An OpenAI Safety Writer Walks Away
For years, David Robinson was one of the people responsible for producing the safety documentation that accompanied OpenAI's biggest model launches. Those reports are meant to be a bridge between the lab and everyone outside it — a written account of what a new system can do, where its limits sit, and what safeguards are in place before it reaches users. This week, Robinson resigned from that role, and rather than leaving quietly he took his argument public in an editorial for The Atlantic.
His case is not that OpenAI needs another few pages of internal policy. It is that the culture running through the frontier AI industry is fundamentally broken — and that fixing it requires something far deeper than new rules or regulatory checklists bolted onto business as usual.
The Culture Problem Robinson Describes
Robinson points to a set of habits that have shaped Silicon Valley for a generation: what he characterizes as extreme confidence, a relentless cycle of sprints, and an unimpeded optimism that treats each new model as an unambiguous step forward. Within that environment, potential harms tend to be ignored, minimized, or assumed away rather than measured and mitigated.
That framing matters because it changes the shape of the debate. If the problem were a single bad decision or one under-resourced team, a targeted intervention might be enough. But if the underlying disposition of the industry is itself the issue, the remedy has to involve changing how these companies think about their own fallibility — something no external rulebook can simply mandate.
Robinson's prescription starts with humility. He argues that AI companies need to look past the insular, move-fast-and-break-things world they grew up in and borrow judgment from fields that have spent decades learning how to manage risk they cannot fully eliminate.
Nuclear Plants, Airports, and the Case for Redundancy
The most concrete part of Robinson's argument is his analogy for how frontier labs ought to operate. Given the scale of the risks involved, he contends, they should run less like startups chasing a launch date and more like nuclear power plants or busy airports.
Those are institutions society already trusts with catastrophic downside. They are designed around the assumption that humans will make mistakes, and they are built with layers of redundancy and deliberate, time-consuming planning so that an occasional and inevitable error does not open a door to disaster. The objective is not flawless performance from any single operator, but a system that absorbs mistakes before they cascade into something irreversible.
Translated to AI development, that logic implies slower release cycles, multiple independent lines of review, and a willingness to spend time that the industry currently treats as pure waste. It also implies that safety work is not a document produced at the end of a build, but a structural property of how the build is conducted from the start.
Why rules alone may not be enough
Robinson's central claim is that this is a deeper issue than simply slapping a few new regulations on how training runs are handled. Rules can constrain behavior at the margins, but they rarely change the incentives or the self-image of the people doing the work. A company that believes it is saving the world is unlikely to be slowed much by a compliance requirement it views as friction.
What he is describing, in effect, is a culture problem that has to be solved from the inside — by leaders who are willing to say out loud that they might be wrong, and by organizations structured so that saying so is not career suicide.
A Growing List of Departures
Robinson is the latest figure in what has become a recognizable pattern: researchers and safety staff leaving prominent AI firms and then speaking publicly about what they saw. The current wave appears to have been set in motion by Jacob Coxon, who resigned from Anthropic and warned that AI could kill us all by the end of the decade.
Others have followed:
- Robert O'Callahan, Bilal Chughtai, and Josh Engels, who all departed Google DeepMind.
- Joe Benton, who left Anthropic.
Each departure adds a data point to the argument that something systemic is pushing safety-minded people out of the labs. When the people hired to think about downside risk keep concluding that they cannot do that job effectively from the inside, it raises obvious questions about how much weight those concerns carry internally.
The Cynicism Problem
There is a fair objection to all of this. It is easy to feel cynical when people who helped build a technology turn around and warn about it — they contributed to the situation they are now describing, and their warnings arrive after the systems are already deployed.
But that reaction, however understandable, does not invalidate the substance of what they are saying. Insider knowledge is precisely what makes these accounts useful. The people who wrote the safety reports, ran the evaluations, and sat in the rooms where release decisions were made are better positioned than outside commentators to describe how those decisions actually get made.
What This Means for the Next Model Release
The practical stakes are straightforward. Every major model launch now arrives with documentation intended to reassure users, enterprises, and regulators. If the person who wrote that documentation concludes the culture producing it is broken, the credibility of the entire disclosure exercise comes into question.
Robinson's answer is not to abandon the work but to rebuild it on a more cautious foundation — one that treats human error as a certainty to be engineered around, rather than an exception to be explained away. Whether frontier labs adopt that posture voluntarily, or only after outside pressure forces it, is now the central question hanging over the industry's next round of releases.
For now, the warnings are accumulating faster than the structural changes. Robinson's resignation is the latest signal that the people closest to the technology are the least comfortable with the pace at which it is being shipped.
This article is based on reporting by The Verge. Read the original article.
Originally published on theverge.com
