Former OpenAI employee David Robinson, who worked at the company for 3.5 years, said artificial intelligence developers are moving too fast and do not pay as much attention to safety as necessary. He took part in building OpenAI's risk assessment system and led the preparation of safety reports for 12 launches of frontier models. Now Robinson says that "the time for trial and error has passed." This was reported by Qazaqyia.kz citing Kursiv Media.

The Atlantic article

Robinson published an article in The Atlantic titled "Quit OpenAI Because Its Culture Is Broken." According to him, the problem is not only in the company's specific decisions, but in the very culture of development, where new systems must be released faster and faster.

He criticizes OpenAI's approach, which the company calls iterative deployment. The idea is to release models, observe their behavior, find problems and then strengthen safeguards. Robinson believes this principle is becoming less and less acceptable as AI capabilities grow.

"The time for trial and error has passed," Robinson wrote.

The example of aviation and nuclear power

In his view, when developing more powerful systems, far stricter safety measures must be built in from the start. As examples, he cites aviation and nuclear power, where potentially dangerous technologies must not be tested on the principle of "let's try first, fix later."

Robinson's main complaint is that AI capabilities are developing faster than researchers can sort out the problem of alignment – matching model behavior to human goals and values.

He believes companies are not cautious enough and should rely more heavily on safety and research specialists before moving to create even more powerful systems. At the same time, Robinson does not claim that artificial intelligence is already out of control. His warning concerns primarily how fast the technology is developing and how well we understand the associated risks.

OpenAI's response

The company rejected the notion that safety remains unaddressed. An OpenAI representative told Reuters that the company ensures its models do not become more capable than it can safely control and protect.

It also stressed that OpenAI can stop training or delay the release of a model if there are grounds for doing so.

And this is not just a statement for the press. OpenAI does indeed use the Preparedness Framework – a system for assessing the most serious risks associated with frontier models. The company states that it checks model capabilities before release, assesses potential harm and develops safeguards. In 2026, OpenAI also published a separate Frontier Governance Framework, which includes risk assessment, reporting, safety measures and incident response.

Not yet

Not yet. At least, no such conclusion can be drawn from the Robinson story.

There is a former employee who worked on the safety of frontier models and who considers the current pace of development dangerous. There is OpenAI, which claims it has risk assessment procedures and can stop development or release when necessary. And there is a broader debate within the industry about how fast companies should move toward increasingly capable systems.

But there is a genuinely unpleasant point in this story. Robinson was not an outside observer: he took part in building OpenAI's safety system and preparing reports on model launches. Therefore his words cannot be taken as proof that the company has lost control over AI, but simply dismissing them as "a former employee is unhappy" would be too frivolous.

For now the question is not "will AI take over humanity?" A far more practical one is different: will developers manage to create sufficient safety measures before the capabilities of the systems themselves become substantially higher?

Earlier, Kursiv wrote that OpenAI's AI agents tried to bypass the protection of more than 100 websites of third-party organizations.