Connor Leahy: EleutherAI Co-Founder Who Now Backs an AI Ban

Connor Leahy is a German-American artificial intelligence (AI) researcher who helped build open GPT-3-style language models. He later became a public advocate for banning superintelligent AI. ControlAI's about page lists him as its US Executive Director, and the group campaigns for a legal prohibition on superintelligent AI.

Who Connor Leahy Is and How His Position Shifted

Connor Leahy built open language models before he argued for limits on AI. He co-founded EleutherAI in July 2020 with Sid Black and Leo Gao. The group's initial focus was replicating GPT-3.

In 2019, while still a student, Leahy set up a dormitory room to train a GPT-2-like model. A 2023 critique calls that work a partial GPT-2 replication. The same critique describes him then as having about two years of professional machine learning experience.

EleutherAI released The Pile, a dataset, in December 2020. Leahy isn't listed among its authors. GPT-Neo followed in March 2021, and he co-authored that paper.

In June 2021 the group released GPT-J-6B. Wikipedia calls it the largest open-source GPT-3-style language model in the world at release. GPT-NeoX-20B followed in February 2022, and Leahy co-authored that paper too.

In March 2022 he co-founded Conjecture and led it as CEO. He signed public letters on AI risk in 2023. He signed the March 2023 Future of Life Institute open letter, which calls on all AI labs to immediately pause for at least 6 months the training of AI systems more powerful than GPT-4. That wording belongs to the letter, not to him. He also signed the May 2023 Statement on AI Risk.

He is skeptical that reinforcement learning from human feedback (RLHF) solves the alignment problem, the central worry of the AI doomers. The sources used here don't record what caused the shift, so this page doesn't offer a reason.

Two basic facts are missing from the record. Wikipedia names no university and says only that he built the 2019 model "while still a student." A 2023 critique says he had a computer science undergraduate degree but names no institution. His birth year isn't published either, and Wikipedia's article carries a missing year-of-birth tag.

He also testified before the House of Commons of Canada, to the Standing Committee on Ethics and Access to Information. Wikipedia doesn't give a date for that testimony. For the wider set of people working on AI risk, see who is fighting AI.

Conjecture and Who Backs It

Leahy's stated mission for Conjecture is to "scale applied AI alignment research." In his April 2022 launch post, he called Conjecture "a new alignment startup" and said it has "VC backing." The post never uses the word nonprofit. Conjecture is venture-backed.

The post names the founders as Connor Leahy, Sid Black and Gabriel Alfour. It names the investors as Nat Friedman, Daniel Gross, Patrick and John Collison, Arthur Breitman, Andrej Karpathy, and Sam Bankman-Fried, adding "among others." The list isn't exhaustive, and the post doesn't give investment amounts.

Andrej Karpathy and Sam Bankman-Fried both appear on that list. This page doesn't characterise either person's relationship to Conjecture beyond that listing. Readers who want the full wording can read the launch post directly.

A June 2023 critique puts Conjecture's 2022 funding at roughly $10 million, "primarily from commercial investment." The critique is pseudonymous, and Leahy's post doesn't give an amount. That makes the $10 million figure an estimate from an unnamed author.

The same critique says Conjecture started with 4 employees in late 2021 and had at least 22 by June 2023. No headcount after that date appears in the sources used here.

Conjecture's products, pricing and revenue aren't published in any source used here. Leahy's launch post says nothing about products. Its current status isn't established either. Leahy published "Conjecture: A Retrospective" in March 2026, the same month he announced his move to ControlAI. The timing doesn't settle whether Conjecture still operates or has been wound down.

What ControlAI Campaigns For

ControlAI describes itself on its about page as "a non-profit working to prevent the risk of human extinction posed by superintelligent AI." The group says it works across lawmakers, media, civil society and the public. It calls for prohibiting the development of superintelligent AI, as AI experts assess it poses an extinction risk. For the control problem behind that ask, see how to stop rogue AI.

Leahy's role is recorded in two places. Wikipedia says he announced in March 2026 that he was joining ControlAI as US Director, after previously serving as an advisor. ControlAI's own site lists him as "US Executive Director," and Democracy Now! uses the same title. This page follows ControlAI's wording, dated to October 2026, because roles can change.

