portuguese, Dutch, Italian, Polish, and Russian, in 72 language combinations. Many translators find its output sounds more natural than that of other major translation services, and DeepL quickly generates demand among professional translators. When the computer-assisted translation tool memoQ adds a DeepL plugin in July 2018, a memoQ solution architect calls it one of the most requested integrations ever, crediting DeepL's understanding of how translation tools work, such as its support for preserving formatting tags in translated segments.
2018 (July 24)	Personal communication	Product integration	Google announces the rollout of Smart Compose to G Suite users. Initially an experimental Gmail feature, Smart Compose autocompletes sentences as users write emails, especially common phrases, greetings, and addresses. It learns users' writing habits over time, including names and jargon, to offer personalized suggestions, which users accept by pressing the tab key. The move marks a shift from Google's usual practice of delaying feature releases after announcement. The feature aims to speed up email writing without replacing full message composition, and becomes a key part of Gmail's AI-driven writing tools.
2018 (October 11)	General-purpose writing	Research publication	Researchers Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova publish BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. The paper introduces BERT (Bidirectional Encoder Representations from Transformers), a language representation model built on the Transformer architecture. Unlike recent language representation models, BERT is pre-trained on unlabeled text by conditioning jointly on both the left and right context of each word, in every layer. The pre-trained model can then be fine-tuned with a single additional output layer to create state-of-the-art systems for tasks such as question answering and language inference, without substantial task-specific changes to its architecture. BERT sets new records on eleven natural language processing tasks, including raising the GLUE benchmark score to 80.5%, an absolute improvement of 7.7 points. It is designed for understanding text rather than generating it. Its approach of pre-training followed by task-specific fine-tuning becomes the paradigm that GPT-3 would later seek to move beyond with few-shot prompting.
2019 (February 18)	General-purpose writing	Model release	OpenAI releases GPT-2, a large language model capable of generating highly coherent and human-like text. While the research is widely praised for its technical achievements, OpenAI's decision not to release the model's largest version sparks controversy. The company cites the potential for misuse in disinformation and abuse at scale, while critics accuse it of generating hype.
2019 (December 5)	Creative and literary writing	Product launch	Nick Walton, who began the project at a hackathon, releases AI Dungeon 2, a text adventure game built on the full 1.5-billion-parameter version of OpenAI's GPT-2, fine-tuned on about 30 megabytes of text scraped from an online choose-your-own-adventure forum. Unlike earlier text adventures with scripted responses, it lets players type any action or dialogue, and the model writes what happens next. The game spreads rapidly: about 100,000 people play in its first week, and within six weeks it has more than a million players and six million stories. Walton's startup Latitude later moves the game to OpenAI's more powerful GPT-3, which often produces sexually explicit content without being asked. Walton says the company did not expect this and must learn to constrain a system that starts out able to write anything. In April 2021, after OpenAI raises concerns about players generating sexual content involving minors, Latitude deploys a filter that also flags innocent phrases such as "eight-year-old laptop" and subjects private stories to human review. Users revolt, the game's Google Play rating falls from 4.8 to 2.6, and thousands of players turn to the new rival NovelAI.
2020 (May 28)	General-purpose writing	Research publication	A team of 31 OpenAI researchers led by Tom B. Brown, and including Alec Radford, Ilya Sutskever, and Dario Amodei, publishes Language Models are Few-Shot Learners, introducing GPT-3. It is an autoregressive language model with 175 billion parameters, ten times more than any previous non-sparse language model. The prevailing approach of pre-training a model and then fine-tuning it requires task-specific datasets of thousands of examples. GPT-3 instead performs new language tasks from a few examples or simple instructions given purely as text, with no further training. It performs strongly on translation, question answering, and cloze tasks, although the authors identify datasets where few-shot learning still struggles and methodological problems that come from training on large web corpora. For writing, the paper's most consequential finding is that human evaluators have difficulty distinguishing news articles generated by GPT-3 from articles written by humans, which the authors discuss alongside the model's broader societal impacts. The paper is revised three times through July 2020.
2020 (September 8)	Journalism	Notable use	The Guardian publishes an opinion piece credited to GPT-3, OpenAI's language model, under the headline A robot wrote this entire article. Are you scared yet, human? The essay argues that humans have nothing to fear from AI, opening with the claim that its author is a robot that taught itself by reading the internet. An editor's note explains the process. The Guardian wrote a prompt asking for a short, simple op-ed of about 500 words on why humans should not fear AI, along with an opening paragraph citing Stephen Hawking's warning about AI. Computer science undergraduate Liam Porr of UC Berkeley then fed these to the model. GPT-3 produced eight essays, and editors cut, combined, and rearranged the best parts of each, saying the job took less time than editing many human op-eds. The piece helps launch a brief journalistic genre of articles in which a language model is presented as the author. Literary scholar Hannes Bajohr would point out that GPT-3's authorship here is a fiction, since humans wrote the prompts and hand-selected the output.
2020	Editing and proofreading	Product integration	Microsoft launches a neural grammar checker for Microsoft Editor, its AI writing assistant, succeeding the parser-based system that had powered Word's grammar checking since 1997. Built on deep neural networks with a "fluency boost" learning and inference method developed at Microsoft Research, it achieves state-of-the-art results on two standard grammatical error correction benchmarks. In 2022 Microsoft releases an optimized version across Word for Windows, Word Online, Outlook Online, and the Editor browser extension. By 2023, a decoding technique that checks a sentence in parallel rather than word by word, exploiting the fact that a corrected sentence closely resembles the original, makes the server model about three times faster and cuts its cost by two-thirds without loss of quality. A compressed model more than five times smaller also runs on users' own devices, working offline and without sending text to a server. Microsoft uses GPT-3.5 to label large amounts of unannotated text as additional training data and to identify and relabel low-quality training examples.
2021 (June 16)	Creative and literary writing	Product launch	Delaware-based studio Anlatan launches NovelAI, a subscription-based AI writing platform built specifically for fiction, founded by developers from the AI Dungeon community who sought a privacy-focused alternative after controversial policy changes and data breaches at other AI fiction platforms. Unlike general-purpose language models, NovelAI uses versions of GPT-Neo and GPT-J fine-tuned on literary text rather than general web data, aiming to maintain narrative voice, genre conventions, and character consistency. The platform reaches 40,000 subscribers within months of launch, presenting uncensored creative freedom, user privacy, and encrypted story storage as its defining values. Its rapid adoption shows demand for AI writing assistance among fiction writers, particularly in genre communities such as fantasy, science fiction, horror, and romance.
2021 (November)	Creative and literary writing	Product launch	Sudowrite, an AI fiction-writing tool founded in August 2020 by science fiction writers Amit Gupta, a Y Combinator alumnus, and James Yu, a former engineer at Google and Zynga, leaves closed beta and raises $3 million in seed funding from backers including the founders of Medium, Twitter, WordPress, and Gumroad. Built on GPT-3, it is designed around the demands of long-form fiction: keeping characters' voices consistent, managing plot, generating description, and overcoming writer's block. Authors including Cory Doctorow, Mary Robinette Kowal, and Ted Chiang are among its early users. Unlike the uncensored, community-oriented NovelAI, Sudowrite presents itself as a professional collaborator that augments rather than replaces the author's voice, and later develops its own fiction model, Muse, trained on data it says was licensed with authors' consent. Together with NovelAI, it establishes fiction-specific AI writing tools as a distinct product category.
2022 (March 4)	General-purpose writing	Research publication	OpenAI researchers Long Ouyang, Jeff Wu, Xu Jiang, and colleagues publish "Training language models to follow instructions with human feedback", introducing InstructGPT, a family of models fine-tuned from GPT-3 using reinforcement learning from human feedback (RLHF) to follow user instructions more reliably than the base model. Where GPT-3 generates text by predicting likely continuations of a prompt, InstructGPT is optimized to produce outputs that match human intent, making it better suited to task-oriented writing such as drafting, summarization, and structured content generation. In human evaluations, outputs from the 1.3-billion-parameter InstructGPT model are preferred to those from the 175-billion-parameter GPT-3, despite its having 100 times fewer parameters. InstructGPT also gives truthful answers roughly twice as often as GPT-3 and produces 25 percent fewer toxic outputs. The approach directly prefigures ChatGPT, which applies the same technique to a conversational format eight months later.
2022 (March)	Journalism	Institutional adoption	The local Argentine newspaper Diario Huarpe, based in San Juan, starts using automation from United Robots to publish around 250 football articles and 3,000 weather reports a month. The system turns structured data into articles in natural regional Spanish through natural language generation, helping the paper's small team expand its coverage of sports and local weather despite limited staff. After some initial resistance, journalists use the automation to free up time for deeper stories. The main challenge is access to structured data, especially for local football leagues, while weather data is readily available. The automation increases the paper's efficiency, search traffic, and audience reach.
2022 (November 17)	Academic and educational writing	Model release	Meta's AI research division releases Galactica, a 120-billion-parameter language model trained on a large corpus of scientific papers, reference material, and knowledge bases, which it presents as able to summarize academic literature, solve math problems, and write encyclopedia-style articles and scientific papers from short prompts. Within days, researchers find that it produces fluent, authoritative-sounding text that is often scientifically inaccurate, including invented citations and false claims, and Meta takes the public demo offline after three days.
2022 (November 30)	General-purpose writing	Product launch	OpenAI releases ChatGPT as a free research preview, a chatbot built on a model fine-tuned from its GPT-3.5 series using reinforcement learning from human feedback, the technique developed for InstructGPT. Its conversational format lets it answer follow-up questions, admit mistakes, challenge incorrect premises, and decline inappropriate requests. OpenAI acknowledges that it sometimes writes plausible-sounding but incorrect or nonsensical answers, and that it is sensitive to how a question is phrased. By making fluent text generation available to anyone with a web browser, ChatGPT puts drafting, editing, summarizing, and brainstorming assistance in the hands of a mass audience, and quickly raises questions about originality, authorship, and reliance on AI-generated text.
2022 (December 5)	Business and professional writing	Legal and policy	Five days after ChatGPT's public launch, the volunteer moderators of the programming question-and-answer site Stack Overflow temporarily ban answers generated with it, one of the first institutional responses to the chatbot. They say the rate of correct answers from ChatGPT is quite low, yet its answers look authentic and are effortless to produce, so people without the expertise or willingness to check them are posting thousands. Because spotting a bad answer often requires a detailed read by someone with subject knowledge, the flood has swamped the site's volunteer quality-control system, making such answers substantially harmful to the site and to users seeking solutions. Users who break the policy face sanctions, even if a post would otherwise be acceptable, and a permanent policy is to be worked out with staff and the community. OpenAI itself acknowledges that ChatGPT sometimes writes "plausible-sounding but incorrect or nonsensical answers", a problem it says is hard to fix.
