Erdosproblems.com succumbs to the AI onslaught (Blog - Changes)
B · Useful, limited Policy post by mathematician Thomas Bloom, who runs the reference site for Erdős problems, with concrete site data and a clear rule change; a first-hand practitioner account, not a study.
Summary based on the full text·Original: about 15 minutes to read
In short
Thomas Bloom, who created erdosproblems.com in May 2023, announced on 6 October 2026 that he is freezing problem comments and proof claims. Most public activity had turned into people posting AI-generated proofs without explanation, as priority claims. He is also removing open/solved status labels and credit language, and shifting the site toward high-quality human expositions. The post is a first-hand account of what happens to a community knowledge resource when AI output floods its contribution channel.
Key numbers
- 1221problems hosted on erdosproblems.com, with over 9000 comments and almost 2000 registered users
- 10,000-25,000unique visitors per day
- 291proof claims on the site at the time of the change, 155 with zero comments (count by a forum user)
As reported by the source.
English·Web page·opens erdosproblems.com
What it reports
- Launched 28 May 2023 with just over 200 problems; comments added August 2025.
- Now 1221 problems, over 9000 comments, almost 2000 registered users, 10,000 to 25,000 unique visitors a day.
- Bloom writes that AI in mathematics went from "essentially useless" to helping on some of the hardest problems in under a year, though much recent progress also came from renewed human effort.
- Main public use of the site became advertising AI-generated proofs, often unexplained; moderation was judged unsustainable because it would reject almost all submissions.
- Four changes: a hiatus on problem comments and proof claims; no problem statuses or solved percentage; no credit or ownership language for future results ("It is known that ... an explanation is provided by ..."); more emphasis on expositions.
- Lean formalisations registered on Palomar will be linked; poorly explained, unformalised proofs are unlikely to be added; suspected plagiarism or undisclosed AI use may be declined.
- A forum user counted 291 proof claims: 155 with zero comments, 14 with only the submitter's comments, 112 with at least one other user's comment; 61 problems had multiple claims.
How they got it
Bloom's own observation as site operator, plus public and private feedback he requested a few weeks earlier. The proof-claim counts come from a commenter, not from Bloom. Comments under the post include disagreement: one regular user argues the status removal only makes progress less visible and that the comment feature produced real collaborations.
Limits
- Opinion and policy, not measurement; no figures on what share of claims were AI-generated or correct.
- Changes are described as experimental and may be reversed.
Why it matters for companies
Internal wikis, idea boxes and expert forums face the same pattern: when generating an answer becomes cheap, unexplained output crowds out the discussion that made the resource useful. Bloom's response is a usable template: separate a repository of machine output from the curated space, reward explanation over first claim, and require disclosure of AI use. The cost is real too: contributors lose recognition and some collaboration stops.
Summary based on the full text. Read the original for the full argument and data.Read the original on erdosproblems.com