Mila - Quebec AI Institute — AI Safety Orgs

Theory of Change

Mila has two theories of change operating in tension.

Mila's institutional theory of change is straightforward: be the world's largest academic deep learning institute, train the next generation of AI researchers, transfer knowledge to industry, and maintain Canada's global AI competitiveness. The about page frames the mission as "a global hub for AI advances that benefit all" through an Open Science approach. Strategic priorities include health, environment, responsible AI, language/image, robotics, and fundamental science. The 2024-2025 Impact Report highlights "economic prosperity," "sustainable technology," and "generational challenges" alongside "trustworthy models" and "global diplomacy."

Bengio's personal theory of change diverges sharply. Yoshua Bengio, who founded Mila in 1993 and stepped down as Scientific Director in March 2025, has become one of the world's most prominent AI safety advocates. His argument: AI capabilities are advancing toward human-level within 5-20 years; we have no reliable method for ensuring alignment; commercial and geopolitical incentives create a dangerous race; and the solution requires both technical innovation (non-agentic "Scientist AI") and international governance (treaties, hardware-enabled verification, mandatory liability insurance). He assigns roughly 20% probability to catastrophic outcomes and considers even 0.1% unacceptable.

Bengio's pivot from capabilities to safety research began in early 2023, triggered by GPT-4's unexpected language mastery and concern for his grandson. He describes the psychological cost: "It is difficult because accepting the logical conclusions that follow means questioning our own role, the value of our work, our own sense of value." In June 2025, he founded LawZero, a separate nonprofit, to pursue the Scientist AI approach with $30M in philanthropic funding.

What They Do

Scale: 1,200+ student researchers, 197 faculty, 700+ peer-reviewed papers per year, 657 active research projects, 180+ papers accepted at NeurIPS/ICLR/ICML. This is overwhelmingly capabilities research. The fraction devoted to AI safety is not reported.

Safety-specific work: AI Safety Studio (chatbot safety in mental health contexts), AI Policy Fellowship (13 policymakers), hosting CAISI (Canadian AI Safety Institute) launch, Bengio's International AI Safety Report (96 experts, 30 countries -- the US declined to back the 2026 edition). The $2M Open Phil grant in 2024 supports safety research. But all of this is small relative to Mila's total output.

Commercialization push: $250M LaSalle campus with 3MW GPU infrastructure ("Sovereign AI Research Hub"). $100M VC fund with Inovia Capital targeting 55+ AI startups. 53 startups founded by Mila researchers. Entrepreneurship Lab (eLab). 176 industry partners.

LawZero (separate nonprofit): Founded June 2025 by Bengio. Developing non-agentic "Scientist AI" -- trained to understand and predict rather than act. Board includes Maria Eitel, Mariano-Florentino Cuellar, Yuval Noah Harari. $30M from Jaan Tallinn, Eric Schmidt, Open Philanthropy, FLI, Gates Foundation. This is where Bengio's safety work actually happens. Mila's Impact Report claims credit for "leading the launch" and CEO Pisano appears on LawZero's team page, but LawZero's financials are excluded from Mila's.

Policy work: Bengio testified before US Senate (2023), published in Science, Journal of Democracy, and Harvard Data Science Review. Joined UK AISI project. VP Benjamin Prud'homme works with OECD, UNESCO, UN on governance.

Key People

Yoshua Bengio (Founder, Scientific Advisor since Mar 2025). Turing Award 2018. Most-cited living scientist. Now focused on LawZero and international policy. His personal trajectory -- 30+ years of capabilities research, then a safety pivot driven by GPT-4 and love for his grandson -- is one of the most well-documented transformations in AI. Co-founded Element AI in 2016 (sold at huge loss in 2020, founders' equity wiped out).

Hugo Larochelle (Scientific Director since Sep 2025). PhD under Bengio, postdoc under Hinton. Led Google Brain/DeepMind Montreal 2016-2025. Priorities: maintaining research quality, commercializing research, supporting startups. Said he is "vigilant" about AI risks but believes AI is "ready to be adopted widely across businesses." His appointment signals Mila's post-Bengio institutional direction.

Valerie Pisano (CEO since Nov 2019). McKinsey, Cirque du Soleil. Operational and talent management background. Appears on LawZero's team page.