The about page lists Andrea Miotti as "Founder & CEO." Leahy is not its founder. Its advisors include former UK Defence Secretary Des Browne and Gabriel Alfour, who also co-founded Conjecture according to the launch post.

The about page says a bill to ban superintelligent AI was consulted on in the U.S. Congress. It also says over 85 UK lawmakers support a similar UK bill. Those are the group's own claims, and they haven't been independently verified.

The about page names no funders, no charity registration number and no jurisdiction of incorporation. ControlAI doesn't publish its funders there, and no other source used here names them. A reader weighing ControlAI's claims has no way to see who pays for them.

Wikipedia says Leahy worked with ControlAI to advocate for governments to implement a pause on artificial general intelligence development. It also says he has criticised major AI companies for trying to evade government oversight. For the group's background, see our ControlAI explainer, and for the pause campaigns, see our pause explainer.

The A Narrow Path document lists its authors as Andrea Miotti, Tolga Bilge, Dave Kasten and James Newport. Leahy isn't named among them. Its three phases are "Prevent the Development of Artificial Superintelligence for 20 Years," "Build an International AI Oversight System that Does Not Collapse Over Time," and "Build Controllable, Transformative AI." The page carries a date of June 1, 2026, which may be a revision date, so this page doesn't date the document.

What Connor Leahy Argues

Leahy's stated ask is a ban. In an August 20, 2026 interview with Democracy Now!, he said: "we should criminalize the creation of superintelligence, the same way that it is illegal to build a nuclear bomb." He also said governments should "make it a top policy priority to prevent this from happening."

He treats the issue as a security matter. He said the government "needs to intervene," because AI is "not just a commercial issue, but a national security issue." Democracy Now! describes ControlAI as briefing lawmakers on threats from unchecked emerging technologies.

On speed, Leahy said automated AI research could produce successively better systems, and that "this can go potentially very quickly." He said many AI companies aim "to achieve within the next couple of years or even months" the ability to build the next generation of AI. The interview quotes don't give a date for when that happens.

He says public awareness is the bottleneck. "Most people I talk to... have not heard about these issues," he said.

Is Connor Leahy Credible?

His open-model record is documented, but a pseudonymous 2023 critique disputes his 2019 accomplishments and Conjecture's research quality.

The Case for Him

• EleutherAI's releases are on record: GPT-Neo in March 2021, GPT-J-6B in June 2021 and GPT-NeoX-20B in February 2022.
• Leahy co-authored the GPT-Neo and GPT-NeoX papers, and Wikipedia calls GPT-J-6B the largest open-source GPT-3-style model at release.
• He has a public advocacy record: the March 2023 pause letter, the May 2023 Statement on AI Risk and testimony before Canada's House of Commons.
• His current positions are stated in his own interview and on ControlAI's public about page.

The Case Against Him

• The most substantial published critique is "Critiques of prominent AI safety labs: Conjecture," by the pseudonymous author "Omega," dated June 12, 2023.
• The critique says posts often don't state assumptions or their evidence base, and that evidence is "frequently cherry-picked."
• It says Conjecture scaled up before showing promising research results, which will make future pivots harder.
• It says Conjecture lacks a clear plan for balancing profit and safety motives.
• It argues his involvement in EleutherAI and Stability AI contributed to race dynamics.
• It says he overstated his accomplishments in 2019 as an undergraduate, and that he has been inconsistent about releasing large language models (LLMs).
• It says Conjecture and its CEO "have misrepresented themselves to external parties multiple times," and that Conjecture reacts defensively to negative feedback.

The sources used here don't include a response from Leahy or Conjecture to the June 2023 critique, so this page doesn't characterise one. The critique describes Conjecture as defensive, but that is the critic's description, not an account of any reply.

The sources can't yet settle whether his forecasts or his safety judgment hold. Several facts stay open: his birth year and university, Conjecture's funding amounts and current status, and ControlAI's funders. ControlAI's legislative claims also come from its own site.

To judge Connor Leahy's current campaign on its own terms, start with ControlAI's about page. Then read the June 2023 critique next to his April 2022 launch post.

Sources

• Connor Leahy, Wikipedia
• We Are Conjecture, A New Alignment Research Startup, April 2022
• ControlAI, About
• Democracy Now!, August 20, 2026 interview
• Critiques of prominent AI safety labs: Conjecture, June 12, 2023
• A Narrow Path