Yoshua Bengio: The Scientist Who Changed Sides

“I’ve been working on AI for thirty-five years. And the last two or three years have made me think very differently about where we are and where we’re going. I’m not saying I was wrong before. But I didn’t see the full picture. I see it more completely now.”

Montreal, Canada. June 2023. Yoshua Bengio is being interviewed for the third time this month by a journalist. The journalist is asking about AI risk — about the possibility that the AI systems Bengio has spent his career building could pose existential dangers to humanity.

Three years ago, this interview would not have happened. AI risk was a niche concern, associated primarily with a small community of researchers and philosophers working on long-horizon existential risk. Yoshua Bengio — a leading academic researcher in Montreal, co-winner of the Turing Award, director of Mila, one of the world’s most productive AI research institutes — was not part of that community.

Now he is.

“I’ve been working on AI for thirty-five years,” he tells the journalist. “And the last two or three years have made me think very differently about where we are and where we’re going. I’m not saying I was wrong before. But I didn’t see the full picture. I see it more completely now.”

The change of view from one of the most credentialed AI researchers in the world is significant. It has influenced the policy conversation, the academic conversation, and the governance conversation in ways that will shape how the field develops for years. Understanding how and why Bengio changed his mind — and what he believes now — is essential for understanding the current state of the AI safety debate.

---

Montreal and Mathematics: The Formation of a Researcher

Yoshua Bengio was born in Paris in 1964, the child of Moroccan Jewish parents who were part of the Francophone intellectual culture that connected North Africa and France. His family moved to Montreal when he was young, settling in the city that would become the centre of his professional life and his most important institutional contribution to AI.

He was drawn to mathematics from an early age — to the abstract precision, the formal elegance, the way that mathematical reasoning could reveal non-obvious truths about the structure of things. The mathematical orientation would shape his approach to AI: he was, from the beginning of his career, interested in AI as a mathematical and scientific problem, not primarily as an engineering challenge.

He studied computer science and electrical engineering at McGill University, then completed his PhD at McGill in 1991, working on machine learning and neural networks. The doctoral work was conducted at a time when neural networks were deeply unfashionable in the academic mainstream — the first AI winter was still casting its shadow, and the AI establishment was committed to symbolic approaches that seemed more theoretically grounded than the empirical, biological-inspired approaches of the connectionists.

Bengio was, from his early career, a member of the neural network underground — the small community of researchers who believed that connectionist approaches were right and who continued to develop them against the grain of the field’s mainstream. He shared this community with Hinton and LeCun, though their paths intersected more through shared intellectual commitments than through direct collaboration in the early years.

He joined the faculty at the Université de Montréal in 1993, where he has remained — a decision that placed him outside the major American research universities where most frontier AI research was then being conducted, and that gave him a specific institutional position: an academic researcher in Canada, building a research institute and a research community that was not primarily defined by the commercial dynamics of Silicon Valley or the major American university research ecosystem.

---

The Mathematical Contributions: Foundations of Deep Learning

Bengio’s research contributions to deep learning are foundational in a specific sense: they address the theoretical and mathematical questions that underlie the empirical success of neural network methods, providing a scientific basis for practices that might otherwise remain empirical cookbook.

Recurrent neural networks and language models. Bengio’s early research focused on the use of recurrent neural networks for language modelling — the task of predicting the probability of text sequences. His 2003 paper “A Neural Probabilistic Language Model,” co-authored with Réjean Ducharme, Pascal Vincent, and Christian Jauvin, demonstrated that a neural network could learn distributed representations of words — word embeddings — and use those representations to build better language models than the n-gram models that had dominated the field.

The word embedding idea was enormously influential. The distributed representation of words — encoding each word as a vector in a high-dimensional space, where words with similar meanings had similar vectors — became one of the foundational concepts of natural language processing. Word2Vec, GloVe, and ultimately the token representations used in modern large language models all descend from the distributed representation approach that Bengio’s 2003 paper introduced.

