Short answer: no. On March 20, 2026, English Wikipedia closed a formal Request for Comment with a 44–2 vote prohibiting the use of large language models to generate or rewrite article content. Two narrow exceptions survived: LLM copyedits of text you wrote yourself, and machine translation of an existing article from another language edition — both under mandatory human review.
The news covered the vote and moved on. Nobody wrote the operator's guide: what a founder or comms lead with a half-finished ChatGPT draft actually does next. This is that guide — what the ban says, how reviewers enforce it, how to recover a rejected draft, and how to tell whether an agency quietly feeds your brief into a chatbot.
One honest note up front: this ban is good news for buyers. It kills the cheapest tier of the market — the $300 "page in 48 hours" offer that was always an AI paste job — and it validates the only approach that ever worked long-term: human-written, source-first, disclosed. If you were about to pay for slop, Wikipedia just saved you the money.
What changed on March 20, 2026
The decision came through Wikipedia's standard governance process: a Request for Comment on the policy page known as WP:LLM, formally Wikipedia:Writing articles with large language models. The outcome, as TechCrunch reported, is blunt: "The use of LLMs to generate or rewrite article content is prohibited."
This replaced earlier, vaguer language that merely discouraged generating articles from scratch — wording that left room to argue a heavily edited AI draft was acceptable. The new policy closes that door. Generating is banned. Rewriting is banned. The model's words do not belong in the encyclopedia, whether they arrived in one paste or over ten revisions.
Two exceptions made it through, and their boundaries matter:
- Copyediting your own text. The policy states that editors "are permitted to use LLMs to suggest basic copyedits to their own writing, and to incorporate some of them after human review, provided the LLM does not introduce content of its own." The text must originate with you; the model may suggest; a human reviews and accepts; nothing new enters. The policy itself warns why: "LLMs can go beyond what you ask of them and change the meaning of the text such that it is not supported by the sources cited."
- Machine translation from another edition. If an article already exists on, say, German or Ukrainian Wikipedia, machine translation can help bring it to English, again under human review. The logic: the content was already written and sourced by humans; the machine only moves it between languages.
Everything else — drafting, expanding, "improving," paraphrasing, summarizing sources into article text — is out.
The two-year escalation that got here
The March vote was not a panic reaction. It closed a two-year escalation that began almost the moment chatbots went public.
First came the volunteers. As machine-written text surfaced in drafts and articles, editors organized WikiProject AI Cleanup to find, flag, and fix it. The project did the unglamorous forensic work: cataloging chatbot tics, tracing fabricated references, building the pattern library reviewers now use daily.
Then came enforcement teeth. On August 4, 2025, English Wikipedia adopted criterion G15, a speedy-deletion rule for LLM-generated pages without human review. Speedy deletion is Wikipedia's strongest tool: an administrator can remove a page on sight, with no seven-day discussion. G15 targets two unmistakable fingerprints: "implausible non-existent references" and "communication intended for the user" — the telltale "I hope this helps" that a careless submitter forgets to delete.
Then came the March 2026 prohibition. Each step targeted the same failure: text that looks like an encyclopedia entry but does not verify. The community did not ban a technology out of distaste; it banned a failure mode it had spent two years documenting. The practical meaning: the enforcement infrastructure — a trained reviewer corps, a pattern guide, an instant-deletion criterion — is in daily use. Assume your draft will be read by someone who has personally deleted a hundred AI submissions.
Allowed vs. banned in 2026
| Use case | Verdict | Policy basis | What reviewers check |
|---|---|---|---|
| Drafting an article with ChatGPT | Banned | WP:LLM RfC, March 2026; G15 for unreviewed output | AI-writing patterns; citation spot-checks; G15 fingerprints |
| Rewriting or expanding existing article text with an LLM | Banned | Same RfC: generating or rewriting content is prohibited | Diffs that drift from what the cited sources say |
| Copyediting prose you wrote yourself | Allowed with human review | Copyedit exception in the March 2026 decision | That no new content or meaning entered the text |
| Machine translation of an article from another Wikipedia edition | Allowed with human review | Translation exception in the same decision | Translation fidelity; sources carried over intact |
| AI-suggested sources | High risk | Verifiability policy, via G15's fabricated-reference criterion | Whether each reference exists and supports the claim attached to it |
| AI-generated images | Avoid for brand and biography pages | Community consensus runs strongly against AI imagery of real people, places, and events | Whether an image misrepresents a real subject |
Two rows deserve a second look. "AI-suggested sources" is not banned outright — no policy forbids asking a model where coverage might exist. But the moment an unverified, model-invented reference lands in a draft, you are in G15 territory. And the copyedit exception is narrower than most want: it covers your prose, not an AI draft you claim as your own.
