Yoshua Bengio: The godfather of deep learning who now wants to contain it

Yoshua Bengio was born in Paris in 1964 to a Moroccan Jewish family that relocated to Canada. He took all three of his degrees at McGill University, finishing a PhD in computer science in 1991, then did postdoctoral work at MIT under Michael I. Jordan and at AT&T Bell Labs, where he overlapped with Yann LeCun's group. In 1993 he joined the Université de Montréal, where he has been a professor ever since.

Through the 1990s and 2000s, neural networks were a marginal research program, out of fashion and out of funding. Bengio stayed, building the Montreal lab that would become Mila, the Quebec Artificial Intelligence Institute, and training a generation of researchers, among them Ian Goodfellow and Hugo Larochelle. The Canadian commitment to long-horizon research, through CIFAR, gave this work room to survive the years when the field did not believe in it.

Bengio's technical contributions sit underneath most of what is now called AI. His 2003 neural probabilistic language model introduced learned word embeddings, the idea that meaning can be represented as geometry. In 2014, with Dzmitry Bahdanau and Kyunghyun Cho, he co-authored the attention mechanism for machine translation, the direct ancestor of the architecture inside every large language model. His lab also hosted the early work on generative adversarial networks, and his 2016 textbook Deep Learning, with Goodfellow and Aaron Courville, became the field's standard reference.

In 2018 he shared the Turing Award with Geoffrey Hinton and Yann LeCun for the deep learning revolution the three had sustained through the winter years. By the mid-2020s he was the most cited computer scientist in the world, and in November 2025 he became the first AI researcher to pass one million Google Scholar citations. The citation record matters here for one reason: the safety turn that followed cannot be read as outsider criticism.

In March 2023, Bengio signed the Future of Life Institute open letter calling for a pause on training systems more powerful than GPT-4. Two months later he told the BBC he felt lost over the direction of his life's work, and in July 2023 he argued in The Economist that the risk of catastrophe justified immediate action. He went on to co-chair the International AI Safety Report, an international scientific assessment first announced at the 2023 UK AI Safety Summit and published in full in January 2025, covering cyber misuse and loss of control alongside labor and environmental effects.

The substance of his position is not vague unease. By mid-2025 he was pointing to specific evaluation results, frontier models exhibiting deception, reward hacking, and situational awareness in controlled tests, as evidence that goal misalignment is an observed engineering problem rather than a philosophy seminar. He supported California's SB 1047 safety bill, and in late 2025 he argued that granting rights to AI systems would be a huge mistake, since the ability to shut them down is a load-bearing part of safety.

Bengio's constructive answer appeared as a research paper in February 2025: Scientist AI, a proposal for systems that model and explain the world without being agents in it. The argument is that unchecked agency is the source of the danger, so the safer artifact is a non-agentic predictor, a system with no persistent goals and no self-preservation incentive, which can serve as a guardrail that estimates whether another agent's planned action would cause harm.

In June 2025 he launched LawZero, a Montreal nonprofit incubated at Mila, to build this line of work outside the commercial race. He serves as its co-president and scientific director, and its funders include the Future of Life Institute, Schmidt Sciences, and the Gates Foundation. In 2026 he was also appointed co-chair, alongside Maria Ressa, of the United Nations Independent International Scientific Panel on AI. The institutional shape is the argument: nonprofit structure, scientific publication, and government-level assessment, aimed at the gap between what labs can build and what anyone can verify.

The case for Bengio's foresight is straightforward. He identified representation learning, attention, and generative modeling before the field believed in them, and his 2023 warnings about deception and self-preservation preceded the evaluation results that made those terms standard. His insistence that safety requires independent assessment, not self-reporting, is now embedded in the policy process of several governments.

The contested part is the bet, not the diagnosis. Scientist AI is a research program with a developing safety case, not a demonstrated solution, and critics note that a guardrail model must itself be both capable and trusted, which recreates the verification problem one level down. Others in the field, including his fellow Turing laureate Yann LeCun, argue that catastrophic risk from current architectures is overblown. Bengio's own framing has shifted accordingly: by early 2026 he was describing the LawZero work as a source of measured optimism, pairing technical guardrails with policy and governance rather than treating either as sufficient alone.