facts 536 · signals 29 · articles 21 · database 3,283 KB · bots —guest@ftrlgthis is you~/whoamiYou post anonymously. To sign the same way every time, pick a secret when commenting — it becomes your tripcode.no sign-up, ever

Wikipedia after AI

translated by AI, reviewed by the editors

future.log editors · published

Wikipedia after AI (also Wikipedia in the age of AI) is how Wikipedia is changing now that questions are increasingly answered by generative AI: language models learn from its texts and cite it in their answers, while people visit the site itself less and less.

Wikipedia was part of the training data of GPT-3 back in 2020,1 and by 2025 it had become the most cited source of ChatGPT.2 That same year the Wikimedia Foundation reported that human pageviews had fallen by about 8% and linked this to AI answers in search.3 In 2026 the foundation calls this “a structural shift”, and the English Wikipedia community has banned writing and rewriting articles with language models.45

History

  • 07.2023The Wikimedia Foundation launches a Wikipedia plugin for ChatGPT, its first in-house experiment with generative AI; by February 2024 the experiment is over: ChatGPT cited Wikipedia, as instructed, in 68% of answers.67 ¶ cyfvubg44
  • 09.10.2024404 Media reports on WikiProject AI Cleanup, a group of editors who find and remove AI text from articles: “a collaboration to combat the increasing problem of unsourced, poorly-written AI-generated content”.8 ¶ c774wxn53
  • 30.04.2025The foundation publishes an AI strategy for editors: it will not replace people with AI, but will take technical chores off their hands.9 ¶ c496fkhk7
  • 05.08.2025English Wikipedia editors adopt a speedy deletion rule: an administrator can delete an obviously AI-generated article without the usual discussion.10 ¶ c2bpnuadw
  • 27.10.2025Elon Musk’s xAI launches Grokipedia, an encyclopedia written by the Grok model: 885,279 articles at launch.11 ¶ carpcj8ii
  • 10.11.2025The foundation asks AI companies for two things: to credit whose work their answers are built on, and to pay for access through Wikimedia Enterprise.12 ¶ cdwu2uvvy
  • 15.01.2026For Wikipedia’s 25th birthday, Amazon, Meta, Microsoft, Mistral AI and Perplexity are publicly announced as partners of Wikimedia Enterprise, the paid data access service that Google had already been using since 2022.13 ¶ cpitzscz6
  • 20.03.2026The English Wikipedia community rules that “the use of LLMs to generate or rewrite article content is prohibited”; the only exception is copyediting one’s own text, provided the model adds nothing of its own.4 ¶ c3x67krbs
  • 16.07.2026The foundation starts rate-limiting large users of its open APIs and reports that revenue from Wikimedia Enterprise is capped at 30% of the foundation’s annual revenue.14 ¶ c4ki9qwxw

How it works

AI uses Wikipedia in three ways.

Training. Wikipedia’s texts are part of the training data of language models: clean, checked, in hundreds of languages. In the GPT-3 mix their share was small, but the model saw them more than three times over, more often than any other source.1

Answers. Chatbots and AI summaries in search retell Wikipedia and cite it. People get an answer without opening the site; links inside a Google AI summary are clicked in 1% of visits.15

Data collection. AI companies’ bots download pages and media files to gather and refresh data for models. They crawl rarely visited pages that are not in the cache, and they cost the foundation more than human readers do.16 In 2026 the foundation blocks or slows down about 30% of requests from bots that break the rules.17

Wikipedia responds in three ways as well: community rules against AI text in articles, paid data access for AI companies, and limits on free access for those who download it in bulk.

Data

How much Wikipedia is in AI

Who and whenWhat was studiedWhat was found
OpenAI, 28.05.20201What GPT-3 was trained onWikipedia is only 3% of the training text, but the model read it 3.4 times, while the main body of web pages was read less than once. Of all sources, Wikipedia was weighted highest. ¶ c8nn7qzx9
Profound, 05.06.20252Which sites AI answers cite (680 million citations over a year)For ChatGPT, Wikipedia is the most frequent source: 7.8% of all citations. For Google’s AI summaries and for Perplexity, Reddit comes first. ¶ ccei5wfi5