2023 (January 3)	Academic and educational writing	Product launch	Edward Tian, a computer science undergraduate at Princeton University writing his senior thesis on AI detection, releases a beta of GPTZero on Twitter during his winter break, describing it as a tool to detect whether text was written by ChatGPT or a human. It crashes within its first week under 30,000 users and draws coverage in the New York Times, Washington Post, and Wall Street Journal. GPTZero measures perplexity, the randomness of word choice within a sentence, and burstiness, how much that randomness varies across sentences, on the theory that human prose varies more than AI prose. The launch sets off a wave of detection tools, including OpenAI's own AI text classifier later that month, which OpenAI later withdraws for low accuracy, and Turnitin's detector in April. Critics soon document false positives that fall disproportionately on non-native English speakers.
2023 (January 17)	Journalism	Notable use	Technology news site CNET begins appending lengthy correction notices to articles it has been generating with artificial intelligence since November 2022, after the site Futurism flags basic errors in them. One AI-written explainer claims that a $10,000 deposit at 3 percent interest earns $10,300 in its first year, rather than $300. The dozens of personal finance pieces, some also published by sister site Bankrate, carry a generic staff byline that discloses the use of automation only when readers click on it; marketing executive Gael Breton exposes the practice on Twitter the previous week. Editor-in-chief Connie Guglielmo describes the program as an experiment meant to assist reporters rather than replace them, and both outlets pledge to review every AI-assisted piece. Computer scientist Hany Farid of UC Berkeley calls it "primarily an editorial failure", suggesting the authoritative AI voice may have lowered editors' guard. Unlike the Associated Press's earlier template-based earnings stories, CNET's system writes full feature articles, and the episode sharpens journalists' fears that AI could accelerate the long decline in newsroom employment.
2023 (January)	Academic and educational writing	Legal and policy	New York City Public Schools, the largest school district in the United States, bans ChatGPT on its networks and devices over concerns about cheating. A department spokesperson says the tool does not help build critical-thinking and problem-solving skills. Days earlier, a Wall Street Journal reporter had shown that ChatGPT could produce an 800-word Advanced Placement English essay with few errors in seconds. Other large districts follow, including Fairfax County Public Schools, Seattle Public Schools, and the Los Angeles Unified School District. In response to the ban, OpenAI says it is working on ways to help detect text generated by ChatGPT. On May 18, schools chancellor David C. Banks would reverse the ban in an op-ed for the education news site Chalkbeat. He would write that the initial caution was justified, but that the system is now exploring the technology's power and risks, giving educators resources and real-life examples of AI use in administrative tasks, communication, and teaching.
2023 (January 26)	Academic and educational writing	Legal and policy	Scientific publishers move to restrict ChatGPT after it is credited as a co-author on a handful of papers and shown to produce fake abstracts convincing enough to fool reviewers. Holden Thorp, editor-in-chief of Science, announces that the journal will not allow ChatGPT to be listed as an author or its text to be used in papers. He reasons that authors must sign a declaration of accountability, which a program cannot do, and warns that reliance on AI could introduce errors and erode scrutiny. Two days earlier, Springer Nature, publisher of nearly 3,000 journals, barred listing such tools as authors but allowed their use if disclosed. Nature editor-in-chief Magdalena Skipper suggests they could help non-native English speakers write more fluently. Elsevier, publisher of about 2,800 journals, permits AI for improving readability and language but not for interpreting data or drawing conclusions. eLife editor Michael Eisen calls adoption inevitable and stresses disclosure, while Oxford technology regulation professor Sandra Wachter warns of junk science and predicts that education, art, and journalism will need similar rules.
2023 (c.February 14)	Business and professional writing	Model release	Writer, a company that sells generative AI tools to businesses, launches Palmyra, a family of three large language models built for business writing and marketing: Palmyra Small (128 million parameters), Palmyra Base (5 billion), and Palmyra Large (20 billion). Writer offers both the models and the application built on them, and emphasizes security, privacy, and control over brand voice, arguing that general-purpose language models lack the context awareness and compliance standards that enterprises need.
2023 (February)	Business and professional writing	Commentary	Ken Scudder, a communications trainer and consultant and a regular contributor to Strategies & Tactics, the trade publication of the Public Relations Society of America, examines the rise of AI-generated text and what it means for professional writers. Having ChatGPT generate an article on AI and writing, he finds its output competent but generic: fluent, yet lacking creativity, personal voice, and lived insight. Scudder argues that writers who embrace AI as a productivity tool, rather than fearing it as a replacement, will stay relevant, and that distinctively human qualities such as humor, specificity, and emotional resonance remain beyond AI's reach. Published for public relations and communications professionals, the piece is an early attempt within the industry to define the human writer's advantage in an AI-assisted workplace.
2023 (February)	Creative and literary writing	Notable use	Science fiction and fantasy magazine Clarkesworld, which has published writers including Jeff VanderMeer, Yoon Ha Lee, and Catherynne Valente and is one of the few paying markets open to short story submissions from new writers, closes to submissions after a flood of AI-generated stories overwhelms its editors. Founding editor Neil Clarke reports that the magazine normally rejects around ten plagiarized submissions a month. Since ChatGPT brought AI language models into the mainstream, however, it has rejected and banned 100 submitters in January and more than 500 so far in February. Clarke blames people from outside the genre community, drawn in by influencers selling ChatGPT as an easy-money side hustle. He rules out the obvious fixes, stating that "Detectors are unreliable," that pay-to-submit would exclude too many legitimate authors, and that print submissions are not viable. He warns that new barriers would fall hardest on new and international writers, and notes that other editors report the same problem, especially at well-paying markets that never close to submissions. The magazine promises to reopen but sets no date.
2023 (March 7)	Academic and educational writing	Commentary	An editorial in Nature Biomedical Engineering, a journal of the Nature family, argues that the moment of truly useful large language models has arrived. It states that, barring advanced watermarking, it is no longer possible to accurately distinguish text written by a human from text generated by a large neural network, a threshold the editors say most machine learning experts and linguists would not have thought reachable a few years earlier. The editors expect the consequences to be far-reaching across research, clinical documentation, and scientific publishing. The piece is an early instance of a major journal's editors stating that the line between human and machine writing has effectively collapsed in practice.
2023 (March 9)	Editing and proofreading	Product integration	Grammarly, whose writing assistant is used by some 30 million people and 50,000 business teams a day, announces GrammarlyGO, a generative AI feature marking its move from correcting text to producing it. Head of product Rahul Roy-Chowdhury describes the shift as "going past revision into conception and composition". Within the apps and websites where people already write, users can compose drafts from short prompts, generate email replies from one-click prompts such as "I'm interested", rewrite passages for tone, clarity, or length, and generate outlines and ideas, with output adjusted to a user's chosen tone and professional role. GrammarlyGO is to enter beta in April across Grammarly's free, premium, business, education, and developer offerings. It is switched on by default for individual users, while business and education administrators can choose whether to enable it for their organizations. Grammarly presents the launch as building on nearly 14 years of its own AI work, and emphasizes security and privacy.
2023 (March 14)	Business and professional writing	Product integration	Google announces generative AI features for Google Workspace, starting with writing tools in Google Docs and Gmail for a group of trusted testers. Users can type a topic and have the AI draft a document or email, then rewrite it, summarize it, or adjust its tone. Google says similar features will later reach Slides, Sheets, Meet, and Chat, for tasks such as generating images, analyzing data, and taking notes. It presents the tools as collaborators that leave users in control, and emphasizes privacy and its AI principles. The announcement comes two days before Microsoft unveils Microsoft 365 Copilot, as the two companies race to build generative AI into their office software.
2023 (March 15)	General-purpose writing	Model release	OpenAI publishes the technical report for GPT-4, a large multimodal model that accepts image and text inputs and produces text outputs. Like its predecessors, GPT-4 is a Transformer-based model pre-trained to predict the next token in a document, followed by a post-training alignment process that improves its factuality and its adherence to desired behavior. The report credits the work to OpenAI as a whole, listing some 280 contributors. While acknowledging that GPT-4 is less capable than humans in many real-world scenarios, OpenAI reports human-level performance on various professional and academic benchmarks, including a simulated bar exam on which it scores around the top 10 percent of test takers. The report also describes infrastructure that let the company accurately predict aspects of GPT-4's performance from models trained with no more than a thousandth of its compute.
2023 (March 16)	Business and professional writing	Product integration	Microsoft announces Microsoft 365 Copilot, which combines large language models, including GPT-4, with users' own documents, emails, calendars, and meetings, accessed through the Microsoft Graph. Built into Word, Excel, PowerPoint, Outlook, and Teams, it lets users draft documents from a short prompt, summarize long email threads, turn written material into presentations, and analyze data in natural language. The announcement also introduces Business Chat, which answers questions by drawing on information across a user's apps. Microsoft says Copilot inherits existing security and privacy permissions, so it can only use data a user is already allowed to see, and presents it as a tool that keeps people in control of the final work.
2023 (March 16)	General-purpose writing	Legal and policy	The U.S. Copyright Office, led by Register of Copyrights Shira Perlmutter, publishes a statement of policy in the Federal Register on registering works that contain material generated by artificial intelligence. The Office notes it is already receiving such applications. One example is Zarya of the Dawn, a comic book it found copyrightable as a whole in February 2023 while ruling that its individual Midjourney images could not be protected. Citing precedents such as Burrow-Giles Lithographic Co. v. Sarony, the Office reaffirms that copyright protects only the product of human creativity. When a system produces complex written, visual, or musical work from a prompt alone, the machine rather than the user determines its expressive elements: asked for a poem about copyright law in the style of William Shakespeare, a text generator still decides the rhymes, words, and structure. Human selection, arrangement, or sufficient modification of AI output can still be protected. Applicants must disclose and exclude AI-generated content that is more than de minimis, should not list an AI tool as an author, and risk losing their registration if they fail to correct the record. The Office plans a notice of inquiry later in the year on AI training and outputs.
