Yoshua Bengio

Who they are

Yoshua Bengio is Co-President and Scientific Director of LawZero — he co-founded Mila in 1993, co-authored the foundational Deep Learning textbook with MIT Press, and shared the 2018 Turing Award with Hinton and LeCun.

Person

Bengio took all three of his degrees — BSc in Electrical Engineering, Master's, and PhD in Computer Science — at McGill University, then joined Université de Montréal where he has stayed for a long tenure that spans the entire arc of modern deep learning. He co-founded Mila (the Quebec Artificial Intelligence Institute) in 1993, originally as LISA, and served as its Scientific Director until March 2025, when he stepped back as the institute had grown to roughly 900 researchers. Along the way he co-founded Element AI in 2016, a Montreal-based company aimed at translating AI research into business applications, which was later acquired; he took an advisory role at Maluuba around its Microsoft acquisition; and he serves as scientific advisor to Valence Discovery, Recursion Pharmaceuticals, and Perceiv AI. He co-authored the Deep Learning textbook (MIT Press, 2016) with Ian Goodfellow and Aaron Courville — still a field standard — and shared the 2018 Turing Award with Geoffrey Hinton and Yann LeCun. He launched LawZero in June 2025, a nonprofit AI safety research organization incubated at Mila, where he now serves as Co-President and Scientific Director; the organization's board includes Yuval Noah Harari, Sir John Rose, and Mariano-Florentino Cuéllar. His public writing at yoshuabengio.org covers AI safety, deep learning, GFlowNets, and what he calls 'Scientist AI' — a design philosophy for AI systems that predict and understand rather than act independently. He has said publicly that he regrets not thinking earlier about the potentially harmful consequences of AI, a candid pivot that now drives his governance and safety work.

Company

In June 2025, Bengio launched LawZero (also written LoiZéro) as a nonprofit AI safety research organization, incubated at Mila and backed by philanthropic funding. He serves as Co-President and Scientific Director. In early 2026, six global leaders — including Yuval Noah Harari, Sir John Rose, Maria Eitel, Valérie Pisano, Mariano-Florentino Cuéllar, and Laurent Charlin — joined the LawZero board, signaling an effort to build institutional credibility and policy reach. The organization's core intellectual bet is the 'Scientist AI' concept: AI architectures that predict and understand the world rather than act as autonomous agents, specifically designed to avoid deception and harmful agency. Bengio represented LawZero at the UN's Safe and Trustworthy AI event in 2026 and has taken it to the World Economic Forum and the Copenhagen Democracy Summit, framing AI safety as a global governance issue rather than a technical niche.

Market

LawZero operates in the nonprofit AI safety research space, which has grown significantly as governments and international bodies — the UN, ITU, and national governments — increase their focus on AI governance. Its closest analogues are organizations like the UK AI Safety Institute and Anthropic's safety teams, though LawZero is explicitly nonprofit and academically anchored. The broader context is rising superpower competition and regulatory nationalism, which shapes how international AI safety initiatives secure funding and political legitimacy — a dynamic Bengio navigates by engaging multilateral forums rather than aligning with any single government.

Network

Bengio's closest intellectual peers are Geoffrey Hinton and Yann LeCun — fellow 2018 Turing Award recipients and co-pioneers of deep learning who are frequently cited alongside him. His LawZero board brings in figures from outside AI research: Yuval Noah Harari, Sir John Rose, Maria Eitel, Valérie Pisano, and Mariano-Florentino Cuéllar. On the research side, Ian Goodfellow and Aaron Courville are long-standing collaborators as co-authors of the Deep Learning textbook, and Laurent Charlin stepped in as Mila's Interim Scientific Director after Bengio's transition.

Geoffrey Hinton· Co-recipient, 2018 Turing Award; deep learning pioneer
Yann LeCun· Co-recipient, 2018 Turing Award; deep learning pioneer
Ian Goodfellow· Co-author, Deep Learning (MIT Press, 2016)
Aaron Courville· Co-author, Deep Learning (MIT Press, 2016)
Yuval Noah Harari· Board member, LawZero
Laurent Charlin· Interim Scientific Director, Mila
Andrew Ng· Deep learning peer; public conversation participant

How they likely show up

Long tenure at Université de Montréal and Mila (founding Mila in 1993, stepping back from Scientific Director in March 2025) → thinks in decade-long arcs; not moved by short-term hype cycles.
Stepped back from Mila's day-to-day to focus on LawZero in 2025 → when he changes direction, it's a considered, public pivot — not incremental drift.
Active writer and speaker at yoshuabengio.org, Davos, the UN, TED, and the Copenhagen Democracy Summit → comfortable in high-stakes public forums and expects conversations to be substantive, not introductory.
Publicly stated he regrets not thinking earlier about AI's harmful consequences → intellectually honest about past errors; will engage criticism more readily than most in his position.
Hybrid role pattern — researcher, institution-builder, nonprofit founder, angel investor, policy advisor simultaneously → operates across multiple registers at once; doesn't silo 'academic' from 'governance' from 'startup' thinking.
Founded LawZero around the 'Scientist AI' concept he developed publicly before launching the org → ideas precede institutions for him; he stress-tests thinking in public before institutionalizing it.

Conversation tips

→ Engage with 'Scientist AI' specifically — it's his current intellectual project, not just a talking point. Ask what distinguishes it architecturally from current agent-based designs.
→ Reference the March 2025 transition from Mila's Scientific Director to LawZero — it was a deliberate, public choice. Asking what drove the timing will get a more honest answer than asking about LawZero's mission in the abstract.
→ He has said publicly he regrets not thinking about AI's harms sooner — don't treat this as a weakness to avoid; it's the most important thing about his current work and he'll engage it directly.
→ Cite a specific post from yoshuabengio.org or a named talk (the 2024 TED talk on catastrophic AI risks, or the Lex Fridman deep learning episode) — he'll know whether you've read it.
→ Don't conflate his safety concerns with pessimism about AI — he remains an active researcher and builder; the goal is safer AI, not less AI.

Toolbox

Openers

Open on the LawZero board expansion in early 2026 — bringing in Yuval Noah Harari, Sir John Rose, and Mariano-Florentino Cuéllar signals he's deliberately building political and cultural legitimacy alongside the technical research. Ask what that board composition is meant to unlock.
Reference his 2024 TED talk, 'The catastrophic risks of AI — and a safer path' — it's his most distilled public statement on why he left day-to-day institute leadership to focus on safety. It's a strong signal you've engaged his actual thinking, not just his Wikipedia entry.
Bring up 'Scientist AI' as a design philosophy — the idea that AI systems should predict and understand rather than act as autonomous agents. He developed this concept publicly before founding LawZero, so it sits at the center of his current intellectual and organizational work.

Discovery questions

You've described 'Scientist AI' as avoiding agency and deception by design — what does that constraint actually rule out in terms of current architectures, and which frontier labs do you think are closest to taking it seriously?
You stepped back from Mila's Scientific Director role in March 2025 after building it to roughly 900 researchers — what did that transition teach you about how to structure LawZero differently from the start?
You've been frank about regretting not thinking earlier about AI's harmful consequences — at what point in the development of deep learning do you think the field had enough information to have changed course, and what would that have required?

Avoid

Don't frame AI safety as a niche concern or a constraint on progress — he has staked his current career on it being the central challenge, and treating it as a sidebar to 'real' AI work will end the conversation fast.

Make it yours