Bengio urges frontier lab employees to leave for the sake of AI safety

Yoshua Bengio, one of the three "godfathers" of deep learning and a 2018 Turing Award winner, has published a column in Transformer with a direct call to action. He advises those for whom AI safety is a real priority to leave the frontier companies. The headline reads like an ultimatum, and the teaser names the audience: employees of the leading labs, whom Bengio asks to join the fight for a safer future. We do not have the full text of the column, only the headline and teaser. What follows is therefore context and the logic of the debate, not a summary of the author's arguments.

What we know about the statement

The published information adds up to the following picture:

• Author: Yoshua Bengio, professor at the Université de Montréal and one of the most cited scientists in computer science.
• Venue: Transformer News, a publication covering AI policy and the AI industry. The column was written exclusively for it.
• Audience: employees of frontier labs, meaning companies that train the most powerful models.
• Main thesis, going by the headline: prioritizing safety and working at such companies are poorly compatible.
• Publication date: October 8, 2026.

The headline alone does not tell us whether Bengio names specific companies or proposes alternatives, such as academia, nonprofits or government institutions. That is worth checking in the original.

Why his words carry more weight than usual

Bengio moved from pure research to a public stance on AI risk some time ago. In 2023 he signed an open letter calling for a pause in training the most powerful systems. He later chaired the international report on the safety of advanced AI, whose first edition came out in January 2025. In June 2025 he launched the nonprofit LawZero with funding of about $30 million. Its goal is to develop systems that do not act as autonomous agents and are therefore easier to control.

In other words, Bengio is not just commenting on the industry from the sidelines. He is building an alternative institutional home and, judging by the column, inviting people to join it. It is easy to read this as both an ideological position and a staffing agenda: talented safety specialists are scarce, and every one of them counts.

The "inside or outside" debate

This debate has a history. In 2021 a group of researchers left OpenAI and founded Anthropic, largely over disagreements about safety. In May 2024 Ilya Sutskever and Jan Leike left the company, and Leike wrote publicly that safety culture had taken a back seat to shiny products. In June 2024 current and former employees of OpenAI and Google DeepMind issued the "Right to Warn" letter, about the right to warn the public of risks. Geoffrey Hinton left Google in 2023 so he could speak freely.

The "stay and influence" position has a strong argument. Only inside a lab can you see the actual models, the evaluation results and the internal release decisions. Leaving costs a person access and leverage, and the replacement may care less about safety. Bengio's side argues the opposite: commercial pressure on a lab outweighs the influence of individual employees, and independent research, regulation and public oversight grow faster when the best specialists are not locked behind nondisclosure agreements.

Practical takeaways

For engineers and researchers, the column is a reason to assess honestly which safety decisions they can actually change in their current role, and whether they have any guarantee that they can speak publicly about risks. Lab leaders should check how their escalation channels, termination terms and whistleblower policies are set up. Otherwise, calls like this one will start to affect hiring and retention.

Investors and regulators should watch whether words are followed by a noticeable exodus of safety specialists. If one begins, it will be a stronger signal than any column. For now, it makes sense for readers to read the original in Transformer in full, to separate Bengio's arguments from the headline, which by its nature is blunter than the text.