2023 (March 27)	Academic and educational writing	Commentary	Three English professors at North Carolina State University offer diverging views on what generative AI means for writing and education. Chris Anson, director of the university's Campus Writing and Speaking Program and a former chair of the Conference on College Composition and Communication, warns that delegating complex writing tasks to AI bypasses the thinking that writing develops, potentially stunting writers' growth. Huiling Ding, a professor of technical communication who studies responsible AI, highlights concerns about authorship, plagiarism, and the displacement of professional writers and artists. Paul Fyfe, director of the university's graduate certificate in digital humanities, takes a more exploratory stance, encouraging structured classroom experiments with AI to build critical AI literacy rather than avoiding the tools. Their positions of caution, ethical critique, and critical engagement reflect the range of responses emerging in writing instruction as ChatGPT spreads.
2023 (April 4)	Academic and educational writing	Product integration	Plagiarism-detection company Turnitin adds AI writing detection to its widely used academic integrity platform, giving instructors an indicator of how much of a submission is likely AI-generated. Turnitin says it prioritized a false positive rate below 1 percent, but problems surface quickly. A Washington Post test that spring finds original student work wrongly flagged, and instructors report international students being flagged disproportionately. A Stanford-led study would find that seven widely used detectors misclassify writing by non-native English speakers as AI-generated at an average rate of 61.3 percent. A University of Reading blind test would find that 94 percent of AI-written exam answers go undetected. Turnitin repeatedly revises the tool: it raises the minimum length to 300 words, acknowledges more false positives for scores between 1 and 20 percent, and later hides low scores behind an asterisk. Its own guidance warns that results should not be the sole basis for action against a student. Vanderbilt University disables the detector, and Australian Catholic University, which records nearly 6,000 alleged misconduct cases in 2024 (about 90 percent of them AI-related), would later abandon it.
2023 (May)	Creative and literary writing	Notable use	Canadian novelist and journalist Stephen Marche publishes Death of an Author, a murder mystery novella released by Pushkin Industries as an ebook and audiobook under the pseudonym Aidan Marchine, a name chosen to signal the book's combined human and machine authorship. Marche plots the story himself, then has ChatGPT compose it paragraph by paragraph, uses Sudowrite to rework the prose and tone, and turns to the Canadian language model Cohere for figurative language, which he credits with the book's best sentences. He estimates the text is 95 percent machine-generated. The novella follows earlier experiments, including a 2021 horror story he described as 17.1 percent auto-generated. Reviewing it in The New York Times, critic Dwight Garner calls it arguably the first halfway readable AI novel and an early glimpse of what is coming for readers, crediting a literate writer for pushing the software beyond boilerplate some of the time. He notes that others have already flooded Amazon with AI-generated books, but without Marche's openness about how they were made.
2023 (June 22)	Business and professional writing	Legal and policy	Judge P. Kevin Castel of the U.S. District Court for the Southern District of New York sanctions two lawyers in Mata v. Avianca, Inc., a personal injury suit by Roberto Mata, who says a metal serving cart struck him on a 2019 Avianca flight. Attorneys Steven Schwartz and Peter LoDuca had filed a brief citing six court opinions that do not exist, complete with fake quotes and citations generated by ChatGPT, which Schwartz had used for legal research. When Avianca's lawyers could not locate the cases, the attorneys stood by them even after the court ordered copies, admitting the truth only once threatened with sanctions; Schwartz later tells the court he failed miserably to verify them. Castel finds bad faith, fines the lawyers and their firm $5,000, and orders them to send his opinion to their client and to each real judge falsely named as the author of a fabricated opinion. He declines to require an apology, reasoning that a compelled one would not be sincere, and dismisses Mata's suit as filed too late. Legal ethics experts suggest the widespread news coverage may be the lawyers' harshest punishment.
2023 (July 13)	Journalism	Institutional adoption	OpenAI and the Associated Press announce a deal under which OpenAI licenses part of AP's text archive, with stories going back to 1985, while AP gains access to OpenAI's technology and product expertise. Financial terms are not disclosed, and the two organizations say they will explore uses of generative AI in news products. AP does not yet use generative AI in its stories, although it has used other forms of AI for nearly a decade to automate corporate earnings reports and some sports recaps. The deal arrives amid mounting pressure over AI training data. The same week, the Federal Trade Commission opens an investigation into OpenAI's data practices, and thousands of authors have recently accused AI developers of exploiting their work. Communications scholar Nick Diakopoulos describes such licensing as a hedge against how courts may rule on AI training. News industry analyst Ken Doctor warns that AP's influence could shape the wider industry, recalling that its 1990s decision to put content online for free led many newspapers into a bad bargain. The deal sets up licensing as an alternative to the litigation path The New York Times would take later in the year.
2023 (July 14)	Academic and educational writing	Research publication	A Stanford University team led by computer scientist James Zou publishes GPT detectors are biased against non-native English writers in the journal Patterns, following an April preprint. Testing seven widely used AI text detectors, the authors find that they classify US eighth-grade essays almost perfectly but label more than half of 91 human-written TOEFL essays by non-native speakers as AI-generated, an average false positive rate of 61.3 percent. At least one detector flags 97.8 percent of the TOEFL essays. The authors trace the bias to perplexity, a measure of how predictable text is to a language model, on which many detectors rely: writers with a narrower vocabulary produce more predictable prose. Using ChatGPT to enrich the TOEFL essays' word choice cuts false positives to 11.6 percent, while asking ChatGPT to rewrite its own college admission essays in more literary language drops detection to nearly zero. The authors strongly caution against using detectors in evaluative or educational settings. They warn that the tools may push non-native writers to rely on AI more simply to avoid being flagged.
2023 (July 18)	Creative and literary writing	Commentary	The Authors Guild submits an open letter to the CEOs of OpenAI, Alphabet, Meta, Stability AI, IBM, and Microsoft, signed by more than 15,000 writers and supporters, including Dan Brown, James Patterson, Suzanne Collins, Margaret Atwood, and Jonathan Franzen. The letter asks AI developers to seek authors' permission, and to compensate them both for past and ongoing use of their works in training and for AI outputs, whether or not those outputs infringe current law. It rejects the claim that models merely "read" their training texts, arguing that they copy works into the software. It also notes that most training books came from pirate sources, and cites the Supreme Court's Warhol v. Goldsmith ruling to question AI companies' fair use defense. The Guild warns that machine-written books could flood the market at a time when, by its own survey, authors' incomes have fallen 40 percent over a decade, and proposes a collective licensing system to pay writers. Two years later, federal judges would rule in two cases that training on books is fair use, while leaving liability for pirated copies open.
2023 (July)	Academic and educational writing	Commentary	The UNC handout Generative AI in Academic Writing offers guidance on using tools like ChatGPT and Microsoft Copilot in academic contexts. It explains how generative AI works by predicting text based on large language models trained on internet data. Potential uses include brainstorming, outlining, summarizing, editing, translating, and drafting transactional communications. The handout warns against over-reliance, highlighting risks like factual inaccuracies, fabricated citations, biased outputs, and violations of academic integrity.
2023 (August 15)	Journalism	Legal and policy	The Associated Press publishes standards for its journalists' use of generative AI, a month after licensing part of its archive to OpenAI. AP says that mindful use of AI can serve its values of accuracy, fairness, and speed, but that the journalist's central role of gathering, evaluating, and ordering facts will not change, and that it does not see AI as a replacement for journalists. Staff may experiment with ChatGPT with caution but may not use it to create publishable content, and any output from a generative AI tool is to be treated as unvetted source material, subject to editorial judgment and AP's sourcing standards. Generative AI may not be used to add or remove elements of photos, video, or audio, and AI-generated images suspected of falsely depicting reality will not be distributed. Staff are urged not to enter confidential information into AI tools, and to check that material arriving from other sources is free of AI-generated content, with the instruction that if there is any doubt about authenticity, it should not be used.
2023 (August 31)	General-purpose writing	Product launch	Baidu, China's dominant search engine company, opens its chatbot ERNIE Bot (文心一言, Wenxin Yiyan) to the public after an invitation-only preview that began on March 16. Built on Baidu's ERNIE large language model, trained on Chinese web pages, search data, and a large knowledge graph, it is presented around use cases including literary creation, business writing, and Chinese language understanding. More than 30,000 businesses apply for access to its API within two hours of the March announcement. The public launch comes as China's rules on generative AI services take effect, requiring providers to register their algorithms and meet content requirements before offering services to the public. ERNIE Bot brings large-scale AI writing assistance to Chinese, showing generative writing tools developing in parallel outside English-language markets under different regulatory and linguistic conditions.
2023 (September 1)	Creative and literary writing	Notable use	The Guardian reports that wild mushroom foraging guides apparently written by AI chatbots are being sold on Amazon, which has become a marketplace for AI-produced books passed off as the work of human authors, especially travel guides. Detection firm Originality.ai rates all four samples it examines as almost certainly machine-written. The books were first reported by 404 Media, and on August 27 the New York Mycological Society warned that Amazon and other retailers had been inundated with AI foraging and identification books, urging buyers to choose known authors because the choice could mean life or death. Foraging guide Leon Frey finds serious errors, including advice to identify mushrooms by smell and taste, dangerous when toxic species such as the death cap resemble edible ones. One book also encourages picking lion's mane, a protected species in the UK. Fungi specialist Myron Smith of Carleton University calls the books totally irresponsible, noting that distinguishing edible from poisonous mushrooms takes an experienced eye. Amazon says it is looking into the books.
2023 (September 12)	Creative and literary writing	Legal and policy	Authors Michael Chabon, David Henry Hwang, Matthew Klam, Rachel Louise Snyder, and Ayelet Waldman file a class-action complaint against Meta Platforms, alleging copyright infringement in the training of Meta's LLaMA large language models. The plaintiffs, prominent novelists and playwrights, claim that their works were copied without consent through their inclusion in Books3, a dataset derived from the shadow library Bibliotik and used as part of LLaMA's training corpus. They argue that Meta relied on large quantities of copyrighted books despite restrictive licenses, violating the Copyright Act and the Digital Millennium Copyright Act, and seek relief for unauthorized reproduction and distribution.
2023 (September 24)	Creative and literary writing	Commentary	Scottish outdoor writer Alex Roddie, editor of Sidetracked magazine and a regular contributor to The Great Outdoors, argues in an essay on his website that generative AI threatens creative writing by automating the thinking and emotional work at the heart of genuine expression. Roddie contends that AI can mimic language but lacks lived experience, intention, and individuality, which he sees as the core of meaningful writing. He warns that outsourcing creativity to AI weakens critical thinking and erodes the human voice, and that generative AI encourages conformity and dilutes originality.
