AI safety dissent must not be left to company promises

The policy also describes internal routes. Compliance may refer safety concerns to senior leadership, the Safety Advisory Group or the board’s Safety and Security Committee. The document describes this review as a channel for input, not a formal investigation. OpenAI’s policy sets out those review channels.

That distinction matters in a dispute over dismissals. OpenAI says an investigation found the researchers violated rules for handling sensitive information. The company’s stated rationale deserves a fair hearing, but the company also controls the process that judged its employees’ conduct. That is exactly where a review independent of ordinary management becomes essential.

OpenAI has framed the issue around trust. In a company statement reported by CBS News, it said, "We cannot do the work in front of us without a high degree of trust." Trust must run both ways. Researchers need confidence that raising a good-faith safety concern will not jeopardize their jobs or their ability to work with independent organizations.

The strongest case for OpenAI’s position is straightforward: research organizations need rules to protect sensitive information, and outside collaboration cannot be a license to disclose restricted material. OpenAI says it investigated and found policy violations, and denies that safety criticism was the reason for the dismissals. Enforcing legitimate information safeguards is not the same as punishing dissent. The way to distinguish the two is a credible review, not simply a contest between competing accounts.

Who can make AI safety oversight credible?

On Thursday, Korbak, Wang and Balesni said the dismissals and the company’s communications about them could make colleagues less willing to speak openly or work with independent safety organizations. They urged OpenAI to preserve outside safety collaboration and the ability to monitor frontier models. The issue is not only how OpenAI handles a personnel dispute; it is whether its rules leave room for the outside collaboration and model monitoring the researchers say safety work requires.

OpenAI says it encourages safety debate, tolerates good-faith mistakes and is engaging third-party safety assessors. The company has also said that keeping models monitorable requires an industry-wide commitment. Those positions create a testable standard: assessors need enough access to do meaningful work, independence from the company’s conclusions and a way to communicate findings. OpenAI should explain how its arrangements meet that standard.

In an AP-NORC survey of 2,140 adults conducted Sept. 24–28, 64% said AI was developing too quickly. That concern about AI’s pace makes safety oversight a public-interest issue, not merely an internal workplace dispute.

Transparency can protect confidential files and private employee details while still giving the public a meaningful account. OpenAI can explain the standards applied, how its process protects good-faith safety reporting and what role independent reviewers can play. It can also set out the access and reporting terms for outside assessors, so the public can judge whether their work is meaningful.

OpenAI’s board Safety and Security Committee should review the dismissal process and publish an explanation that protects sensitive information while clarifying how the company distinguishes safety reporting from policy violations. Congress should establish a binding baseline for good-faith AI safety reporting and independent scrutiny.