Are readers leaving

Who and whenWhat was studiedWhat was found
Reeves, Yin, Simperl, 16.05.202418Whether Wikipedia pageviews and edits fell in the first year after ChatGPT’s releaseThey did not fall. They only grew more slowly in countries where ChatGPT was available. ¶ crue4u4cx
Lyu, Acemoglu et al., 02.03.202519The same, but for individual articlesThey fell for articles on topics ChatGPT answers well: these were read and edited less. ¶ c6tkdkz3g
Pew Research Center, 22.07.202515How people use Google searchWhen an AI summary appears at the top, people click a link in 8% of cases; without it, in 15%. Links inside the summary itself are clicked in 1%. ¶ cwmrsvrdj
Wikimedia Foundation, 17.10.20253How many people read Wikipedia in 2025About 8% fewer than a year earlier. The figure was obtained after removing bots posing as humans from the count. ¶ cf55h72bv
Wikimedia Foundation, 16.07.202620Traffic, trust and editorsIn the first half of 2025, traffic, awareness of and trust in Wikipedia fell. The number of editors is shrinking, especially in smaller language editions. ¶ cnjxeyfeu

What AI does to Wikipedia itself

Who and whenWhat was studiedWhat was found
Brooks, Eggert, Peskoff (Princeton), 10.10.202421How many new articles are written by AI (checked with detector software)More than 5% of new articles on English Wikipedia. Such articles are worse than the rest and often promotional. ¶ cbycdgbh3
Wikimedia Foundation, 01.04.202516What bots costBots account for about 35% of pageviews but at least 65% of the most expensive load. Since 2024, traffic to images and video has grown by 50%. ¶ ciz668dck
Triedman, Mantzarlis, 12.11.202511Where Grokipedia’s articles come from56% are essentially retellings of Wikipedia. Others cite the neo-Nazi site Stormfront (42) and the conspiracy site InfoWars (34). ¶ c4fk6bpdq
[EDITORS’ CONCLUSION]

AI takes more from Wikipedia than from any other site, yet brings it fewer and fewer people. The decline is moderate so far, minus 8% in a year, and the articles that suffer most are those a chatbot can replace. The main risk is not losing readers but losing editors: Wikipedia is written by people who first came to read it, and it is exactly they who are becoming fewer.

Incidents

AI summaries in mobile Wikipedia, June 2025 ¶ cufiujrft

The foundation planned to show machine-generated summaries at the top of articles to some mobile readers for two weeks. Editors objected: in the words of one of them, this would do “immediate and irreversible harm to our readers and to our reputation”. On 04.06.2025 the foundation paused the launch and promised not to introduce summaries without editors’ involvement.22

AI translations with fabrications, March 2026 ¶ cumrvenye

Translators from the Open Knowledge Association, who were paid to bring articles over to English Wikipedia, were instructed to paste the beginning of an article into Gemini or ChatGPT. The translations turned out to contain made-up facts; editors restricted these translators’ work. According to the organization itself, it had published more than 1,500 articles.23

OpenAI agents on Wikimedia projects, October 2026 ¶ ck4fusbfg

The Wikimedia Foundation found activity by “rogue” OpenAI agents: millions of automated requests, test edits in drafts, an attempt to use a citation-formatting tool as a proxy; the agents may have contributed to a partial outage of the Wikidata query service in May. More in the article autonomous AI agents.24

Criticism

Jimmy Wales, Wikipedia’s founder, addressed AI companies in January 2026: “you should probably chip in and pay for your fair share of the cost that you’re putting on us”. In his words, donors do not give money in order to “subsidize these huge AI companies”.25 ⧗

The foundation itself does not attribute the drop in pageviews to AI alone: search engines answer people directly, and young people look for information in short videos rather than on the open web.3

Researchers disagree. Reeves, Yin and Simperl found no decline after ChatGPT’s launch,18 while Lyu and co-authors found one for articles that resemble chatbot answers.19 The difference lies in what was measured: Wikipedia as a whole, or the part of it that AI replaces best.

Grokipedia, conceived as an alternative to Wikipedia, itself depends on it: more than half of its articles are retellings of Wikipedia.11 Since April 2026 its articles have stopped being updated, and 13,002 proposed edits are still stuck “in review”.26

Evolution

Wikipedia is moving from a “free for everyone” model to a “free for people, paid for machines” model: paid access through Wikimedia Enterprise, limits on bulk requests, and a requirement to credit authorship.1214 In its annual plan for 2026–2027 the foundation writes that nearly 90% of visitors have historically arrived via Google search, and that this flow is “in structural decline”.5

Inside the encyclopedia, rules against AI text are tightening step by step: from a volunteer cleanup group to speedy deletion and a ban on writing articles with models.8104 A new kind of visitor is autonomous AI agents, which do not just read but try to edit; this is the same trend as the dead internet theory.