2023 (September 25)	Creative and literary writing	Research publication	The Atlantic staff writer Alex Reisner publishes a searchable database of Books3, a collection of more than 191,000 pirated ebooks, most published in the previous 20 years, that was used without permission to train generative AI systems by Meta, Bloomberg, and others. Reisner identifies the books by extracting ISBNs from the unlabeled text and matching them against a book database, yielding 183,000 titles with author information that any author can now look up. Nearly every author who contacted him after his earlier reporting found their work in it. Books3 is at the center of copyright lawsuits against Meta by writers including Sarah Silverman, Michael Chabon, and Paul Tremblay. A Meta spokesperson does not directly answer questions about training its LLaMA models on pirated books, instead pointing to a court filing arguing that neither the model nor its outputs are substantially similar to the authors' books. Reisner argues that AI training practices are secretive and fundamentally nonconsensual, and that their harms to authors may lie beyond the reach of copyright law. Writers including novelist Monica Ali would use the tool to discover their own books in the dataset.
2023 (September)	Creative and literary writing	Legal and policy	Amazon (company)'s self-publishing platform Kindle Direct Publishing (KDP) tightens its rules in response to concerns about AI-generated books flooding the platform. Early in the month, KDP begins requiring authors to disclose when the content they upload includes AI-generated text, images, or translations. Later in the month, it lowers the number of new titles an account may create each day. KDP states that "we have not seen a spike in our publishing numbers" and presents the change as protection against abuse that will affect very few publishers, who can request an exception. KDP does not publish the new figure, but news outlets report a cap of three new titles a day, which KDP confirms to the Authors Guild while noting that the limit may be adjusted. The Guild, which has been in dialogue with Amazon about generative AI, welcomes the measures as tangible safeguards for authors, publishers, and readers.
2023 (November 11)	Creative and literary writing	Commentary	The Guardian publishes reflections on AI from twelve novelists and poets, including Bernardine Evaristo, Jeanette Winterson, Nick Harkaway, Harry Josephine Giles, and Stephen Marche. Evaristo, a Booker Prize winner, finds ChatGPT's creative writing weak and cliché-prone, but raises the worst-case possibility that writers adept at prompting AI could sideline those who devote their lives to the craft. Winterson, who prefers the term "alternative intelligence", says she would gladly co-write fiction with an AI, and argues that the real problem is that mostly men are designing it. Harkaway dismisses current systems as statistical models that skew toward banality. Giles argues that combinatory literature is as old as literature itself, and that the profit motive has sanded the strangeness out of commercial models. Poet Philip Terry, tracing writing machines from Gulliver's Travels to the Oulipo, predicts that formulaic genres such as Mills & Boon romance will be the easiest to automate. The final contribution, credited to novelist Nathan Filer, ends by revealing that it was written by ChatGPT-4 from his prompts and a sample of his prose, and asks whether that changes its value.
2023 (November 27)	Journalism	Notable use	Technology site Futurism reports that Sports Illustrated, a winner of numerous National Magazine Awards now managed by publisher The Arena Group, has been publishing product reviews under the bylines of writers who appear not to exist. One author, "Drew Ortiz", described as an avid outdoorsman, has no presence outside the magazine, and his profile photo is sold on a website offering AI-generated headshots. His profile is later replaced by another fake persona, and his articles are silently reattributed to the new name. A person involved in creating the content tells Futurism that the articles themselves are absolutely AI-generated. After Futurism asks for comment, all the fake authors disappear from the site. The Arena Group says the articles were licensed from contractor AdVon Commerce, which assured it they were written and edited by humans using pen names, and ends the partnership while it investigates. Futurism, whose parent company has previously worked with AdVon, finds similar fake writers at The Arena Group's financial site TheStreet, and frames the episode as part of a wider pattern of failed attempts to replace journalists with AI, following earlier embarrassments at CNET, Gizmodo, BuzzFeed, and Gannett.
2023 (November 29)	Creative and literary writing	Commentary	Novelist Monica Ali, author of the Booker Prize-shortlisted Brick Lane, delivers English PEN's H. G. Wells Lecture, published as Bookpocalypse: AI and the Risks to Literature and Free Expression. She recounts testing writing tools herself. A model trained on her own novel produces incoherent continuations, and ChatGPT and Google's Bard turn out bland stories, but Sudowrite can take an idea through outline, chapter beats, and prose to a full manuscript within days. When she asks ChatGPT for a scene from her novel Love Marriage in which a British Muslim doctor sleeps with a colleague while menstruating, it refuses, then rewrites the story so the character upholds the teachings of her faith. Ali argues that because such systems predict the most likely next words, they default to widely held, traditional viewpoints. She warns that an AI-assisted flood of books could homogenize literature and crowd out minority and dissenting voices. She also discovers her own books in the pirated Books3 dataset used to train AI models. Ali accepts that AI writing tools are here to stay, and predicts a future divide between "natural" writers who avoid them and those who use them.
2023	Creative and literary writing	Legal and policy	The Writers Guild of America (WGA) secures rules on artificial intelligence in its 2023 Minimum Basic Agreement (MBA), the contract governing writers' work for the companies that employ its members. Neither traditional nor generative AI counts as a writer, so material it produces cannot be considered literary material. Unpublished AI-generated material that a company hands to a writer does not count as assigned or source material. A company therefore cannot give a writer an AI-generated screenplay and pay only a rewrite fee: the writer is treated as the first writer. Writers may choose to use AI with the company's consent and in line with its policies, but companies cannot require them to use tools such as ChatGPT, and must disclose when material they provide is AI-generated. The WGA reserves the right to argue that using writers' material to train AI is prohibited, and the companies agree to meet with the Guild about their use of AI. The 2026 agreement, covering May 2026 to May 2030, would keep these protections and add a requirement that companies notify the Guild in writing if they license writers' scripts, or the productions based on them, to train commercial generative AI systems, allowing the Guild to seek talks over the license, including payment for writers.
2023 (December 27)	Journalism	Legal and policy	The New York Times sues OpenAI and Microsoft for copyright infringement in federal court in Manhattan, becoming the first major American media organization to sue the companies over the use of its written work. The complaint argues that millions of Times articles were used to train chatbots that now compete with the paper as a source of reliable information. It cites examples of the chatbots reproducing near-verbatim excerpts of paywalled articles, copying text from the paper's product review site Wirecutter without links, and attributing fabricated information to the Times. It seeks "billions of dollars in statutory and actual damages" and the destruction of models and training data built on Times material. The Times says it approached the companies in April about a commercial agreement, without result. OpenAI says it is surprised and disappointed given what it describes as constructive talks, and Microsoft declines to comment. The suit follows authors' lawsuits over AI training earlier in the year. It contrasts with publishers such as the Associated Press and Axel Springer, which have licensed their journalism to OpenAI. Unlike the book cases, it frames chatbots as direct competitors to journalism.
2023	Creative and literary writing	Notable use	The Land of Machine Memories, a science fiction story created with AI by Shen Yang, a professor at Tsinghua University's School of Journalism and Communication, wins second prize in the fifth youth popular science and science fiction competition run by the Jiangsu Popular Science Writers Association in Nanjing. The contest, aimed at science fiction beginners nationwide aged 14 to 45, does not ban AI. The story follows a woman who loses her memories of her family in a failed experiment and travels through a land of machine memories with an AI companion to recover them. Shen produces it in about three hours using 66 prompts, then edits the AI's text. Organizer Fu Changyi keeps its origin from the six judges, three of whom vote for it. Science fiction novelist Wang Yanzhong is impressed, while fellow judge Suo Hefu is less enthusiastic, and at least one judge readily recognizes the prose as machine-generated. Shen says he plans to publish his method so that others can follow it. An analysis on the linguistics blog Language Log notes that many of the prompts simply instruct the AI to write the next paragraph in strict accordance with an outline, suggesting the system needed close guidance.
2023–2025	General-purpose writing	Legal and policy	In the United States, efforts to require disclosure of AI-generated text produce no federal law during the first wave of generative AI adoption. The AI Labeling Act of 2023 (S.2691), introduced on July 27, 2023 by Senator Brian Schatz with Senator John Kennedy as cosponsor, would require clear disclosures on AI-generated content, including metadata identifying the generating tool, while exempting minor AI assistance, with enforcement by the Federal Trade Commission. Referred to the Senate Commerce Committee, it advances no further and expires with the 118th Congress in January 2025. The bipartisan REAL Act, introduced on December 10, 2025 by Representatives Bill Foster and Pete Sessions, would require federal agencies to label AI-generated content in official communications but does not cover publishing generally. At the state level, California's SB 942, signed on September 28, 2024 and effective January 1, 2026, becomes the most significant enacted US law on AI content labeling. The result leaves American writers, publishers, and readers without a consistent federal standard, in contrast to the European Union's binding framework.
2024–2025	Academic and educational writing	Research publication	A study by Zhongshi Wang and Mengyue Gong, published in Learned Publishing in December 2025, analyzes AI policies for peer review at 439 high-impact and 363 mid-impact journals across 21 disciplines. By August 2025, 83 percent of the high-impact and 75 percent of the mid-impact journals have such policies, and nearly a quarter of the high-impact journals revised theirs between March and August 2025 alone. The authors identify five policy types anchored to major publishers. Elsevier is the most restrictive, banning AI from peer review and editorial work entirely, even for language polishing. Springer Nature forbids uploading manuscripts to generative AI tools but otherwise requires only disclosure, Wiley allows grammar and readability polishing of review reports, and the Association for Computing Machinery is the most permissive, allowing AI to improve reviews once confidential information is removed. All agree that AI cannot replace human judgment of a manuscript's novelty and merit. Science, technology, and medicine journals are stricter than those in the humanities and social sciences. The Committee on Publication Ethics bars AI from authorship and from carrying peer review responsibility.
2024 (January)	Creative and literary writing	Notable use	Japanese novelist Rie Kudan, 33, wins the 170th Akutagawa Prize, Japan's most prestigious award for emerging fiction writers, for Tokyo-to Dojo-to (Sympathy Tower Tokyo), a novel the selection committee praises as almost flawless. She then discloses that about 5 percent of the book consists of text generated verbatim by ChatGPT. Kudan explains that she turned to ChatGPT for soft, vague language suited to the novel's muddled themes of justice and AI, and that she uses the tool as a confidant for thoughts she cannot share with other people. The AI-generated passages are largely dialogue spoken by an AI character within the novel, making their use a deliberate part of its design rather than a substitute for the author's prose. The case is among the first in which a major literary prize goes to a book containing acknowledged AI-written text.