How reviewers detect AI text
Wikipedia does not run a detection algorithm and does not need one. It has something more durable: the community-maintained field guide Signs of AI writing, in its own words "a list of writing and formatting conventions typical of AI chatbots such as ChatGPT, with real examples taken from Wikipedia articles, drafts, comments, and other content."
The guide catalogs patterns across seven categories. The ones that catch corporate drafts most often:
- Stock significance inflation. Phrases like "pivotal moment in the evolution," "stands as a testament," "rich tapestry." LLMs reach for importance language because they are trained to be engaging. Encyclopedias are trained to be flat.
- Formula structures. The "not only X but also Y" construction; the closing paragraph that begins "Despite its challenges…" — outline-shaped writing that summarizes instead of informing.
- Punctuation and style tics. Em-dash overuse, title case in headings, bolded keyword lists, Markdown formatting leaking into wikitext.
- Communication intended for the user. "I hope this helps." "Certainly — here is the revised draft." G15-level giveaways that appear in real submissions more often than you would believe.
- Citation anomalies. Broken links, invalid DOIs, references that do not exist. More on this below, because it is the one that ends projects.
The part operators misunderstand: reviewers do not need to prove you used AI. At Articles for Creation (AfC) — the review queue where company and biography drafts are evaluated — the burden sits entirely on the submission. A reviewer who sees three stylistic flags and one reference that does not check out declines the draft — correctly, regardless of how the text was produced. "I wrote it myself" is a conversation you can have — not one you can win after a fabricated citation is found.
The hallucinated-citation trap
If one section is worth memorizing, it is this one. Fabricated references are the fastest way to lose a draft, and the failure LLMs commit most confidently.
A language model asked to "write a Wikipedia article about X with references" will produce citations that look perfect: a real newspaper, a plausible journalist, a credible headline, a date that fits the narrative — and an article that was never published. G15's first listed fingerprint is exactly this: "implausible non-existent references". One discovered fabrication does not cost you one reference. It costs the reviewer's trust in every other reference, and usually the draft itself.
If any AI touched your research at any stage, audit every reference before anything goes near Wikipedia:
- Open every link. Not skim — open. A reference you cannot open is a reference you do not have.
- Verify the source says what the draft claims. The second-order hallucination: the article exists, but it never makes the claim attached to it.
- Resolve every DOI and ISBN. Invalid identifiers are a listed AI sign and take seconds to check.
- Match author, outlet, and date independently. Search the headline on the publisher's own site, not just in a search engine.
- Delete anything you cannot personally verify. A thinner, verified list beats a padded one. The base rate shows the stakes: in a study of 1,009 AfC submissions, 68% were rejected, with sourcing and notability dominating decline reasons.
If your draft was already rejected
Suppose the bad news already arrived: your draft was declined, or deleted under G15. Recovery starts with an honest diagnosis, in this order.
First, read the decline reason, not just the verdict. AfC declines are templated and specific. A decline for AI-generated content or sourcing is a process failure — fixable. A decline for notability is a case failure, and no amount of rewriting fixes it: the same study found 57% of rejections cited notability. If notability is the underlying problem, stop drafting and read our Risk Report 2026 on scoring your sources before spending anything further. A formal notability audit settles the question for a fixed fee.
Second, do not "humanize" the AI draft. This is the most tempting mistake. Paraphrasing tools and "AI humanizers" change the words and keep the bones: same structure, same significance inflation, same unverified references. Reviewers who declined the first version recognize the skeleton in the second, and a cosmetic resubmission burns goodwill you will want later.
Third, rebuild from sources, not from the draft. The sequence that works: set the AI text aside entirely; assemble and verify the real source list; outline only what the sources support; write new prose by hand. The six decline patterns and the fix for each are mapped in why Wikipedia drafts get rejected. If a live page needs repair rather than recreation — stale facts, tone problems, a maintenance tag — that is editing work, which follows different rules than new-page creation.
