Kevin Weil is no longer OpenAI’s vice president for science. He left the company in April 2026 after serving first as chief product officer and then building OpenAI for Science. His public record supports a narrower conclusion than the old title of this article suggested: Weil helped turn frontier models into products and later organized early experiments with scientists, but neither he nor OpenAI demonstrated an autonomous system that can replace scientific method, peer review, or laboratory validation.

The clearest evidence from his science tenure is a collection of documented collaborations in which researchers used GPT-5 for literature work, calculations, candidate mechanisms, and parts of proofs. OpenAI itself described those cases as selected examples and warned that the model could invent citations, mechanisms, and proofs. That limitation is central to evaluating Weil’s thesis, not a footnote.

This article was checked on September 13, 2026. OpenAI pages are primary sources for the program’s claims and stated limitations. TechCrunch provides the dated departure report. Stoke Space and Colossal provide current biographical and board information, but each organization controls its own profile.

His OpenAI roles changed quickly

Weil joined OpenAI in 2024 as chief product officer. A current biographical profile published by Colossal describes his remit as leading teams that turned models into ChatGPT, Codex, and the OpenAI API. The same profile says he later served as vice president of OpenAI for Science. Because this is a biography hosted by an organization associated with Weil, it is useful for the career chronology but not an independent appraisal of his performance.

His earlier career helps explain the product emphasis. The profile lists senior roles at Twitter, Instagram, Meta’s Novi effort, and Planet. It also records undergraduate study in physics and mathematics at Harvard and a master’s degree in physics from Stanford. Public biographies do not support the stronger claim, repeated in some profiles, that he completed a physics PhD. This article therefore does not describe him as a physicist by doctorate.

Moving from chief product officer to a science-specific program was a material change in scope. Product leadership asks how a model reaches users, how an interface reveals capability, and how a service can be operated. Scientific work adds a stricter requirement: a useful answer must survive domain review and, where applicable, mathematical or experimental verification. A fluent output is not a result.

OpenAI for Science produced case studies, not a general proof

In November 2025, Weil authored OpenAI’s account of early experiments in accelerating science with GPT-5. The page describes collaborations across mathematics, physics, biology, computer science, astronomy, and materials science. In the examples, researchers remained responsible for defining questions, checking arguments, and deciding which ideas deserved follow-up.

OpenAI reported cases in which GPT-5 helped refine a proof, search a literature, or propose a mechanism that a scientist then tested. Those examples show a plausible workflow benefit. They do not measure an average success rate across all scientific questions, compare the tool against a controlled baseline, or establish that a model can run a research program alone.

OpenAI says this directly in the limitations section. It calls the examples curated illustrations rather than a systematic sample and notes that GPT-5 can hallucinate citations, mechanisms, or proofs. It also says expert oversight remains essential. Any assessment that repeats the successes but drops those qualifications changes the meaning of the source.

A later OpenAI report, AI as a Scientific Collaborator, compiled more examples and identified Weil by his science title. It remains a company-authored report. The work is useful evidence that the program existed and engaged named researchers, but its performance claims should not be read as an independent evaluation of GPT-5.

A government memorandum was a framework, not a delivered project

OpenAI and the U.S. Department of Energy announced a memorandum of understanding in December 2025. The OpenAI announcement said the parties would exchange expertise and explore projects connected to national laboratories and the Genesis Mission.

The wording matters. OpenAI said specific work would require follow-on agreements. A memorandum that creates a path to discuss projects is different from a procurement award, a deployed laboratory system, or a peer-reviewed scientific result. The announcement is evidence of institutional access and intent. It is not evidence that Weil’s organization completed the possible fusion, computing, or national-laboratory projects discussed in the release.

That distinction is especially important in AI-for-science coverage. Partnership announcements can arrive months or years before a reproducible outcome. A sound review should track the research question, baseline, model contribution, human validation, and eventual publication separately.

The departure closed a short operating chapter

TechCrunch reported on April 17, 2026 that Weil and Sora leader Bill Peebles had left OpenAI. Its departure report quotes Weil’s public note and says OpenAI for Science was being absorbed into other research teams. That report is the best dated public evidence for the end of his role.

The report also describes controversy around an overstated claim about mathematical problems. The responsible lesson is not to infer Weil’s motives for leaving. No public source reviewed here establishes an internal dispute, performance decision, or private negotiation. The verifiable facts are that he departed, the dedicated group did not continue in the same form, and OpenAI retained science-related work elsewhere.

OpenAI’s current science pages continue to present scientific discovery as a company priority. That continuity means Weil’s departure did not end the broader idea. It did end the basis for calling him OpenAI’s current science leader.

His public record after OpenAI is limited but specific

Stoke Space announced on July 8, 2026 that Weil had joined its board of directors. The company’s board announcement identifies his OpenAI science position as a former role and describes his experience in product strategy and public companies. It is reliable for Stoke’s appointment, while its praise of the appointee is naturally promotional.

Colossal’s biography also describes the OpenAI role in the past tense and lists boards including Cisco, Stoke Space, and The Nature Conservancy. Neither page reviewed for this update announces a new full-time operating position. This article does not infer one from an advisory profile, a board seat, or press speculation.

The separation between confirmed role and possible next move prevents a common profile error. An executive can remain active in several institutions without every affiliation being an employment relationship.

A useful AI-for-science evidence standard

Weil’s strongest contribution was framing model use as part of a workflow rather than as a substitute for scientists. OpenAI’s own case-study page says researchers set the agenda, challenge outputs, and validate results. That is a sound operating model even when individual claims later require correction.

Teams evaluating scientific assistants should ask five concrete questions:

  1. Was the problem defined before the model saw the answer or result?
  2. Is there a human or computational check that can falsify the output?
  3. Does the report include failures and discarded suggestions, not only selected wins?
  4. Can another qualified group reproduce the result from the disclosed method?
  5. Did the model change the final scientific conclusion, or only reduce time spent searching and drafting?

These questions separate research acceleration from persuasive demonstration. A model that finds relevant papers faster can be useful without discovering new science. A model that proposes a valid proof step may contribute more directly, but the proof still needs independent checking. A laboratory suggestion has no empirical standing until the experiment is run and reproduced.

Weil’s OpenAI chapter is best read as an early product experiment in this stricter environment. The public record has named collaborations, institutional agreements, explicit limitations, and a clear end date. It does not support a claim that AI had become an autonomous scientist, or that a brief executive program had already transformed discovery at scale.

Source note

Sources were checked on September 13, 2026. OpenAI materials document company positions, selected case studies, and their stated limitations. TechCrunch documents the April departure. Stoke Space and Colossal document current public affiliations. No source reviewed here establishes Weil’s private reasons for leaving or a new full-time role.