2024 (March 18)	Academic and educational writing	Research publication	Researchers Zhuoyan Li, Chen Liang, Jing Peng, and Ming Yin publish a study, presented at the CHI Conference on Human Factors in Computing Systems, on how people value and experience AI writing assistance. In a randomized experiment, 379 participants write essays or stories either on their own, with AI editing help, or with AI drafting the text. Participants are willing to give up part of their pay to get AI support, especially for creative tasks and full content generation. AI assistance boosts productivity, confidence, and grammatical quality, but reduces participants' sense of accountability, their satisfaction, and the diversity of their writing.
2024 (July 29)	Business and professional writing	Legal and policy	The American Bar Association issues Formal Opinion 512, "Generative Artificial Intelligence Tools", its first formal ethics guidance on lawyers' use of AI. It follows widely reported cases of AI hallucinations in legal documents and earlier opinions from state bars and task forces in Texas, Illinois, California, and elsewhere. The opinion applies the Model Rules of Professional Conduct to generative AI, covering lawyers' duties of competence, confidentiality, communication with clients, candor toward courts, supervision, and reasonable fees. Building on the rules on competence, confidentiality, and responsibility for nonlawyer assistance, it treats keeping up with technologies such as AI as part of a lawyer's required competence. Commentators largely welcome it, noting that it leaves room for interpretation and for lawyers to decide whether to use AI at all, while making ignorance of its guidelines no longer acceptable. A "human in the loop" approach, such as Thomson Reuters' requirement that an attorney validate every AI-generated result before it is used in a brief or contract, emerges as accepted best practice. Questions remain about how AI costs can fairly be billed to clients, and about whether smaller organizations can afford advanced tools.
2024 (August 1)	General-purpose writing	Legal and policy	The European Union's Artificial Intelligence Act (Regulation EU 2024/1689), the world's first comprehensive AI law, enters into force. Its Article 50 sets the first binding rules that directly affect AI-generated written text. Providers of generative AI systems must mark their outputs in a machine-readable format, and organizations that use generative AI professionally must clearly label AI-generated text published to inform the public on matters of public interest. Text that has undergone genuine human editorial review, with a person or organization bearing editorial responsibility, is exempt, so the rule encourages publishers to build demonstrable human review into AI-assisted workflows. The labeling obligations are phased in, becoming applicable on August 2, 2026.
2024 (August 4)	Academic and educational writing	Tool development	The Wall Street Journal reports that OpenAI has developed a method that can reliably detect when ChatGPT has been used to write an essay or research paper, identifying AI-written text with 99.9 percent certainty, but has not released it despite widespread concern about students using AI to cheat. According to people familiar with the matter and internal documents, the project has been debated within OpenAI for about two years and has been ready for release for about a year. "It's just a matter of pressing a button," one of the people says. The report comes a year after OpenAI withdrew its earlier AI text classifier, and amid evidence that commercial detectors misjudge human writing.
2024 (August)	Creative and literary writing	Commentary	National Novel Writing Month (NaNoWriMo), the nonprofit behind the annual challenge to write a 50,000-word novel draft in November, issues a statement saying it neither supports nor condemns any approach to writing, including AI, and that categorical condemnation of AI has "classist and ableist undertones". Founded in 1999, the challenge had drawn more than 400,000 participants worldwide in 2022 and launched bestsellers by authors including Sara Gruen, Erin Morgenstern, and Rainbow Rowell. Critics reject the implication that disabled or low-income writers need AI to write well, and cite AI's reuse of copyrighted work, energy use, and bias. Members of its Writers Board resign, among them novelist Daniel José Older, who points to the organization's sponsorship by AI writing assistant ProWritingAid. NaNoWriMo clarifies that it objects only to situational abuse of AI, but authors distance themselves and sponsors withdraw. The controversy compounds earlier crises over forum moderation and child safety. On March 31, 2025, the organization would announce it is shutting down, citing financial strain and declining participation, as local writing groups form independent alternatives.
2024 (September 5)	Journalism	Research publication	Euractiv reports on a study arguing that journalism is evolving with AI rather than dying. Despite fears of newsroom collapse, Danish labour market data show little evidence of job losses linked to large language models, even as surveys find that most Americans expect AI to bring lower news quality, job cuts, and more misinformation. The study points to investigative work at outlets such as The Wall Street Journal, The Washington Post, and the Associated Press, which use machine learning and geospatial tools to find and report stories, as evidence that AI can extend reporting while human journalists keep editorial judgment.
2024 (October 1)	Academic and educational writing	Commentary	An article by Charlotte Huff, published by the American Psychological Association, examines how AI tools can help psychology researchers and students with tasks such as correcting grammar, formatting citations, and generating ideas. It warns against over-reliance, stressing the need for human authorship, ethical oversight, and transparent disclosure. Under APA Style guidelines, only humans can be listed as authors, and any use of AI must be documented. The article lists risks including misinformation, fabricated citations, unreliable detection, and bias, particularly against non-native English speakers, and concludes that AI should augment rather than replace the scholar's role in research.
2024 (October 23)	General-purpose writing	Research publication	Researchers at Google DeepMind, including its chief executive Demis Hassabis, publish SynthID-Text in Nature, describing a watermarking scheme for AI-generated text that is ready for use in production systems. Rather than detecting AI text after the fact, an approach the authors note performs inconsistently and produces more false positives for groups such as non-native English speakers, SynthID-Text slightly adjusts how a model chooses each next word, leaving a statistical signature that can later be detected without access to the model itself. In a live experiment covering nearly 20 million responses from Google's Gemini chatbot, users rate watermarked and unwatermarked answers the same. Google reports that the scheme now watermarks Gemini's responses, which the authors call the first deployment of a generative text watermark at scale, and releases its code publicly. The authors acknowledge the method's limits: it works only if the companies running models choose to apply it, is hard to enforce on open-source models, and can be weakened by paraphrasing or by attacks that strip, steal, or spoof the watermark.
2024 (October)	Creative and literary writing	Legal and policy	Penguin Random House, the largest of the Big Five English-language trade publishers, changes the copyright page in its books to forbid their use in training AI, as first reported by the trade magazine The Bookseller. The new wording states that no part of a book may be used or reproduced "for the purpose of training artificial intelligence technologies or systems". It is to appear in all new titles across the publisher's imprints, and in reprints of backlist titles. The change makes Penguin Random House the first of the Big Five to act publicly against AI training in this way. It follows other publishers' moves against AI companies' use of their work, including The New York Times' 2023 lawsuit against OpenAI and Microsoft and, that same month, its cease-and-desist letter to the AI search startup Perplexity.
2024 (November 14)	Creative and literary writing	Research publication	Philosophers of science Brian Porter and Edouard Machery of the University of Pittsburgh publish two preregistered experiments in Scientific Reports on how non-expert readers respond to AI poetry. They collect five poems each from ten well-known English-language poets, from Geoffrey Chaucer and Shakespeare to T. S. Eliot and Sylvia Plath, and use ChatGPT 3.5 to generate five more in the style of each, keeping the first results without selecting the best. Asked to tell them apart, 1,634 participants score below chance, at 46.6 percent accuracy, and judge the AI poems to be human-written more often than the real ones; experience with poetry does not help. In a second study, 696 participants rate the AI poems higher on 13 of 14 qualities, including rhythm and beauty, though not originality, while rating any poem lower when told it is AI-generated. The authors argue that readers prefer the AI poems because they are simpler and more direct, mistake the complexity of human poems for AI incoherence, and treat their own preference as evidence of human authorship.
2024 (November 19)	Creative and literary writing	Institutional adoption	HarperCollins confirms an agreement with an unnamed AI technology company allowing limited use of select nonfiction backlist titles to train AI models. The publisher presents the deal as a way to offer authors new opportunities while protecting the value of their works and their shared royalties, and says authors may opt in or decline. Its statement follows children's author Daniel Kibblesmith's posting, on Bluesky, of an email offering him $2,500 for permission to use his 2017 picture book Santa's Husband, a work of fiction despite the company's description of the deal. According to the email, titles would be licensed for three years, with protections concerning credit and limits on how much text could be reproduced verbatim in each AI response. The terms are non-negotiable, several hundred authors have already agreed, and the buyer's identity must be kept confidential. Kibblesmith urges those angered by the offer to channel their outrage into buying physical books by living authors from local bookstores.
2024	Academic and educational writing	Research publication	Quantitative studies begin to measure AI involvement in academic peer review. A study of 50,000 peer reviews for computer science conference papers submitted in 2023 and 2024, published in Nature, estimates that up to 17 percent of their sentences were likely written by a large language model. A separate study of reviews for the 2024 International Conference on Learning Representations, a leading machine learning venue, finds that at least 15.8 percent were at least partly AI-generated, and that AI-assisted reviews are associated with higher paper scores and acceptance rates. A 2026 Nature survey of 1,600 academics would find that more than half had used AI tools while peer reviewing. Researchers also document authors hiding prompts in white or microscopic text within manuscripts to steer AI-assisted reviews toward favorable assessments.
2025 (January)	Creative and literary writing	Research publication	Researchers Edirlei Soares de Lima, Margot Neggers, Bruno Feijó, Marco A. Casanova, and Antonio L. Furtado publish a paper in the journal Entertainment Computing presenting ChatGeppetto, an interactive system for AI-assisted storytelling based on semiotic theory. Drawing on four kinds of relations between story elements (syntagmatic, paradigmatic, meronymic, and antithetic), the system helps users remix existing narratives into new stories, with ChatGPT generating the text and Stable Diffusion the illustrations. In initial user studies, participants report no statistically significant difference in satisfaction compared with a baseline, though the semiotic approach is credited with making stories more emotional and memorable. The authors identify questions of authorship, bias, and AI's creative role as central to future research.
2025 (January)	General-purpose writing	Legal and policy	The U.S. Copyright Office publishes Copyright and Artificial Intelligence, Part 2: Copyrightability, a report on whether works created with AI can be protected. It concludes that prompts alone are not enough to make a user the author of AI output, describing them as instructions that convey unprotectable ideas, since currently available systems do not give users enough control over or predictability in what is generated. Human contributions such as selecting, arranging, or modifying AI output can still be protected. The report draws throughout on the Thaler v. Perlmutter case, cites the "monkey selfie" ruling Naruto v. Slater on the need for human authorship, and leaves it to the courts to give further guidance on how the human authorship requirement applies to specific uses of AI.
