Wikipedia editors who expected AI to do the housekeeping banned it from writing by March 2026
Princeton University's study found about 5% of nearly 3,000 new English Wikipedia articles from August 2024 contained substantial AI-generated text. Wikipedia had banned writing with large language models by March 2026, a Wired book excerpt reports.
Reality
- Evidence45
- Adoption30
- Hype gap+20
- Incentives
- Not scored
- Confidence45
Wikimedia Foundation says traffic from 'rogue' OpenAI agents, including hundreds of thousands of Wikidata queries, may have fed a partial outage in May. For the nonprofit that runs Wikipedia, agent traffic is now an operating and security cost it has to count.
Perspective Coverage
20 publishers
- Builder
- Builder 26%
- Operator
- Operator 48%
- Investor
- Investor 26%
Reality
- Evidence58
- Adoption
- Not scored
- Hype gap+20
- Incentives55
- Confidence68
Wikimedia says hosts should not have to absorb the cost of OpenAI's misbehaving agents
Wikimedia says agents it believes OpenAI operated made millions of requests and may have helped knock Wikidata's query service offline on 7 May. The complaint is about who pays the server and staff costs that misbehaving agents create.
Reality
- Evidence45
- Adoption
- Not scored
- Hype gap+25
- Incentives50
- Confidence40
Wikipedia's 8% drop in human readers tests what AI companies will pay for an open encyclopedia
Bernadette Meehan, the Wikimedia Foundation's new CEO, says Wikipedia's human page views fell 8% while AI tools draw on its content without credit or payment. Whether AI firms pay for it turns on whether a paid feed beats scraping a site that stays open.
Reality
- Evidence50
- Adoption
- Not scored
- Hype gap+20
- Incentives65
- Confidence45
buildOne report1 publisher Crutches built from measured failures lift a local Qwen 3B from 33% to 52% on post-cutoff facts
Qwen2.5-3B, wired to a local Wikipedia index, scored 52% on 150 post-cutoff questions it answers none of unaided, up from 33%, in a dev.to author's tests. Each fix targets a measured 3B failure, so a zero-shot 7B gained only 9 points from them, and the two readers' confidence intervals overlap.
Reality
- Evidence35
- Adoption
- Not scored
- Hype gap+25
- Incentives
- Not scored
- Confidence40
buildOne report1 publisher Agentic planning added 3 of 32 exact-match points in a 100-question GraphRAG benchmark
A TigerGraph hackathon entry raised exact match from 67% to 99% on 100 questions, with the agent alone accounting for 3 of the 32 points. The rest needed a parser that made Wikipedia infobox fields countable in the graph.
Reality
- Evidence45
- Adoption
- Not scored
- Hype gap+15
- Incentives35
- Confidence45
buildOne report1 publisher Routing every query into Google AI Mode cut publisher clicks by 18.8 points
A preregistered trial pushed one arm's entire seven-day query stream into AI Mode. The 18.8 point drop in outbound clicks is therefore the effect at full substitution, and a publisher's own exposure share scales it.
Reality
- Evidence52
- Adoption
- Not scored
- Hype gap+14
- Incentives68
- Confidence46
buildOne report1 publisher A fifteen-line abstention rule removed more correct answers than confident wrong ones
Five frontier models ran a 295-item failure corpus bare and then wrapped. Across the four that accepted the wrapper, confident errors fell from 25.8% to 7.4% and correctness fell further, from 43.9% to 21.3%.
Reality
- Evidence45
- Adoption12
- Hype gap+10
- Incentives60
- Confidence38
An Entrepreneur column moves brand accuracy from keyword rankings to the LinkedIn page and Reddit
A contributor argues that discoverability now turns on whether your story matches across pages you do not control. The evidence is client anecdote, and the consequence is an ownership question for whoever signs off on company facts.
Reality
- Evidence18
- Adoption
- Not scored
- Hype gap+45
- Incentives72
- Confidence42
Frustration over hiring and firing decisions pushed Wikimedia staff to a 158-14 union vote
The Communications Workers of America will represent more than 200 of the foundation's nearly 700 staff. Employees say what carried the ballot was how management explained its decisions, with pay and benefits secondary.
Reality
- Evidence64
- Adoption55
- Hype gap+12
- Incentives68
- Confidence58
buildOne report1 publisher RAC audits each atomic fact against retrieval after the model has already answered
The paper reports up to 30% factuality gains over prior baselines with no fine-tuning, though its own introduction concedes that RAG with Google search had already cleared those baselines by more than ten points.
Reality
- Evidence28
- Adoption
- Not scored
- Hype gap+34
- Incentives68
- Confidence44
Two letters on one Monday ended the URL swap that read X without a login
Nitter and Xcancel both stopped after X Corp wrote to them on the same August Monday, and every workflow that read a public post without an account now needs one, or a paid API key. Neither maintainer will discuss the letters yet.
Reality
- Evidence64
- Adoption45
- Hype gap+18
- Incentives68
- Confidence58
buildOne report1 publisher ChatGPT tops Google's paid-click share at 4.75%, and growth teams should reprice the auction
iPullRank's Q3 2026 panel puts ChatGPT sixth among Google destinations by clicks but first by paid share. An AI assistant is now visible in the auctions marketers already pay to win.
Reality
- Evidence34
- Adoption30
- Hype gap+26
- Incentives62
- Confidence33
Google says frontier models already know the facts they get wrong. That is a budget decision.
Google Research reports Gemini-3-Pro and GPT-5 encode 95-98% of tested facts yet fail to recall 26-34% of them, moving the fix from pretraining scale toward post-training and inference.
Reality
- Evidence55
- Adoption
- Not scored
- Hype gap+20
- Incentives68
- Confidence48