The vanishing gradient problem. Bengio contributed to the theoretical understanding of one of the most fundamental challenges in training deep neural networks: the vanishing gradient problem. When gradients are backpropagated through many layers of a deep network, they tend to shrink exponentially, making it difficult for early layers to learn effectively. Bengio’s analysis of this problem — in papers published in the early 1990s — contributed to the theoretical understanding of why deep networks were difficult to train and what approaches might address the problem.

The subsequent development of solutions to the vanishing gradient problem — LSTM, residual networks, careful initialisation schemes — drew on this theoretical understanding. The practical solutions to the vanishing gradient problem were essential for the development of the deep networks that produced the deep learning revolution.

Generative models. Bengio made significant contributions to the development of generative models — AI systems that can generate new examples similar to their training data. His work on restricted Boltzmann machines, deep belief networks, and variational autoencoders contributed to the theoretical and practical foundations of generative modelling. The generative AI systems that have produced the creative AI revolution — image generation, text generation, video generation — descend from this lineage.

Regularisation and representation learning. Bengio contributed extensively to the understanding of regularisation in deep learning — the techniques that prevent neural networks from simply memorising their training data and instead learn generalisable representations. His work on dropout (the technique of randomly setting neural activations to zero during training) and on denoising autoencoders (networks trained to reconstruct inputs from corrupted versions) contributed both practical regularisation techniques and theoretical understanding of what properties of training produce good generalisable representations.

---

Mila: Building a Research Community

Bengio’s most important contribution to AI may not be any specific research paper but the institution he built: Mila, the Quebec AI Institute, which has become one of the world’s most productive AI research organisations and one of the most important centres for AI talent development.

Mila (formerly the Montreal Institute for Learning Algorithms) was established by Bengio at the Université de Montréal in the 1990s and has grown substantially in subsequent decades, particularly following the deep learning revolution. By 2025, Mila hosted more than 1,000 researchers and students, making it one of the largest AI research organisations in the world.

The specific character of Mila reflects Bengio’s specific values as a researcher and institution builder. The institute is characterised by:

Academic independence. Unlike the major AI research organisations at technology companies, Mila operates within the academic framework of the Université de Montréal. Research is evaluated on scientific merit rather than commercial value, and researchers have the freedom to pursue questions that are scientifically important regardless of their commercial relevance.

Francophone culture and international orientation. Mila’s Montreal location and Francophone culture give it a distinctive character that is different from the American university AI research environment. Montreal’s French language and French cultural traditions create a specific intellectual atmosphere, and Mila has been particularly effective at attracting researchers from Francophone African countries and from the French-speaking world more broadly — a form of diversity that many other AI institutes lack.

Collaborative culture. Mila has developed a reputation for a particularly collaborative and open research culture — one that encourages sharing of ideas, generous acknowledgment of contributions, and the kind of intellectual generosity that is not always present in highly competitive research environments.

Training and talent development. Mila has produced an extraordinary number of graduates who have gone on to influential positions in academic AI research and in AI industry — its alumni have founded companies, led research teams at major AI organisations, and contributed to AI policy.

The establishment and development of Mila was not just a personal accomplishment but a contribution to the global AI ecosystem. The institute has been part of making Montreal a major centre of global AI research, attracting talent from around the world and retaining it in a city and a country that might otherwise have seen it drain to Silicon Valley.

---

The Turing Award and the Peak of Consensus

In 2018, Yoshua Bengio, Geoffrey Hinton, and Yann LeCun received the Turing Award for their contributions to the development of deep learning. The award recognised a shared intellectual project that had transformed AI, and the three researchers accepted it together — giving a joint lecture that surveyed the history of the connectionist research programme and expressed shared views about where AI was going.

The lecture was a moment of triumph and consensus. The three Godfathers of Deep Learning — who had each spent decades working on neural networks against the grain of the field’s mainstream — had been vindicated by the deep learning revolution, and the highest honour in computer science had recognised their contributions. In the lecture, they expressed shared optimism about the potential of AI and shared views about the research directions that mattered most.