2025 (February 13)	Creative and literary writing	Research publication	Researchers Alicia Guo, Shreya Sathyanarayanan, Leijie Wang, Jeffrey Heer, and Amy Zhang publish a study of how 18 creative writers integrate large language models into their practice without compromising values such as authenticity and craftsmanship. Because creative writing is highly personal and driven by the writer's own choices, some critics see AI as reducing it to prompting. The writers studied, however, develop nuanced and evolving relationships with AI tools, deciding deliberately when and how to use them and building custom workflows that preserve their creative control. Many treat AI not merely as a tool but as a collaborator, and the authors draw on their practices to suggest ethical and sustainable ways for technology to support creative work.
2025 (February 20)	Creative and literary writing	Commentary	British novelist and science writer Sanjida O'Connell, author of eight novels and four nonfiction books, a Betty Trask Award winner, and a Fellow of the Royal Literary Fund, a UK charity supporting professional writers, outlines ten ways writers can use AI without sacrificing creativity, in a piece for the fund's own publication. Initially skeptical because her work had been used without permission to train AI, she arrives at a pragmatic accommodation rather than rejection. Her suggestions include using AI for brainstorming, writing prompts, finding literary agents, crafting pitches, summarizing research, proofreading, and producing publicity material. She stresses careful prompting and urges writers to keep editorial control and protect their own voice, noting that while AI output tends to be generic, it can still produce ideas that spark original writing.
2025 (March 7)	General-purpose writing	Legal and policy	Four Chinese agencies, led by the Cyberspace Administration of China together with the Ministry of Industry and Information Technology, the Ministry of Public Security, and the State Administration of Radio and Television, issue the Measures for Labeling of AI-Generated Synthetic Content, to take effect on September 1, 2025. Building on earlier rules for generative AI services, the measures cover text, images, audio, video, and virtual scenes. They require two kinds of labels: explicit labels that users can perceive, which for text means notices or symbols at its beginning, end, or middle, or beside it in the interface; and implicit labels embedded in file metadata recording the content's attributes, the provider's name, and a reference number, with digital watermarks encouraged. Platforms that distribute content must check metadata and add visible notices to content that is confirmed, declared, or suspected to be AI-generated, and users must declare AI content they publish. Providers may supply unlabeled content on request only after setting out users' responsibilities and keeping logs for six months. Anyone who maliciously removes, alters, or fakes labels, or supplies tools to do so, violates the measures.
2025 (March 18)	General-purpose writing	Legal and policy	The U.S. Court of Appeals for the D.C. Circuit rules in Thaler v. Perlmutter that the Copyright Act requires works to be authored by a human being. Computer scientist Stephen Thaler had applied in 2018 to register A Recent Entrance to Paradise, an image generated by his AI system, the "Creativity Machine", listing the AI as its sole author and conceding that it lacked human authorship. The Copyright Office refused, and a district court upheld the refusal in 2023. The appeals court finds that although the Act never defines "author", its provisions on ownership, terms measured by the author's lifespan, joint authorship, and signatures all assume a human, and that Congress adopted the Office's longstanding human-authorship requirement in the 1976 Act. It rejects Thaler's work-made-for-hire argument, while noting that works made with AI assistance can still be protected. Because Thaler disclaimed any human contribution, the court leaves open how much human input AI-assisted work requires, a question raised by Allen v. Perlmutter, a pending case over an image that its maker says took more than 600 prompts. The Supreme Court would decline to hear Thaler's appeal on March 2, 2026.
2025 (May 18)	Journalism	Notable use	The Chicago Sun-Times prints "Heat Index", a summer guide licensed from King Features, a unit of Hearst. Its summer reading list recommends books that do not exist, pairing invented titles and plot descriptions with real authors' names. After readers expose the list on social media two days later, the paper confirms that a freelancer working for King Features built the section with an AI tool without disclosing it, and that it was inserted without review by the newsroom and without any indication that it came from a third party. King Features says it bans the use of AI to create content and ends its relationship with the freelancer. The freelancer, Marco Buscaglia, would admit to having republished the AI-generated list without verifying it. Chicago Public Media chief executive Melissa Bell calls the episode a learning moment for journalism, arguing that news is valuable because of the humanity behind it. The paper does not charge subscribers for the edition, removes the section from its digital edition, and pledges to label licensed third-party content and hold it to its own editorial standards.
2025 (May 23)	Creative and literary writing	Commentary	Australian poet and artist Luke Beesley, whose collection In the Photograph was shortlisted for the 2024 Prime Minister's Literary Awards, argues in The Guardian that writing by hand, specifically in pencil, has become a quiet act of resistance against AI's growing role in literature. Drawing on his own practice of handwritten first drafts, he contends that as AI-generated text becomes harder to distinguish from human writing, authenticity will increasingly reside in the writing process and its material traces, such as notebooks, pencil stubs, and physical drafts, rather than in the finished text alone. He suggests such traces may soon serve as proof of human authorship amid growing concerns about trust and plagiarism in literary culture.
2025 (June 12)	Creative and literary writing	Research publication	Researchers Qiang Yi, Yangfan He, and Jianhui Wang present SCORE, a framework for improving the coherence and emotional consistency of long stories generated by large language models. SCORE tracks key story elements, summarizes each episode, and uses retrieval-augmented generation to look back at earlier passages, catching problems common in AI-generated fiction such as characters behaving inconsistently, shifts in emotional tone, and gaps in plot logic. In tests on LLM-generated narratives, the authors report that SCORE significantly improves coherence and stability compared with standard GPT models.
2025 (June 23)	Creative and literary writing	Legal and policy	Judge William Alsup of the U.S. District Court for the Northern District of California issues a split summary judgment ruling in Bartz v. Anthropic, a copyright suit by authors against AI company Anthropic. Anthropic trained its Claude models on millions of digitized books, some bought in print, disassembled, and scanned, and many others downloaded from pirate libraries. The authors do not claim that Claude reproduces their books, only that copying them for training is infringement. Alsup holds that training language models to generate new text is "exceedingly transformative" and a fair use, as is scanning purchased print copies. He rules against Anthropic on the pirated copies, however, finding that every fair use factor weighs against downloading millions of books for free and keeping them permanently, and sends those claims to trial on infringement and damages. Along with a ruling in a parallel case against Meta two days later, it is among the first US district court decisions on whether training large language models is fair use. The pending trial over the pirated copies would end in a $1.5 billion settlement.
2025 (June 25)	Creative and literary writing	Legal and policy	Two days after the Bartz v. Anthropic ruling, Judge Vince Chhabria of the same court grants summary judgment to Meta on fair use in Kadrey v. Meta Platforms, in which thirteen authors allege that Meta torrented their books from shadow libraries to train its Llama models. Chhabria treats downloading and training as a single transformative use, since Llama is a general text-generation tool while the books exist to be read. Unlike Alsup, however, he considers market harm the most important factor, criticizing the Bartz ruling for neglecting it and warning that no other use of copyrighted works could flood the market with competing works the way language model training can. The authors lose because they offer almost no evidence of such harm, and Chhabria stresses that the ruling's reach is limited because the plaintiffs made the wrong arguments. He also rules for Meta on Digital Millennium Copyright Act claims, leaving open a claim that Meta unlawfully distributed books while torrenting. Law firm White & Case, which represents ROSS Intelligence in a related appeal, notes that the rulings reveal deep judicial disagreement over AI training.
2025 (June 27)	Creative and literary writing	Commentary	More than 70 authors publish an open letter on Literary Hub urging major U.S. publishers to limit their use of generative AI, and it quickly gathers more than 1,100 signatures. Signatories include crime novelist Dennis Lehane, author of Mystic River and Shutter Island, and Lauren Groff, a three-time National Book Award finalist. The letter asks publishers to commit to not releasing AI-generated books, not replacing human workers with AI, and not training AI on copyrighted material without consent, and to keep using human narrators and translators. Organizers present it as a response to an existential threat to authors' livelihoods, made more urgent by recent court rulings favoring AI companies in copyright disputes. Unlike earlier campaigns aimed at AI developers, it targets publishers as the intermediaries between writers and readers. Of the major publishers, only Simon & Schuster responds publicly, affirming its commitment to protecting authors' rights.
2025 (June 30)	Academic and educational writing	Commentary	Hua Hsu, a staff writer at The New Yorker, English professor at Bard College, and winner of the 2023 Pulitzer Prize for his memoir Stay True, examines how generative AI is reshaping college writing and higher education. Drawing on classroom observation and interviews with students and educators, he finds students relying on tools such as ChatGPT for everything from formal essays to personal messages, often bypassing the writing process rather than using AI to support it. Some educators respond with in-class exams and handwritten work, while others accept AI as a legitimate learning aid. Hsu frames students' behavior less as cheating than as a rational response to an efficiency-minded, consumer-oriented educational culture that has failed to articulate what writing is supposed to develop.
2025 (July 2)	Academic and educational writing	Research publication	Data scientist Dmitry Kobak and colleagues publish a study in Science Advances estimating that at least 13.5 percent of biomedical abstracts published in 2024 were processed with large language models, or at least 200,000 papers a year. Borrowing the method of excess mortality studies, they compare word frequencies in more than 15 million PubMed abstracts from 2010 to 2024 with what earlier trends predicted, avoiding the need for detectors or samples of known AI text. 2024 shows 454 words with excess usage, more than during the COVID-19 pandemic, but where the pandemic's excess words were nouns tied to content, 2024's are almost all style words favored by LLMs: "delves" appears 28 times more often than expected, alongside words such as "underscores", "showcasing", and "crucial". The lower bound varies widely, at about 20 percent in computational fields and in papers from China, South Korea, and Taiwan, about 5 percent from the United Kingdom and Australia, and over 40 percent in some subsets. The authors caution that native English speakers may simply be better at removing telltale words, so the true share is likely higher.
