Former OpenAI safety writer says the company culture is broken as departures mount

David Robinson, a former OpenAI safety writer who helped prepare risk reports for major model launches, has resigned and is publicly warning that the company’s culture has drifted into something he says is unsafe. His criticism matters because it adds to a growing pattern of departures from major AI labs by people who were once responsible for evaluating the risks of ever-more powerful systems.

Robinson’s argument is not simply that one company needs stricter rules. He says the entire frontier AI industry has embraced speed, scale, and confidence at the expense of caution, and that the sector now needs the kind of disciplined safety engineering used in nuclear power and aviation. His exit, and the message that followed, sharpen fresh questions about whether leading AI companies are moving faster than their own safeguards can keep up.

Why Robinson’s resignation is resonating

Robinson is not an outside critic. He worked inside OpenAI on the safety reporting process that accompanied major model releases, which gives his warning extra weight for investors, policymakers, and employees across the AI sector. When someone with direct experience in the launch process says the system is broken, it suggests the concerns are about operations and organizational culture, not just public relations.

His departure also lands at a moment when the AI industry is under intense scrutiny over model capability jumps, competitive pressure, and the pace of deployment. The concern is not merely that models may produce errors or bias. It is that companies building increasingly powerful systems may be operating with a mindset better suited to consumer software than to technologies whose failures could have broad social consequences.

Robinson argued that the frontier AI sector has been shaped by “extreme confidence” and a relentless race to ship bigger models, leaving too little room for humility, redundancy, and deliberate planning.

That warning reflects a wider unease inside the AI field: the more capable systems become, the less forgiving the consequences of mistakes may be. A missed edge case in a chatbot is one thing. A missed safety failure in a widely deployed frontier model is something else entirely.

What did David Robinson say about OpenAI and the industry?

Robinson says the problem goes beyond a few weak policies or incomplete checklists. In his view, the culture surrounding frontier AI development has been shaped by a “move fast” mentality that assumes progress itself is the main objective, while danger is treated as a problem to be patched later.

He describes the prevailing approach in Silicon Valley as one marked by endless urgency and optimism, where teams push ahead with ambitious releases and often underestimate the likelihood of serious mistakes. That, he says, creates a setting in which the most important risks are not just technical but institutional: how decisions are made, how dissent is handled, and how much power safety teams actually have when deadlines loom.

His prescription is stark. Frontier AI organizations, he argues, should resemble highly regulated high-stakes operations rather than normal software startups. In practice, that means more checks, more separation of duties, slower release cycles, and a design philosophy that assumes human error will happen and must be contained.

How does he think AI labs should operate?

He believes the right model is closer to a nuclear facility or a major airport than a traditional tech company. In that framing, the goal is not to eliminate all risk — that is impossible — but to create layered protections so that one mistake does not cascade into catastrophe.

That comparison matters because it highlights a structural conflict at the heart of AI development. The industry is rewarded for speed, product launches, and market dominance. Safety engineering, by contrast, tends to reward patience, redundancy, and the willingness to delay a release when uncertainty remains.

• Redundancy: multiple independent checks before a launch.
• Planning: slower, more formal review of worst-case scenarios.
• Humility: acknowledging that model behavior can surprise even experienced teams.
• Containment: designing systems so a single failure does not become a wider incident.

Why are so many AI safety workers leaving?

Robinson is the latest in a series of departures that has drawn attention across the AI world. The pattern suggests that concern about frontier AI risk is not confined to academics or external critics; it is increasingly being voiced by people who worked closest to the systems themselves.

Some former employees have gone public with especially dire warnings. Others have simply stepped away, signaling discomfort with the direction the industry has taken. Together, these exits create a picture of internal tension between safety-minded staff and organizations under pressure to ship more capable models in a highly competitive market.

The names associated with these exits span several major labs, underscoring that the issue is not limited to one company. Anthropic and Google DeepMind have each seen departures from researchers or safety personnel who later spoke critically about the risks of advanced AI.

Who else has raised alarms after leaving major labs?

Among the most prominent examples is Jacob Coxon, who left Anthropic and then warned publicly that AI systems could pose catastrophic risks before the end of the decade. At Google DeepMind, Robert O’Callahan, Bilal Chughtai, and Josh Engels have all departed, while Joe Benton left Anthropic as well.

These exits do not prove that a single consensus exists inside the industry. But they do show that some of the people who understood the systems best have become less willing to stay quiet about what they see as dangerous tradeoffs.

