How Wikipedia Tries to Detect AI-Generated Content

Pangram flagged perhaps 20 percent of new Wikipedia articles and drafts as AI-generated in late 2025; by early 2026, the figure was 15 or 16 percent. That movement is worth tracking, but it does not yet tell us whether Wikipedia’s rule against AI-written articles is working.

AI-generated content is harder to spot by style alone

Volunteers in WikiProject AI Cleanup are finding that the obvious signals are giving way to subtler patterns. Instead of conspicuous phrases, they are looking at structures such as the “rule of three” and negative parallelisms. Those patterns can be clues, but they are not the same as proof that a person used AI. The more detection relies on style, the more it risks confusing a recognizable writing habit with evidence of how a passage was made.

That makes this a moving target for anyone trying to build a dependable detector. A checklist based on yesterday’s telltale phrases can lose value as models change their output. The figures from Pangram show that the share it flagged shifted over a few months, but the source account says it is unclear whether that reflects policy effects or changes in AI models. The numbers are a signal to investigate, not a clean measure of enforcement success.

Misleading citations are the harder content problem

A polished paragraph is only part of the problem. AI systems may locate real sources and then misrepresent what those sources actually say, leaving citations that look plausible but do not support the accompanying claims. The source account describes misleading citations as a major worry, and that shifts the practical focus from “does this sound machine-written?” to “does this claim follow from its evidence?”

A volunteer-created AI source verifier checks whether sources support claims. That is a more useful direction for editorial work than relying on style clues alone: a citation can be checked against a claim even when the prose gives no obvious indication of its origin. Still, the existence of a verifier does not establish how accurate or comprehensive it is. In my view, the valuable unit to check is the claim-source pair, not a whole article’s supposedly machine-like voice.

Detection tools need to be treated as signals, not verdicts

Pangram is being used to flag possible AI-generated writing, while the Wikimedia Foundation is developing Paste Check to flag text pasted from some AI chatbots. They address different clues: one flags writing, while the other is described as looking for pasted text. Neither description supports treating a flag as a definitive answer about authorship or accuracy.

For a production workflow, I would use these tools to prioritize review, then verify citations and claims directly. That is a practical interpretation, not a report of a settled Wikipedia process. False certainty in either direction would be costly: a stylistic pattern could draw attention away from a factual problem, while a clean-looking passage could still contain a citation that does not support its claim.

The fall in Pangram’s flagged share is not enough to show that the rule is succeeding, and Lebleu says it is too early to determine whether it is. The useful next signal is whether citation checks and tool flags help volunteers catch unsupported claims consistently—not simply whether a detector’s percentage falls. Until that is clearer, the rule’s impact remains an open question.