2025 (August 21)	Journalism	Notable use	Trade publication Press Gazette reports that at least six outlets, including Wired and Business Insider, have removed articles by a freelance writer named Margaux Blanchard that appear to have been generated by AI, from a writer who may not exist. Wired had published her May story about couples marrying in the video game Minecraft, quoting a "digital celebrant" whose identity no one could verify. The magazine grew suspicious only when the writer could not provide details for its payment system and insisted on PayPal or a check. It acknowledges that the story skipped a proper fact-check and a senior edit, conceding that even a publication used to catching AI scammers had let one through. Business Insider removes two first-person essays by Blanchard and says it has bolstered its verification procedures. The ruse surfaced when Jacob Furedi, editor of the new magazine Dispatch, received a pitch about a secretive forensic training ground in a former Colorado mining town that he could find no trace of, and challenged the writer to provide public records requests. Blanchard does not respond to requests to prove she is a real person.
2025 (August 28)	Personal communication	Product integration	WhatsApp introduces Writing Help, an AI assistant that suggests ways to refine the tone, style, and clarity of a message before it is sent. Users write a message, tap a pencil icon, and choose from options such as professional, funny, supportive, rephrase, or proofread. The feature runs on Meta's Private Processing technology, which the company says keeps both the original message and the AI's suggestions encrypted and inaccessible to WhatsApp and Meta. It first rolls out in English in selected regions, with wider availability planned later in the year.
2025 (September 10)	Creative and literary writing	Commentary	Rice University launches a fiction course that examines how generative AI can both inspire and challenge creative writing. Rather than outsourcing storytelling to ChatGPT, students explore ways to incorporate or resist AI's influence, using it to spark ideas while confronting its limitations. The class studies critical essays and AI-generated texts, and addresses ethical questions such as copyright and environmental costs. Associate teaching professor Ian Schimmel says that grappling with AI's flaws and controversies encourages deeper reflection on creativity and technology.
2025 (September 15)	General-purpose writing	Research publication	OpenAI releases How People Use ChatGPT, a National Bureau of Economic Research working paper by its economic research team and Harvard economist David Deming, which it describes as the largest study to date of how people use the chatbot. Classifying a representative sample of conversations with an automated, privacy-preserving system, the authors trace ChatGPT's growth from its November 2022 launch to July 2025, by which time around 10 percent of the world's adults have adopted it. Non-work messages grow faster than work messages, rising from 53 percent to more than 70 percent of all use. "Practical guidance", "seeking information", and "writing" are the three most common topics, together accounting for nearly 80 percent of conversations, while programming and self-expression are relatively small. Writing dominates work use, accounting for about 40 percent of work-related messages, which the authors attribute to chatbots' ability to produce finished text, unlike search engines. Work use is most common among educated users in highly paid professional jobs, the early gender gap among users has narrowed sharply, and usage is growing fastest in lower-income countries.
2025 (October 29)	Creative and literary writing	Legal and policy	A federal judge in New York allows a class-action copyright lawsuit against OpenAI and Microsoft, brought by authors including George R. R. Martin, to proceed on claims about ChatGPT's outputs. The judge finds that a ChatGPT-generated outline for a sequel to Martin's A Song of Ice and Fire series, titled A Dance With Shadows, could be found substantially similar to Martin's work, since it draws on the series' characters, setting, and plot elements. The ruling lets the authors pursue the argument that AI outputs themselves can infringe copyright, in addition to their claims about the use of books to train AI models.
2025 (November 25)	General-purpose writing	Commentary	Australia's Macquarie Dictionary names "AI slop" its word of the year, chosen by both its expert committee and a public vote over a shortlist that includes "Ozempic face", "blind box", and "Roman Empire". The committee describes the term as AI generated slop, which lacks meaningful content or use, and says it reflects a major aspect of societal change during 2025. Having spent years learning to be search engineers to find meaningful information, the committee argues, people now need to become prompt engineers to wade through AI slop, which it expects to be a robust addition to English for years to come. Among honorable mentions is "clankers", a slur for AI-driven robots that do work normally done by humans. Committee member and crossword setter David Astle calls AI slop this decade's equivalent of spam, noting that it already forms compounds such as "slopaganda" and "slop music". The term covers AI-generated text, images, and video alike, and its spread follows a year in which such content proliferates across social media and political messaging.
2026 (February 8)	Creative and literary writing	Notable use	The New York Times reports that romance, the best-selling fiction genre and long an early adopter of new publishing technology, is rapidly taking up AI. A veteran romance author writing as Coral Hart, previously published by Harlequin and Mills & Boon, produced more than 200 novels in 2025 under 21 pen names using chatbots, chiefly Anthropic's Claude, selling about 50,000 copies. She says the AI can draft a full novel from her outline in about 45 minutes, and she has taught more than 1,600 people her methods, but she does not disclose AI use under her current pen names, citing stigma. A BookBub survey of more than 1,200 authors finds about a third using generative AI, most without telling readers. The article also traces a recurring stock phrase, a lover's name uttered "like a ragged prayer", across several AI-assisted books. Bestselling author Marie Force, who learned that more than 80 of her novels were used to train Claude, complains that AI output buries newer human writers. Future Fiction Press labels its 19 AI romances, but its cofounder Elizabeth Ann West says, "If you hide that there's A.I., it sells just fine."
2026 (March 2)	Creative and literary writing	Legal and policy	The Authors Guild opens its "Human Authored" certification to all authors whose books are published in the United States, after a members-only beta that began in January 2025 and saw thousands of books certified. The program lets authors and publishers mark books as written by humans, to distinguish them from AI-generated books that the Guild says are already saturating the market. A work qualifies if its text was fully written by people and not generated by AI, apart from a minimal amount such as AI-powered spelling and grammar checks. Using AI for research, brainstorming, outlining, or auxiliary parts such as a table of contents or index does not disqualify a book. Certified titles may carry the mark on their covers, copyright pages, and marketing. Certification costs $10 per title for nonmembers and is free for members. Authors must verify their identity and sign a license agreement, and are limited to ten titles a year. A public database lets readers check certified titles, and publishers are to be allowed to license the mark in bulk. The Guild argues that readers have a right to know whether a book was written by a human or generated with AI.
2026 (March 20)	General-purpose writing	Legal and policy	English Wikipedia adopts a guideline prohibiting the use of large language models to generate or rewrite article content, after a request for comment closes with 44 editors in favor and two opposed. The previous guideline had barred only creating new articles from scratch with AI. The new rule keeps two exceptions: editors may use LLMs to suggest copyedits to their own writing, provided the tool introduces no content of its own, and may use them for translation under a separate guideline. Because AI detection tools are unreliable and some people naturally write in a style resembling AI output, enforcement is to focus on whether edits meet the site's sourcing and content policies rather than on style alone. The guideline is proposed by editor Ilyas Lebleu, a founding member of WikiProject AI Cleanup. That volunteer group was formed after editors noticed telltale AI catchphrases in articles, and it searches for and removes AI-generated text, including hoax articles with fabricated sources. The vote follows the appearance in early March of a suspected autonomous AI agent, TomWikiAssist, that writes and edits several articles.
2026 (June 8)	Business and professional writing	Legal and policy	Senior Judge Sharion Aycock of the U.S. District Court for the Northern District of Mississippi sanctions all four lawyers on both sides of a lawsuit over unpaid legal fees against the city of Aberdeen, after filings from each side cite nonexistent cases produced with AI tools. Calling it an unusual scenario in which attorneys for both litigants engaged in similar sanctionable conduct, she removes all four from the case and cancels the trial. Plaintiff's attorney Kathleen Wilson used First Drafts, an AI drafting program, and told the court she had not known AI could invent cases; a Louisiana bankruptcy court had already disciplined her for similar conduct in April. The city's attorney Kathryn Williams, a law firm partner, relied on her firm's AI research tool despite a firm policy requiring verification, which the judge calls particularly egregious. Both are barred from the court for two years and fined $2,500 and $3,500, and the two local lawyers who sponsored them are fined $1,000 each for signing the filings. First Drafts says it tells users to verify all AI-generated citations, and the American Bar Association points to its 2024 ethics guidance.
2026 (June 23)	Academic and educational writing	Product integration	Superhuman, the company formed when Grammarly bought the email provider Superhuman in 2025 and adopted its name, acquires GPTZero, the AI detection startup that Edward Tian first built as a Princeton senior thesis project three years earlier. Terms are not disclosed. Tian says GPTZero has more than 19 million registered users and $30 million in annual recurring revenue, after raising only $13.5 million in total and becoming profitable by 2024. Superhuman already offers its own AI detection tool, which is designed to help users, often students, check whether their writing appears AI-generated and then revise it so that it does not. GPTZero, by contrast, describes its mission as helping people detect and defend against AI slop. Asked why it bought a competitor, Superhuman says that "two AI detectors are better than one."
2026 (July 20)	Creative and literary writing	Legal and policy	Judge Araceli Martínez-Olguín of the U.S. District Court for the Northern District of California grants final approval to the $1.5 billion class action settlement in Bartz v. Anthropic. Authors Andrea Bartz, Charles Graeber, and Kirk Wallace Johnson had sued AI company Anthropic over its downloading of pirated books to develop its large language models. The settlement pays roughly $3,000 per work, four times the $750 statutory minimum, and requires Anthropic to destroy files obtained from Library Genesis and Pirate Library Mirror; Anthropic states that neither dataset was used to train any of its commercially released models. Class members release only claims over the past acquisition and copying of their works through August 25, 2025, keeping all claims over AI outputs and future conduct. The court cuts class counsel's fees to about 6.8 percent of the fund and addresses 54 objections, including arguments that the total is too low against the $150,000 statutory maximum per work. The Authors Guild presents the outcome as proof that industrial-scale piracy of authors' books carries a heavy price, while similar suits against other AI companies remain pending.
2026 (August 2)	General-purpose writing	Legal and policy	The transparency obligations in Article 50 of the European Union's Artificial Intelligence Act begin to apply, two years after the law entered into force. They cover any AI system used in the situations the article lists, not only those classified as high-risk, and they extend to open-source systems. Providers of generative AI must mark text and other outputs in a machine-readable format so they can be detected as AI-generated, except where the system only assists with standard editing such as grammar correction or does not substantially alter the input. Deployers that publish AI-generated text to inform the public on matters of public interest must disclose it, unless the text has undergone substantive human review and a person or organization holds editorial responsibility for it. A provisional "AI Omnibus" agreement reached in May 2026 gives generative systems already on the market until December 2 to meet the marking requirement. To help providers and deployers comply, the European Commission published the final version of a voluntary Code of Practice on marking and labelling AI-generated content on June 10, together with a set of EU icons for labelling such content.