A rejection is recoverable. A pattern of low-effort resubmissions is much less so.
How to vet whether your agency uses AI
The March 2026 decision split the market into providers whose workflow was already human and source-first, and providers whose unit economics depend on a chatbot. The second group did not announce the change. Ask, and listen for specifics:
- "Who writes the draft, and can I see drafting history?" A human process produces artifacts — research notes, outlines, redlined revisions. "Our system generates it and an editor polishes" is a description of the banned workflow.
- "How do you verify citations?" The right answer describes opening and reading every source. Hesitation here is disqualifying.
- "What changed for you after the March 2026 RfC?" A competent provider has a precise answer. A provider who has not heard of WP:LLM is not following the policies their work lives under.
- "Do you disclose paid editing?" Unrelated to AI, but the same honesty test. Disclosure is mandatory under the Wikimedia Terms of Use; an operator who hides one compliance obligation will hide another.
- "Will you show me the source list before drafting begins?" Source-first providers do this by default, because sources determine whether the project should exist at all.
Red flags cluster predictably: delivery promised in 24–48 hours, prices around $300, "proprietary AI-assisted pipeline" as a selling point, and reluctance to show anything before the finished draft. Human research and writing have a cost floor. Anyone far below it is automating — and what they are selling you is now, literally, a speedy-deletion criterion.
The compliant 2026 workflow
What survives review in 2026 is the workflow that survived review in every previous year, with the boundaries now written into policy:
- Notability before anything. Establish that independent, reliable, in-depth coverage exists. If it does not, the honest move is to wait — not to draft.
- A human-verified source dossier. Every source opened, read, and graded by a person. AI may help you search; a human confirms each item exists and says what you think it says.
- An outline that follows the sources. The article contains what independent coverage supports — nothing else, however true.
- Human drafting. A person writes the prose, in neutral register, citation by citation. The step the policy now explicitly reserves for humans.
- A citation audit. The five-step reference check above, run as a gate before submission.
- Conflict-of-interest disclosure. Paid work is declared per the Terms of Use, and the draft goes through AfC review rather than being slipped directly into the encyclopedia.
- Human responses to human reviewers. Questions at AfC get answered by a person who actually read the sources.
Where can AI legitimately sit? Two places. As a private research assistant — surfacing coverage you then verify by hand — and, under the copyedit exception, as a grammar pass over text you wrote, with every suggestion reviewed. The policy's own caution applies: models "can go beyond what you ask of them and change the meaning of the text such that it is not supported by the sources cited." The text of record stays human. This is how our page creation service is structured, and it predates the ban — the rules caught up with the workflow, not the other way around.
Three questions everyone asks
Will Wikipedia know if I used AI? Often, yes — and the more honest answer is that it does not matter. Detection is probabilistic; enforcement is not. Reviewers act on artifacts: stylistic patterns from the Signs of AI writing guide, references that fail verification, leftover chatbot phrasing. A G15 deletion does not require proving which tool you used. If the artifacts are present, the draft fails; if your process was genuinely human, they are absent. Betting on slipping through is betting against a reviewer corps with two years of practice.
Can AI help with research? Yes, under one rule: nothing a model tells you is true until a human verifies it against the source. Brainstorming where coverage might exist, summarizing an article you then read yourself, organizing notes — fine; that work never touches Wikipedia. Citing AI-suggested references without opening them is how fabricated citations enter drafts, and that is the one mistake with no recovery path.
Does the ban apply to other languages? The March 2026 decision is English Wikipedia policy. Each of the 300-plus other language editions governs itself, and rules genuinely vary — some have their own LLM restrictions, some none yet. Two practical notes: English Wikipedia is harshest on company and biography content, so it sets the de facto standard for commercial work; and the machine translation exception cuts across editions — a human-written article on one edition can be machine-translated to another under human review, making a well-sourced home-market page a strategic asset rather than a consolation prize.
The AI ban did not make Wikipedia harder. It made the shortcuts visible and the honest path official. If your page matters enough to exist, it matters enough to be built the way that holds: human-written, disclosed, source-first — the way that survives 2026 review.