Does AI scooping force researchers to hide work in progress?

Erik Hoel argues that AI has turned intellectual culture into what he calls a "dark forest," a state in which anyone who shares an idea or an unfinished piece of work risks having it absorbed and completed by someone else's AI before they can finish it themselves. His test case is the Navier-Stokes Millennium Problem: NYU mathematician Tristan Buckmaster and an Anthropic researcher spent a year advancing a line of attack originated by Diego Córdoba and Luis Martínez-Zoroa, using AI "as tools" rather than autonomously. Buckmaster says OpenAI's team "rushed to scoop them," forcing an early, partly unfinished release, and that OpenAI's earlier claim that "very little human input" went into its own competing result "turned out not to be true" once an entire human team was revealed. Buckmaster also asked OpenAI directly whether its models had been trained on, or had access to, the pair's draft chats in Codex, and reports getting no clear answer on the training question.

Hoel's mechanism is competitive and economic, not technical: because AI can now take a half-formed public idea and push it most of the way to completion, the normal incentive to share work-in-progress — presenting at a conference, posting a draft online, floating an idea on social media — now exposes it to being finished by someone else first. He generalizes past math to "all subjects," arguing the old walls of ordinary human effort, which made scooping require matching that effort, have fallen, so "now all thinking must be in stealth." He borrows "wallfacer," from Liu Cixin's The Three-Body Problem, for the posture this forces: work alone, leak nothing, and only ever surface the finished result.

This sits alongside Did GPT-5 really solve previously unsolved math problems?: both are cases where a company's claim of a largely autonomous AI math result did not survive scrutiny of the human effort behind it, feeding exactly the credit anxiety Hoel describes. It also cuts against the open culture that Why did Erdős problems become a popular AI testing ground? describes, where Bloom's public problem list let strangers collaborate — the kind of commons Hoel says the dark forest dynamic destroys. Hoel's citation of Tao's worry that AI proofs may be "odorless," lacking the insight a human proof would carry, extends rather than restates Can opaque machine learning models help prove new mathematics?, which addresses validation of a tool's output rather than the loss of a shared, legible proving culture. And where Does AI help individual scientists while narrowing scientific focus? measures a narrowing of topic coverage in published science, Hoel's claim is the softer, harder-to-measure version of the same contraction: work that would once have been shared mid-process now simply isn't shared at all.

The excerpt is a single essay built on one disputed case plus Hoel's own extrapolation from it; it does not measure how many researchers or creators have actually gone silent, and it does not establish that AI-driven scooping, rather than ordinary competitive pressure, caused Buckmaster's team to rush its release. Buckmaster's own question about whether OpenAI trained on or accessed the pair's chats is reported as unanswered, not confirmed. If Hoel's extrapolation holds even partly, the implication is that venues built for sharing unfinished work — conference talks, open problem lists, draft posts — lose their function as AI capability rises, without any formal retraction of openness as a norm.