Wikipedia built the dataset. AI is reading it for free.

Wikipedia's operating budget is roughly 80% small donations, prompted by banners shown to human readers. Human pageviews are down about 8% year over year for March through August 2025, compared with the same months in 2024, according to the Wikimedia Foundation's own updated analysis. At the same time, bandwidth for media downloads has climbed 50% since January 2024, driven mainly by AI scrapers and crawlers, per the foundation's Swiss chapter reporting. The same content is being read by AI instead of by humans, but only the humans pay.

Two of the largest AI labs, OpenAI and Anthropic, do not appear on the foundation's public list of paying enterprise customers, while Amazon, Google, Microsoft, Meta, and Perplexity do. Last week, the foundation's new chief executive, Bernadette Meehan, named AI chatbots as a material cause of the traffic decline in her first major interview on the question, with Axios, and called on labs that scrape the site to pay for it. The foundation's Swiss chapter had already framed the stakes in a February 2026 roundtable report: "Without action, it risks becoming the backbone of AI systems while losing its own sustainability."

The mechanism is a tax on human attention. Wikipedia is free to read, but the banner asking for $3 is not free to show. A chatbot that reads the article, summarizes it, and answers the user with no link back has collected the content and skipped the tax. Roughly 80% of the foundation's operating budget comes from those small donations.

Wikimedia already has a paid product for this. Wikimedia Enterprise sells bulk, real-time access to the same content for commercial reuse. Its public customer list includes Amazon, Google, Microsoft, Meta, and Perplexity. Five of the largest AI vendors in the world are paying. OpenAI and Anthropic are not on the list. Meehan declined to confirm or deny whether either lab has an undisclosed agreement, but the omission is the question.

A reproducible per-article analysis by the researcher Martin Monperrus, run against the official Wikimedia Pageviews API, finds a median decline of 33% for the 31 articles he tracked between March 2022 and March 2026, against a 6% decline for the site as a whole over the same window. The variance is large. Articles on Music fell 81.7%, Internet culture 81.5%, Blockchain 74.2%, Vaccines 71.9%, and Climate change 66.5%. The aggregate is being held up by articles the chatbots are not yet summarizing well, or that users still need to read in full to verify.

One caveat belongs in the number. The 8% year-over-year decline only became visible after the foundation upgraded its bot detection in May 2025. Much of the traffic that had been counted as human, especially from Brazil, was reclassified as bot traffic once the new model ran. Without the methodology note, the 8% overstates the drop in actual human readership.

The four designs for the new contract are visible from the precedents. Paid licensing, the Wikimedia Enterprise model, works for the companies that sign up, but it depends on each lab agreeing that scraping a free encyclopedia without paying is the kind of thing a contract can resolve. Attribution, the "summarize and link back" pattern that some chatbots now use for news publishers, recycles some of the human-attention tax by sending readers to the source. Traffic referral would formalize that: a chatbot answer about vaccines would carry a link to the Wikipedia article, and Wikipedia would get the pageview and the donation banner. Direct underwriting, the kind of flat-fee data-layer deals that news publishers and forum operators have struck with the major labs, would have the AI vendors pay a fixed sum to support the data they sit on top of.

None of these is free. Paid licensing leaves the open question of what to do about the labs that will not sign. Attribution depends on the labs agreeing to a citation standard they do not yet share. Traffic referral asks the labs to give up part of the answer surface their product is built on. Direct underwriting concentrates the funding in fewer hands, which is its own risk for an organization that describes itself as a public good.

The next test is the next Wikimedia Enterprise renewal cycle and the EU AI Act provisions on training-data transparency, both of which fall inside the same fiscal year. The foundation's new CEO has named the question publicly. The labs not on the public list will have to answer it.