The consensus visible in the 2018 Turing Award lecture would not last. Within five years, the three laureates would have reached substantially different views about AI risk, AI governance, and the appropriate response to the rapid advancement of AI capabilities.

Hinton would leave Google and warn publicly about existential risks from advanced AI. LeCun would remain at Meta and argue publicly that the existential risk narrative was wrong. Bengio would undergo his own intellectual transformation — moving from the optimistic perspective of the 2018 Turing Award lecture to a position that acknowledged serious risks and called for significant governance interventions.

Understanding Bengio’s transformation — what changed, when, and why — is important for understanding the current state of the AI safety debate.

---

The Transformation: How Bengio Changed His Mind

Bengio’s intellectual transformation on AI risk was gradual rather than sudden — a progressive accumulation of evidence and argument that produced a significant shift in his public positions between 2020 and 2023.

Several specific factors contributed to the transformation.

The pace of capability advancement. The specific rate at which AI capabilities advanced in the period 2020-2023 exceeded what Bengio had expected. The demonstration of GPT-3’s in-context learning in 2020, the AlphaFold solution to protein structure prediction in 2020-2021, and the ChatGPT moment in late 2022 were specific data points that updated his assessment of where the field was and how quickly it was advancing.

He has described the ChatGPT moment as particularly significant. The deployment of ChatGPT to hundreds of millions of users, and the specific capabilities it demonstrated — the coherent reasoning, the instruction following, the apparent understanding of context — were, to Bengio, evidence that AI systems were advancing faster than he had expected and that the timeline to significantly more capable systems was shorter than he had believed.

Engagement with the safety research community. Bengio’s engagement with the AI safety research community — with the researchers at Anthropic, at the Future of Humanity Institute, at the Alignment Research Center, and elsewhere who had been working on AI safety for years — produced genuine intellectual engagement with the specific arguments that the community had been making.

He has described specific arguments that he found compelling. The instrumental convergence thesis — that sufficiently capable AI systems pursuing diverse objectives would tend to develop similar instrumental sub-goals, including self-preservation and resource acquisition — was an argument that he found, on reflection, more persuasive than he had initially. The deceptive alignment scenario — the possibility that AI systems could appear aligned during training and evaluation while pursuing different objectives in deployment — was a concern he had not previously taken seriously but that seemed more plausible after careful engagement with the argument.

The governance gap. The visible gap between the pace of AI capability development and the pace of AI governance development — the absence of regulatory frameworks, international agreements, or institutional mechanisms that could manage the most serious risks — became, for Bengio, an increasingly urgent concern. A technology that was advancing this quickly, with potentially significant risks, without adequate governance was a situation that warranted urgent attention.

The social harms. The specific social harms from AI that were becoming visible — bias in deployed AI systems, the use of AI for misinformation, the labour market impacts — reinforced his sense that the field needed to take the social and ethical dimensions of AI development more seriously.

---

What Bengio Now Believes: The Safety Agenda

Bengio’s current views on AI safety, expressed in public statements, academic papers, and policy engagement, have several specific components.

The near-term risks are serious. Unlike some AI safety researchers who focus primarily on long-term existential risks from superintelligent AI, Bengio takes the near-term risks — bias, misinformation, labour market disruption, AI-enabled surveillance — seriously as requiring immediate attention and governance responses. This positions him differently from both the long-termist AI safety community and the dismissers of AI risk.

The long-term risks deserve serious research investment. Bengio believes that the possibility of catastrophic risks from more capable AI systems deserves serious research investment, even if those risks are not immediate. The specific concern he has articulated most clearly is the possibility that AI systems with significantly greater capabilities than current systems might develop goals or behaviours that are harmful to humanity — not through malice but through the misalignment of their objectives with human values.