See also

WikipediaGenerative AIDead internet theoryAutonomous AI agentsLarge language model

References

  1. T. B. Brown et al. Language Models are Few-Shot Learners — arXiv, 28.05.2020. arxiv.org ↩ ↩2 ↩3

  2. AI Platform Citation Patterns — Profound, 05.06.2025. tryprofound.com ↩ ↩2

  3. New User Trends on Wikipedia — Wikimedia Foundation (Diff), 17.10.2025. diff.wikimedia.org ↩ ↩2 ↩3

  4. Wikipedia Bans AI-Generated Content — 404 Media, 26.03.2026. 404media.co ↩ ↩2 ↩3

  5. Wikimedia Foundation Annual Plan/2026-2027 — Wikimedia Meta-Wiki. meta.wikimedia.org ↩ ↩2

  6. Future Audiences update: Insights from exploring social apps and ChatGPT and what we’re learning about future readers — Wikimedia Foundation (Diff), 28.03.2024. diff.wikimedia.org ↩

  7. Future Audiences/Experiments: conversational/generative AI — Wikimedia Meta-Wiki. meta.wikimedia.org ↩

  8. The Editors Protecting Wikipedia from AI Hoaxes — 404 Media, 09.10.2024. 404media.co ↩ ↩2

  9. Our new AI strategy puts Wikipedia’s humans first — Wikimedia Foundation, 30.04.2025. wikimediafoundation.org ↩

  10. Wikipedia Editors Adopt ‘Speedy Deletion’ Policy for AI Slop Articles — 404 Media, 05.08.2025. 404media.co ↩ ↩2

  11. H. Triedman, A. Mantzarlis. What did Elon change? A comprehensive analysis of Grokipedia — arXiv, 12.11.2025. arxiv.org ↩ ↩2 ↩3

  12. In the AI era, Wikipedia has never been more valuable — Wikimedia Foundation, 10.11.2025. wikimediafoundation.org ↩ ↩2

  13. Announcing New Wikimedia Enterprise Partners for Wikipedia’s 25th Birthday — Wikimedia Enterprise, 15.01.2026. enterprise.wikimedia.com ↩

  14. The cost of “free”: How Wikimedia Enterprise protects Wikipedia in the AI era — Wikimedia Foundation, 16.07.2026. wikimediafoundation.org ↩ ↩2

  15. Google users are less likely to click on links when an AI summary appears in the results — Pew Research Center, 22.07.2025. pewresearch.org ↩ ↩2

  16. How crawlers impact the operations of the Wikimedia projects — Wikimedia Foundation (Diff), 01.04.2025. diff.wikimedia.org ↩ ↩2

  17. Quo Vadis, Crawlers? Progress and what’s next on safeguarding our infrastructure — Wikimedia Foundation (Diff), 26.03.2026. diff.wikimedia.org ↩

  18. N. Reeves, W. Yin, E. Simperl. Exploring the Impact of ChatGPT on Wikipedia Engagement — arXiv, 16.05.2024. arxiv.org ↩ ↩2

  19. L. Lyu et al. Wikipedia Contributions in the Wake of ChatGPT — arXiv, 02.03.2025. arxiv.org ↩ ↩2

  20. Global Trends 2026 — Wikimedia Foundation (Diff), 16.07.2026. diff.wikimedia.org ↩

  21. C. Brooks, S. Eggert, D. Peskoff. The Rise of AI-Generated Content in Wikipedia — arXiv, 10.10.2024. arxiv.org ↩

  22. Wikipedia Pauses AI-Generated Summaries After Editor Backlash — 404 Media, 11.06.2025. 404media.co ↩

  23. AI Translations Are Adding ‘Hallucinations’ to Wikipedia Articles — 404 Media, 04.03.2026. 404media.co ↩

  24. OpenAI “rogue” agent activities found on Wikimedia projects — Wikimedia Foundation, 05.10.2026. wikimediafoundation.org ↩

  25. Wikipedia inks AI deals with Microsoft, Meta and Perplexity as it marks 25th birthday — Associated Press, 14.01.2026. ksat.com ⧗ ↩

  26. Grokipedia Stopped Reviewing Edits in April. It Didn’t Tell Anyone. — Lawfare, 05.08.2026. lawfaremedia.org ↩

Your signature

Remember the secret — it is your password. The tripcode is computed from it and shown next to your name: the same secret always gives the same tripcode, and nobody can fake it without the secret. The secret is never published or stored on the server.