Bengio's Call: If You Care About AI Safety, Leave — Is Staying at Frontier Labs Still Worth It?

An "Open Letter of Choice" to Frontier AI Employees

Turing Award winner Yoshua Bengio has published a sharply positioned op-ed in Transformer, and its core argument can be summed up in one sentence: If you put safety first, you should leave frontier AI companies.

This is not an emotional slogan. Bengio's framing is one of "choice" — he describes the situation facing frontier AI practitioners as a personal decision that must be made before time runs out, not an industry debate that can be discussed at leisure. He supports this position with two key judgments: first, we genuinely have room to choose; second, we no longer have much time to hesitate.

One detail worth noting: Bengio does not single out any specific company or individual. His article is fundamentally an interrogation of professional identity: the code you write every day, the models you train, the alignment strategies you tune — are they actually pushing the technology in a safer direction, or merely accelerating the expansion of capability boundaries? No third party can answer that question for you.

Yoshua Bengio is a professor at the University of Montreal and scientific director of the Mila research institute. Along with Geoffrey Hinton and Yann LeCun, he is known as one of the "three giants" of deep learning, and the three jointly received the 2018 Turing Award. In 2023, Bengio signed an open letter calling for a pause on large-scale AI training, and he has continued to speak out on AI safety and governance since then, including work related to chairing the International AI Safety Report. His shift in stance is symbolically significant: a founding figure of deep learning publicly expressing reservations about the frontier R&D path carries more industry impact than the voices of outside critics.

Why "Stay Inside and Push for Change" No Longer Works This Time

Over the past few years, a widely accepted course of action has existed within the AI safety community: stay inside the most frontier labs and influence decisions from within. The logic behind this path is that if safety researchers all leave, it will only let capability-first voices monopolize the conversation.

Bengio's article essentially challenges that premise. When the pace of the capability race becomes so fast that safety assessments struggle to keep up, both the organizational resources and the time window needed for "internal checks and balances" are being compressed. His argument points to an uncomfortable conclusion: staying at a frontier lab may no longer be an effective safety strategy, but simply a way of lending legitimacy to the capability race.

The point of disagreement here is very real. Those who support staying inside would point out that leaving means handing over the seats of safety and governance to others, and that frontier labs have the most complete model access — external researchers find it difficult even to conduct independent evaluations. Bengio's position is closer to "voting with your feet" — when the moral cost at the individual level is clearly marked, talent mobility itself becomes a constraining force.

Both logics have a basis in reality, and Bengio chooses the more radical side. He does not offer a transitional plan, nor does he provide a "checklist of conditions for staying." This uncompromising expression is itself the article's strategy: pushing the urgency of the decision directly in front of the reader.

The term "frontier AI company" usually refers to the small number of labs that invest the greatest compute in pushing the boundaries of model capability and pursuing the most advanced performance, such as OpenAI, Anthropic, Google DeepMind, and Meta AI. Their common characteristics include: possessing tens of thousands of high-end training chips, controlling the largest-scale data and model access, and iterating on a monthly cycle. The definition of "frontier" is itself dynamic — today's boundary capabilities will be caught up to broadly across the industry within a year or two, so these labs are always under structural pressure to "accelerate or fall behind," and safety assessments are often forced to be completed within shorter time windows.

What Realistic Options Exist for Safety-First Practitioners

If you accept Bengio's framework, the question shifts from "should I leave?" to "where should I go?" This turn is not easy, because frontier labs are almost the only places with access to the largest-scale compute and the newest models.

One path is to move to research institutions and nonprofits whose organizational goal is safety. Such institutions continue to produce work on model evaluation, alignment research, and policy formulation, but their resources and data access are usually not comparable to those of frontier labs. Another path is returning to academia, at the cost of a significant drop in compute scale. A third path is rarer: entering regulatory or standards-setting bodies and translating technical understanding into the language of rules.

Each path means ceding some kind of capability. Bengio's article does not avoid this point, and his judgment is that such a trade-off is worthwhile at the present moment. Whether that conclusion holds depends on a deeper question: does the influence of safety research come more from access inside the model, or from public, independent, and credible evaluation conclusions?

Alignment research focuses on how to keep the behavior of AI systems consistent with human intent and values, including directions such as interpretability, reward modeling, and supervision methods. It is highly relevant to the discussion of "staying at frontier labs" because alignment experiments usually require access to the most powerful, most frontier models — and those models exist only inside a small number of frontier labs. External researchers often can only get interfaces to released models and have difficulty touching the training process and internal representations, which naturally limits independent evaluation and alignment research in terms of resources.

What This Debate Is Really Wagering

Bengio's article has drawn attention not only because it comes from a Turing Award winner, but because it pushes a question that had long remained in academic conferences and internal memos down to the level of individual career choice.

When top researchers begin to publicly discuss "leaving," the talent moat of frontier labs faces real pressure. The signal this sends to the entire industry may exceed the strength of the article's own argument — it means the tension between safety and capability has evolved from a dispute over technical路线 into a dispute over organizational loyalty.

For ordinary practitioners, this article does not necessarily provide a direct answer, but it does offer a clear way of asking the question: Is the position you occupy making the technology safer, or merely making it faster? There is no standard solution to this question, but the cost of avoiding it is rising.

Bengio's article was published in the technical publication Transformer, whose readership is highly concentrated among frontline AI research and engineering personnel, so its distribution path itself is "for insiders." What is called "voting with your feet" manifests in the talent market as the hiring flows and employer choices of top researchers, which directly affects labs' ability to attract safety talent. When talent supply begins to tilt toward non-frontier institutions, frontier labs must either adjust their safety commitments to retain people or bear long-term costs to their safety reputation — this is precisely the mechanism Bengio hopes to trigger.

One-line Summary

Bengio argues that safety-focused practitioners should leave frontier AI companies, pushing the tension between safety and capability down to the level of individual career choice.

Paragraph Summary

In an op-ed published in Transformer, Yoshua Bengio explicitly argues that practitioners who put safety first should leave frontier AI companies, describing it as an urgent personal choice rather than an industry debate. The article effectively questions the long-defaulted path of "staying inside labs to push for safety change," arguing that the capability race has compressed the time and resources needed for internal checks and balances. Bengio does not name any company; instead, he asks practitioners to examine the direction of their own work. Safety-first practitioners face real trade-offs: moving to safety research institutions, academia, or regulatory bodies all require ceding compute and model access. The deeper wager in this debate is that the tension between safety and capability is shifting from a dispute over technical路线 to a dispute over organizational loyalty.