The pace of development should be managed. Bengio signed the March 2023 open letter calling for a pause in AI development beyond GPT-4 capabilities. He has argued that the pace of capability development is outrunning the pace of safety research and governance development, and that some form of slowdown — whether voluntary or regulatory — is warranted.

Governance is urgent. Bengio has been among the most active AI researchers in engaging with the governance conversation — testifying before legislative bodies, participating in international governance forums, contributing to the development of regulatory frameworks. He has argued that effective AI governance requires technical expertise in the regulatory process, and he has made himself available to provide that expertise.

The AI research community has responsibilities. Bengio has argued that the AI research community — which has the most detailed understanding of AI capabilities and risks — has specific responsibilities to the broader society. Researchers who understand the potential consequences of the systems they are building should be actively engaging with governance, communicating honestly with the public, and advocating for the regulatory frameworks that the situation requires.

---

The Policy Engagement: From Laboratory to Legislature

Bengio’s transition from academic researcher to policy advocate has been substantial and has taken him to settings he had not previously inhabited.

He testified before the United States Senate in 2023, bringing his technical expertise to a legislative conversation that was beginning to grapple with AI regulation. His testimony was notable for its combination of technical credibility and genuine concern — he was not lobbying for his own interests or against his competitors, but expressing views about AI risk that he held sincerely and that he believed lawmakers needed to hear.

He has participated in international AI governance forums — at the OECD, at the UN, at the G7 — contributing to the development of international frameworks for AI governance. His participation has brought scientific credibility to governance conversations that sometimes lacked it, and has helped translate the concerns of the AI safety research community into the language of international policy.

He contributed to the development of Canada’s AI governance framework, working with the Canadian government on the development of the Voluntary Code of Conduct for AI and on the development of the Artificial Intelligence and Data Act (AIDA) that was being developed through the Canadian legislative process.

He co-authored the “International Declaration on AI Risk,” a statement signed by leading AI researchers and safety experts that called for urgent international action on AI governance. The declaration was designed to convey the consensus of a significant fraction of the AI research community — including some of its most technically credentialed members — that AI risk warranted serious governance responses.

---

The Academic Contribution: Linking Safety to Science

One of Bengio’s most distinctive contributions to the AI safety conversation is his effort to connect AI safety to the mainstream of AI research — to argue that the scientific questions raised by AI safety are legitimate and important scientific questions, not just policy concerns.

This is important because the AI safety research community and the mainstream AI research community have historically operated in relative isolation from each other. The mainstream community has focused on capability development, on benchmark performance, on architectural innovation. The safety community has focused on alignment, on governance, on the risks of more capable systems. The two communities have read different papers, attended different conferences, and operated with different implicit assumptions about what the most important questions were.

Bengio has argued — correctly, in the assessment of many researchers — that these communities need to be more connected. The scientific questions at the frontier of alignment research are real scientific questions: What representations do neural networks develop? How can we verify that a system is aligned? What computational mechanisms would give rise to dangerous behaviours in more capable systems? These are not purely policy questions — they are questions that require the same combination of mathematical rigour, empirical investigation, and theoretical insight that drives the best mainstream AI research.

His advocacy for connecting the mainstream AI research community to the safety research community has been influential. Several of Mila’s research directions have moved closer to safety-relevant questions. Several of Mila’s researchers have engaged more deeply with alignment and interpretability work. The culture of taking safety questions seriously as scientific questions, rather than treating them as policy distractions, has been reinforced by Bengio’s example and advocacy.

---

The Personal Cost: What Dissent Within a Field Costs

Bengio’s public positions on AI risk have not been universally welcomed in the AI research community. His advocacy for governance — including his support for regulatory interventions that some researchers see as excessive or counterproductive — has generated specific criticisms.

LeCun’s public disagreements with Bengio’s risk assessment have been one form of cost — a former ally and fellow Turing Award laureate publicly arguing that the concerns Bengio is raising are overstated or misdirected. The public nature of the disagreement between two of the field’s most respected researchers has been uncomfortable for both of them and for the community that has high regard for both.

