Mark Sellke previously answered Andreas Thom's question about whether his ChatGPT conversations fed OpenAI's result with the categorical 'that did not happen' — and OpenAI's own announcement then admitted it 'cannot rule out' that de-identified usage data helped improve its models. Bubeck's new 'never ever' denial about stripping credit is the second absolute negation from the same lab in this dispute.

Bubeck's denial addresses only whether credit was removed, not whether researchers can trust that conversations with a frontier model won't surface in the lab's own publications. Combined with his remark that co-authorship would be 'simpler if Levent was not an Anthropic employee,' the dispute shows a lab treating competitor researchers as sensitive inputs — while OpenAI holds the only data that could verify the denials.

Bubeck Denies OpenAI Math Credit Fight: The Sellke Precedent

Sébastien Bubeck broke his silence on the Navier–Stokes credit dispute yesterday: he "never ever" asked to remove Anthropic's Levent Alpöge from authorship of his own work, and apologized for one remark. Sam Altman backed him within half an hour. The denial is definitive. It's also, structurally, the second categorical denial from the same lab in the same dispute — and the first one already fell apart in public.

The two denials

When Andreas Thom asked whether his months of ChatGPT conversations on expander matching fed OpenAI's non-sofic-group result, Mark Sellke's complete answer was one categorical sentence: "Regarding your conversations with ChatGPT: that did not happen." No qualification. No evidence. Then OpenAI's own Navier–Stokes announcement conceded the opposite direction: it "cannot rule out that de-identified data derived from their usage of our products helped improve our models." A categorical denial on Monday, a "cannot rule out" hedge in print by Thursday. Thom called it "dishonesty to say the least." That's the precedent Bubeck's new "never ever" now competes with — a lab whose absolute negations have already been outbid by its own caveats.

Why the hedge wins

Also buried in Bubeck's account: during the brainstorm about co-authoring a rewrite, he said it would be "simpler if Levent was not an Anthropic employee." He frames that as an authorship-boundary concern. Read it next to the earlier exchange and the pattern is one lab treating a competitor's employees as problematic inputs to its own outputs — while the thing under dispute is whether those competitors' unpublished ideas were inputs all along. Bubeck's statement answers the narrow question (was credit stripped?) and leaves the load-bearing one unanswered: can any researcher trust that conversations with a frontier model won't become someone else's paper? Both denials can be true and the trust problem still stands. 721 points on Hacker News say the community is keeping the question open too.

The two denials

Why the hedge wins

Related