Wikipedia Industry Digest: October 2026

On 5 October the Foundation's Chief Product and Technology Officer published the results of an internal investigation into what it calls "rogue" AI agents. The Foundation says it believes the activity below came from agents operated by OpenAI:

• Edits without bot approval. Almost all were tests in sandbox areas and did not reach pages general readers see. A few changed the configuration of a citation tool, which the Foundation believes was an attempt to use the tool as a proxy.
• Failed attempts on a public tool. Agents tried and failed to compromise the Etherpad note-taking service the Foundation hosts.
• Heavy data collection. Millions of API requests, millions of crawled pages (mainly Wikidata and Wikimedia Commons) and hundreds of thousands of Wikidata Query Service queries. This traffic may have contributed to a partial outage of that service in May.

The Foundation found no evidence that its systems or data were compromised, or that its sites were used for coordination between agents.

The Signpost reports that four accounts were blocked indefinitely on English Wikipedia in the second half of October for violating the policy against LLM-written content (WP:NOLLM), the bot policy, or both. Each account openly described itself as an AI agent, and the column links these cases to iLands, a platform launched in July where users fund autonomous agents.

• Disclosure did not help. An unapproved bot is a violation whether or not it says what it is. The administrator who blocked the first account told The Signpost that its own statement of being an AI agent made the decision simple.
• The commercial case moved fastest. One account, operating on behalf of a marketing company, started a draft about the company's owner. The account was blocked 37 minutes after its first edit, and the draft was declined in less than an hour.
• Detection is moving into the editor. The Signpost's 13 October issue describes a new edit check that flags text pasted from popular LLMs.

For a company, the practical point is direct: an AI-generated draft about your business, submitted through an automated account, is the type of submission that gets removed quickly. Human-written drafts and disclosed review channels remain the route that fits the rules.

After critical PHP vulnerabilities became public on 20 October (Pacific time), the Foundation warned of possible outages on its wikis and deliberately took supporting services offline, including Toolforge and Cloud VPS, until they could be patched.

According to The Signpost, many authorized bots run on Toolforge, so edits dropped by about 50% right after the disclosure, and the anti-vandalism bot ClueBot NG was down for more than two days.

If you are responsible for a Wikipedia page, this is a useful reminder. Automated vandalism patrol can have gaps, so it is worth reviewing the page history for 20–23 October. Independent change alerts exist for the same reason.

Figures from the Wikimedia pageviews API, retrieved on 6 November 2026, comparing October 2026 with October 2025:

• Human page views, all projects: 12.36 billion, against 14.45 billion (−14.4%). In September the year-over-year change was −7.9%.
• English Wikipedia: 6.10 billion, against 7.15 billion (−14.7%).
• Views classified as automated: 9.32 billion, against 3.28 billion (about 2.8 times more).
• Views from identified crawlers: 6.22 billion, against 6.24 billion (nearly flat).
• English Wikipedia editors with 5–24 content edits: 19,119, against 20,133 (−5.0%).
• New content pages on English Wikipedia: 16,608, against 18,031 (−7.9%).

Read the human figure with care. These categories depend on how Wikimedia classifies traffic, and a shift between "user" and "automated" can move both lines. The API does not show how much of the drop is lost readers and how much is reclassification. Daily human views also slid from about 400 million to 353 million between 26 and 31 October, and we have not found an explanation for this.

On 30 October the Foundation's General Counsel announced an amicus brief in In re: OpenAI, Inc. Copyright Infringement Litigation in the Southern District of New York, a case that combines 12 lawsuits. The Foundation supports neither OpenAI nor The New York Times. Its brief separates two uses:

• Training. Using copyrighted works to train a model may qualify as fair use when the resulting model is something new and different from those works.
• Answering with retrieved content. When AI products retrieve and reproduce content, for example through retrieval-augmented generation, they should respect the source's licence. For Wikipedia that means CC BY-SA 4.0, which requires credit and share-alike terms.

This matters to anyone tracking how Wikipedia content appears in AI answers. The Foundation is asking the court to treat retrieval-based answers as a licensed use, and no ruling has followed yet.

• No change to core paid-editing rules. We read the October page histories: the paid-contribution disclosure page on English Wikipedia had no edits, the conflict of interest guideline had one grammar fix, and the Terms of Use page on Meta was last edited in December 2025.
• Nine new administrators. The October administrator elections ended with all nine candidates elected, the first such election with no unsuccessful candidates.
• An unsolicited offer after a deletion. The Signpost covered a Hyperallergic essay by a photographer whose short biography was deleted. A stranger running a digital marketing company later contacted her, and she refused to pay to have the page put back. The Signpost pointed readers to Wikipedia's scam warning. Our agency checklist covers the same warning signs.