Pattern: Three of Mila's most prominent researchers after Bengio hold or recently held primary positions at frontier labs -- Doina Precup (Google DeepMind Research Director), Joelle Pineau (former Meta FAIR VP, now Cohere Chief AI Officer), Hugo Larochelle (ex-Google DeepMind). The talent pipeline runs from Mila to frontier labs. Bengio is the exception who went the other direction.

Money and Incentives

Dominant funder: Quebec government -- $137M+ documented over 7 years ($80M in 2018, $21M in 2023, $36M in 2025). These grants are explicitly for AI competitiveness, talent retention, and economic development. The political framing is nationalistic: ministers celebrate "Quebec's leadership" and "making Canada shine."

Federal/CIFAR: ~$20M of $60M Pan-Canadian AI Strategy (shared with Amii and Vector Institute). Canada CIFAR AI Chairs program. Similarly oriented toward competitiveness.

Industry partners: Google ~$6.75M, Microsoft $7M, DeepMind $4.5M, plus Samsung and 170+ other partners. These are the same companies whose safety practices Bengio publicly criticizes. The partnerships create direct talent pipeline incentives.

Safety-specific funding: Open Philanthropy/CG $4.7M across 4 grants (2017-2024). LawZero $30M (separate entity). These are small fractions of total funding. Safety funding represents perhaps 3-4% of Quebec government funding alone.

Commercialization: $250M LaSalle campus (compute infrastructure), $100M VC fund with Inovia, 53 startups. Mila is directly entering venture capital.

Incentive structure: Mila's money comes from sources that want AI competitiveness (government), talent for their labs (industry), and commercial returns (VC fund). The incentive to prioritize safety is structurally weak. Safety funding goes primarily to LawZero, not Mila. The Element AI failure ($600-700M valuation to $230M fire sale, founders' equity wiped) demonstrates the risks of commercialization without clear product-market fit.

Budget: Exact figures unavailable (Canadian nonprofit, no 990 filings). Estimated $30-50M+ annually based on government grants alone.

What Others Say

Yann LeCun (Turing Award co-laureate, Munk Debate): "This is just another engineering problem." Intelligence does not equal desire for domination. "We're going to build machines to be subservient to us." Has called Bengio's concerns "complete B.S."

Melanie Mitchell (Santa Fe Institute, Munk Debate): "These systems have no agency of their own. They have no desires. They don't want anything." Argues x-risk framing harms productive AI discussion and could "wipe out some of its potential benefits for humanity." Only the "bad actor" scenario is plausible. The Munk Debate audience actually shifted 6 points toward skepticism after hearing both sides.

Joelle Pineau (senior Mila researcher, former Meta FAIR VP): "I become more skeptical when the conversation descends into science-fiction scenarios. There is no realistic mechanism to make us think that this is the most likely trajectory." Called alarmist messaging "irresponsible."

Eryk Salvaggio (TechPolicy.Press): AGI framing "undermines the greatest risk that emerges from AI, which is a false sense of faith in the robustness and over-evaluation of the capacities of extremely brittle systems." Warns of regulatory capture where doomer narratives serve Big Tech.

Helen Toner (former OpenAI board member): "I worry that some aggressive AGI timeline estimates are setting them up for a boy-who-cried-wolf moment."

Stuart Russell (Berkeley, defending Bengio's approach): "The safety case isn't about imminence. It's about the fact that we still don't have a solution to the control problem." Even if AGI is decades away, urgency is justified.

US government: David Sacks (White House AI and crypto czar): "The Doomer narratives were wrong." The US declined to back the 2026 International AI Safety Report.

What's Absent

The fraction of Mila's 700+ annual papers devoted to safety vs. capabilities is unreported. No institutional statement on AI existential risk. No conflict of interest policies despite pervasive dual appointments with frontier labs. No Canadian financial disclosures (no 990 equivalent). No safety-focused hiring data. No institutional endorsement of the Scientist AI approach. No evidence of internal safety culture change beyond the AI Safety Studio. No notable researcher departures over the safety-capabilities tension. No forum posts from the AI safety community specifically evaluating Mila (as distinct from evaluating Bengio personally).

Mila - Quebec AI Institute