Three researchers fired by OpenAI have written to the company’s board and its safety committee with a warning: as the industry races to build more capable systems, it is in danger of losing the ability to see how those systems think. The letter, disclosed by The Wall Street Journal on Oct. 7, asks the company to work with outside safety auditors and to preserve visibility into the reasoning behind its models.
The researchers argue that certain paths of development would make it harder to monitor what a model is doing, and they want the company to refrain from advancing down those paths. Their core concern is that capability and oversight are moving in opposite directions, and that oversight is losing.
The letter lands at an awkward moment for OpenAI, which has spent the past year describing itself as committed to safety even as it races rivals toward larger and more capable models. The signatories are former employees, and their warning carries the particular weight of people who once worked on the inside.
What the researchers are asking for is specific. They want OpenAI to cooperate with external auditors who can examine its systems from the outside, rather than relying on internal reviews alone. They also want the company to keep the reasoning process of its models legible, so that researchers and regulators can trace how a model reached a decision.
That demand speaks to a shift in how the field has evolved. The models that have attracted the most attention recently are those that reason through a problem step by step before answering. That reasoning is valuable, but it has also become more opaque, and the researchers fear the industry is trading away the ability to audit its own systems in exchange for performance.
OpenAI has not said how it will respond to the letter, and the company has generally resisted calls to open its most advanced systems to outside inspection, arguing that doing so risks security. The letter frames that caution as a retreat, a willingness to accept less transparency in the name of competitive advantage.
The dispute is part of a larger argument running through the AI industry about who gets to check the work of the companies building the technology. Regulators have pressed for audits and disclosure. Companies have pushed back, warning that revealing too much hands a roadmap to rivals and to bad actors.
The researchers’ position is that the balance has tipped too far toward speed. They contend that if the industry keeps building systems whose internal reasoning cannot be examined, it will reach a point where the people running the systems no longer understand them well enough to keep them safe.
Their letter is a bid to change that course before the industry arrives there. By asking the board and safety committee directly, the researchers are appealing to the parts of OpenAI whose formal responsibility is oversight, and testing whether those bodies will act independently of the company’s commercial goals.
The concern the researchers raise is not hypothetical to them. Models that reason in long chains of internal steps produce answers whose path is hard to reconstruct after the fact, and the more capable those models become, the harder the reconstruction gets. Once the trail is gone, they argue, neither the company nor a regulator can say with confidence why a system did what it did, or whether it can be trusted to do it again.
OpenAI’s safety committee has been restructured more than once as the company has grown, and its independence from management has been a subject of dispute inside the company. A letter addressed to that committee is, among other things, a test of whether it can still act as a check rather than a formality.
The three are not the first employees to leave OpenAI over safety concerns, and the company has weathered a series of such departures while continuing to grow. Each departure raises the same question: whether the people who know the most about a system are the ones who raise the sharpest warnings, and whether anyone is listening.
The letter’s timing matters. OpenAI is pushing forward on a range of products while telling the public that safety is its top priority. A letter from inside the company, delivered to its own oversight bodies, is the kind of document that forces the company to answer the question in public rather than in private.
What the researchers want is not a halt to development. They want development to proceed on terms that keep the systems legible, so that safety can keep pace with capability. The warning is that the industry is not doing that, and that the gap is widening.
The outcome of the letter will say something about how much the company’s stated commitment to safety constrains its actual choices. Boards receive warnings all the time. What matters is whether this one changes anything.