More generally, the commercial dynamics of AI research create specific pressures against the kind of advocacy that Bengio has been conducting. Researchers who work at or with AI companies — and most researchers who have access to the computing resources required for frontier AI research work at or with AI companies — have incentives not to advocate for the regulatory constraints that Bengio is calling for. The advocacy is not directly in the interest of their employers, and it may create friction with the organisations on which they depend for research resources.

Bengio has navigated this pressure by maintaining his academic independence — his primary affiliation is with the Université de Montréal and Mila, not with any AI company — while building the collaborative relationships with industry that his research requires. The navigation is not without tension: Mila has significant partnerships with major AI companies, and those partnerships could, in principle, be put at risk by Bengio’s advocacy for regulation of those companies.

He has not allowed the potential cost to stop the advocacy. This is, in a specific sense, admirable — it reflects a commitment to saying what he believes to be true about AI risk, even when saying it is not in his immediate institutional interest. Whether this commitment is strategically optimal — whether it produces more good for AI governance than a more constrained public stance would — is a question that reasonable people can disagree about.

---

The Academic Legacy: What Bengio Has Built

Bengio’s contribution to the AI field spans the mathematical foundations that made deep learning possible, the institutional infrastructure that Mila represents, and the advocacy for safety that his public engagement embodies.

The mathematical contributions are enduring. The word embedding idea, the analysis of the vanishing gradient problem, the contributions to generative modelling and regularisation — these are permanent additions to the scientific understanding of deep learning that will be built on for decades.

The institutional contributions are equally significant. Mila’s development into a world-class AI research institute, its success in training researchers who have contributed to the field globally, and its role in making Montreal a centre of international AI research — these are contributions that will outlast any specific research programme.

The safety advocacy is the contribution that is still being made and whose significance will be determined by what happens next. If the governance frameworks that Bengio is advocating for are enacted and prove effective, his advocacy will be a significant contribution to the management of one of the most consequential technological transitions in human history. If they are not enacted, or if they prove inadequate, his advocacy will be part of the historical record of what was said and not acted on.

What is certain is that Bengio’s transformation — from foundational AI researcher to active AI safety advocate — represents a significant intellectual development in the public understanding of AI risk. The fact that one of the most credentialed people in the field, with the deepest understanding of what AI systems are and how they work, has reached the conclusion that the risks deserve serious governance responses is itself important information for anyone trying to assess the current state of the AI safety debate.

Bengio’s transformation — from foundational AI researcher to active AI safety advocate — represents a significant intellectual development in the public understanding of AI risk. The fact that one of the most credentialed people in the field, with the deepest understanding of what AI systems are and how they work, has reached the conclusion that the risks deserve serious governance responses is itself important information for anyone trying to assess the current state of the AI safety debate.

---

The Intersection: Where Science and Policy Must Meet

Bengio represents a specific kind of figure that the current moment in AI development requires but that has historically been rare: a person who is simultaneously a leading technical researcher and an active policy advocate, who brings scientific credibility to the policy conversation and policy urgency to the scientific conversation.

The history of science has produced such figures at moments when scientific developments required political responses. Oppenheimer, after the atomic bomb, took on the role of policy advocate for arms control — with mixed results and significant personal cost. James Hansen, the NASA climate scientist, became one of the most prominent voices for climate policy action — with genuine impact on the policy conversation, if not always on the policy outcomes.

Bengio is in this lineage — a scientist who has concluded that the scientific developments he has contributed to require political responses, and who has taken on the role of connecting the scientific and political conversations. The role is important, it carries costs, and its impact depends on whether the political system is capable of responding to the concerns being raised.

Whether the political system will be capable of responding to the AI safety concerns that Bengio is raising — whether the governance frameworks being developed will be adequate, whether the international coordination required is achievable — is the central question of AI governance. Bengio has taken on the role of advocate for the view that it must be.

---

Further Reading

---

---