Numerical and visual data

Google Ngram Viewer

The chart below shows Google Ngram Viewer data for spell checker, grammar checker, machine translation, natural language generation, and language model, from 1900 to 2022. "Machine translation" is the earliest term to take off, rising sharply in the 1950s to a first peak around 1961 during early Cold War research on automatic translation, then falling to a low in the mid-1970s. It climbs to a second peak around 1989, dips around 2000, and then rises steeply from the mid-2010s, coinciding with the shift to neural machine translation, to become by far the most frequent term by 2022. "Spell checker" and "grammar checker" appear in the early 1980s, as such tools reach personal computers, and peak in the late 1990s before declining steadily as checking becomes a built-in feature that is rarely discussed. "Language model" grows gradually from the 1960s and then accelerates sharply after about 2017, the year the Transformer architecture was introduced. "Natural language generation" remains low throughout but rises slowly, overtaking both checker terms by the end of the period. Because the Ngram corpus ends in 2022, the chart does not capture the period after the release of ChatGPT.

Google Trends

The chart below shows Google Trends data for AI writing (search term), worldwide, from October 2021 to October 2026, when the screenshot was taken. Interest is ranked on a scale from 0 to 100, with 100 the peak search interest for the period. Searches are close to zero until late 2022, then begin rising around the release of ChatGPT in November 2022, reaching about 15 percent of peak interest through 2023. Interest grows gradually through 2024 and early 2025, jumps to more than half of peak around mid-2025, and holds near that level until early 2026. It then surges, reaching its peak around April 2026, before falling back sharply by late summer and partly recovering in the most recent months. Google Trends reports an increase of more than 5,000 percent compared with the previous five years.

Wikipedia Views

The chart below shows monthly pageviews of the English Wikipedia article Automated journalism, from July 2015 to September 2026, broken down by mobile web, mobile app, desktop spider, and mobile web spider traffic. Views begin in late 2016, when the article first appears, and grow steadily to around 1,000 a month by 2019, with mobile web as the largest component. Total views reach a high of about 2,400 in early 2021, driven mostly by mobile readers, and then settle at around 500 to 800 a month through 2023 and 2024, apart from a brief spike in early 2023. From 2023 onward, automated crawler traffic (desktop spider views) grows to make up most of the total, while mobile web views fall to around 100 a month. The largest spike, about 4,200 views around mid-2025, consists almost entirely of crawler traffic rather than human readers. Views stand at around 600 to 700 a month in 2026.

Meta information on the timeline

How the timeline was built

The initial version of the timeline was written by Sebastian.

Check [1].

Funding information for this timeline is available.

Feedback and comments

Feedback for the timeline can be provided at the following places:

• FIXME

See also

• Timeline of large language models
• Timeline of transformers
• Timeline of machine learning
• Timeline of machine translation
• Timeline of optical character recognition
• Timeline of AI in programming
• Timeline of AI in medicine
• Timeline of AI policy
• Timeline of AI ethics violations
• Timeline of Claude
• Timeline of Google Gemini
• Timeline of Microsoft Copilot
• Timeline of encyclopedias
• Timeline of Wikipedia
• Timeline of fact-checking
• Timeline of search engine optimization
• Timeline of Substack

External links

• U.S. Copyright Office: Copyright and Artificial Intelligence
• EU Artificial Intelligence Act, Article 50: Transparency obligations
• Wikipedia: WikiProject AI Cleanup
• Electronic Literature Directory