What makes frontier AI safety different from ordinary product risk?

Frontier AI safety is different because the systems are not just feature updates or incremental tools. They are large-scale models trained with enormous compute and deployed into products that can influence search, coding, education, customer support, and other parts of daily life at global scale.

As model capability rises, so does uncertainty about how systems will behave in unfamiliar settings. That uncertainty makes pre-release evaluation harder and increases the value of slow, conservative decision-making. A company can recover from a buggy app. It may not be able to easily recover from a badly understood capability release in a frontier model that is quickly integrated across products and platforms.

The tension is also organizational. Safety teams often need independence to stop or delay releases when they detect unresolved problems. If the broader company culture treats those pauses as obstacles rather than necessities, safety review can become performative instead of protective.

How do safety reports fit into model launches?

Safety reports are meant to provide a structured assessment of what a model can do, where it may fail, and what precautions are still needed before release. In principle, they are the internal equivalent of a pre-flight or pre-plant inspection: a disciplined checkpoint before something widely used goes live.

Robinson’s role in writing those reports is important because it means he was part of the process designed to catch problems before they reached users. His decision to leave suggests that, in his view, the process is no longer robust enough to match the risks involved.

How the AI industry’s culture became part of the story

The debate is increasingly about culture as much as technology. For years, Silicon Valley has celebrated speed, experimentation, and aggressive execution. That approach helped create a wave of consumer products and cloud services, but critics say it does not translate well to frontier AI, where mistakes can scale quickly and where even developers may not fully understand a model’s emergent behavior.

Robinson’s critique suggests that the industry may have inherited a startup ethos that values momentum above restraint. In the past, that attitude was often defended as necessary for innovation. Now, opponents argue, it is becoming a liability because the systems being built are too consequential for improvisation.

That shift in language is notable. The issue is no longer framed simply as “Can we build this?” but also as “Should we release it now, and under what controls?” Those are fundamentally different questions, and the second one is becoming harder for many companies to answer convincingly.

Robinson urged AI companies to adopt the mindset of high-reliability industries, where the possibility of human error is assumed rather than ignored.

Why this warning lands at a sensitive moment

The timing matters because the frontier AI race is accelerating. Companies continue to spend heavily on model training, infrastructure, and product integration, even as governments and researchers debate what guardrails should exist. The result is a widening gap between technological progress and the systems meant to supervise it.

That gap helps explain why internal dissent attracts so much attention. When a former safety worker says the culture is broken, readers hear more than one person’s frustration. They hear a signal that the institutions responsible for managing high-risk technology may not be keeping pace with the technology itself.

For OpenAI specifically, the resignation adds another layer to its already closely watched reputation. The company has repeatedly argued that it takes safety seriously, but statements from departing insiders can complicate that message, especially when they describe a culture that privileges speed over caution.

For the broader sector, the story may intensify pressure on companies to show not just that they have safety teams, but that those teams can actually slow or reshape decisions when necessary. In a competitive market, that may be one of the hardest promises to keep.

Timeline of the escalating safety concern

The recent wave of departures helps explain why Robinson’s warning is getting attention now. The following timeline summarizes the broader sequence of events that put frontier AI safety back in the spotlight.

What happens next?

Robinson’s resignation will not by itself change how frontier AI is built, but it may influence how the public, regulators, and even employees interpret future safety claims. Once a former insider says the culture is broken, companies may face a higher burden to prove that their safety systems are more than symbolic.

For OpenAI and its rivals, the practical question is whether they can convince talent inside their organizations that caution will be rewarded rather than sidelined. If they cannot, more departures are likely. And if the people responsible for evaluating risk keep leaving, the credibility of the industry’s safety promises will only grow weaker.

For now, Robinson’s message is simple: frontier AI is too important to be run like a typical startup. He believes the sector needs slower, more disciplined processes before its speed outruns its safeguards.

That warning is likely to keep echoing because it comes from inside the system it criticizes. In AI, that may be the most unsettling kind of alarm of all.

Key facts at a glance

• David Robinson has resigned from OpenAI after working on safety reports for major model releases.
• He says the company culture, and the broader AI industry culture, is fundamentally broken.
• Robinson argues frontier AI labs should follow high-reliability models like nuclear plants and airports.
• His exit joins a growing list of departures from OpenAI, Anthropic, and Google DeepMind.
• The warning adds pressure on AI firms to demonstrate that safety reviews can actually restrain launches.