Former OpenAI safety employee says company culture has broken down
In short
David Robinson, a recently departed OpenAI employee, published a signed article in The Atlantic on Oct. 3 saying the company’s culture has broken down and that it relies too heavily on releasing systems first and fixing problems later. He said he helped draft OpenAI’s risk response framework and oversaw safety reports for 12 frontier model releases, and called on AI labs to adopt multiple safeguards like those in aviation or nuclear power. An OpenAI spokesperson said the company pauses training or delays releases when it needs to slow down.
Read the full story 2 min read
A recently departed OpenAI employee, David Robinson, published a signed article in The Atlantic on Oct. 3 saying the company’s safety culture has broken down, according to IT Home, CLS, Yicai and Jiemian. The Chinese outlets reported the article on Oct. 4. [ 1 , 2 , 3 , 4 ]
Robinson said he worked at OpenAI for more than three years and was among its longest-serving employees; IT Home put his tenure at three and a half years. He wrote that he helped draft the company’s risk response framework and oversaw safety evaluation reports for 12 frontier model releases, and that “the company keeps jumping from one product launch to the next” without the level of careful control he believed was needed. [ 1 , 2 , 3 , 4 ]
He wrote that “the era of trial and error is over”, and that frontier AI labs should follow aviation or nuclear power in adopting multiple safeguards and careful planning, so that an occasional and unavoidable human error does not become a disaster. [ 1 , 3 , 4 ]
IT Home, citing Reuters, reported that Robinson also warned that AI capabilities are developing faster than researchers’ understanding of alignment, the field that seeks to keep AI systems acting in line with human goals and values. It said he described OpenAI as relying heavily on “iterative deployment” — releasing systems first and hardening safety afterwards — and that his remarks intensified an industry debate over whether companies are moving too fast, after incidents of failed safety controls and abnormal behaviour in experimental systems that have brought OpenAI and its rival Anthropic under scrutiny. [ 1 ]
An OpenAI spokesperson said in a statement: “We have been controlling the model to avoid its capabilities exceeding what we can safely manage; when we need to slow down, we pause training, or hold back the release of a model.” Only IT Home carried the company’s response, and it does not address Robinson’s claim about the company’s culture. [ 1 ]
The outlets differ on one detail: IT Home said Robinson worked at OpenAI for three and a half years, while CLS, Yicai and Jiemian said more than three years. CLS, Yicai and Jiemian attributed their reports to Xinhua and CCTV News and did not include a company statement. [ 1 , 2 , 3 , 4 ]
Why it matters
The article feeds an industry debate over whether AI companies are moving too fast in building more capable systems, a debate the documents say has already put OpenAI and its rival Anthropic under scrutiny after earlier safety-control failures and abnormal behaviour in experimental systems. The documents describe no regulatory, legal or company action following the article, and give no indication of what OpenAI might change.
Key facts
- A recently departed OpenAI employee, David Robinson, published a signed article in The Atlantic on Oct. 3 saying the company’s culture has broken down, according to IT Home, CLS, Yicai and Jiemian. [ 1 , 2 , 3 , 4 ]
- Robinson said he worked at OpenAI for more than three years and was among its longest-serving employees; IT Home put his tenure at three and a half years. [ 1 , 2 , 3 , 4 ]
- Robinson said he helped draft the company’s risk response framework and oversaw safety evaluation reports for 12 frontier model releases. [ 1 , 2 , 3 , 4 ]
- He wrote that the era of trial and error is over and that frontier AI labs should follow aviation or nuclear power in adopting multiple safeguards and careful planning. [ 1 , 3 , 4 ]
- IT Home, citing Reuters, reported that Robinson warned AI capabilities are developing faster than researchers’ understanding of alignment, the effort to keep AI systems acting in line with human goals and values. [ 1 ]
- An OpenAI spokesperson said the company keeps models within what it can safely manage and pauses training or delays a model’s release when it needs to slow down. [ 1 ]
- IT Home said the remarks intensified an industry debate over the pace of development, after incidents of failed safety controls and abnormal behaviour in experimental systems that put OpenAI and Anthropic under scrutiny. [ 1 ]
Confirmed by several sources
- A recently departed OpenAI employee, David Robinson, published a signed article in The Atlantic on Oct. 3 criticising OpenAI’s approach to AI safety. [ 1 , 2 , 3 , 4 ]
- Robinson said he helped draft OpenAI’s risk response framework and oversaw safety evaluation reports for 12 frontier model releases. [ 1 , 2 , 3 , 4 ]
- Robinson said the era of trial and error is over and that frontier AI labs should adopt multiple safeguards of the kind used in aviation or nuclear power, so that an inevitable human error does not become a disaster. [ 1 , 3 , 4 ]
Still unclear
- Robinson’s exact length of service at OpenAI. IT Home said he worked there for three and a half years, while CLS, Yicai and Jiemian said more than three years and described him as one of its longest-serving employees; no document gives exact dates.
- Robinson’s job title and the date he left OpenAI. IT Home described him as a former employee in the safety department; no document states his title or departure date.
- The earlier incidents of failed safety controls and abnormal behaviour in experimental systems that IT Home referred to. Only IT Home mentions them, without details, dates or confirmation from other outlets.
- Whether OpenAI disputes Robinson’s characterisation of its culture. Only IT Home carries a company statement, and it addresses model release practices rather than responding directly to the culture claim.
What local media are saying
Timeline, local time
- David Robinson publishes a signed article in The Atlantic titled “I quit my job at OpenAI because the corporate culture is broken”, criticising the company’s safety practices. No publication time is given. [ 1 , 2 , 3 , 4 ]
- IT Home reports the article, citing Reuters, and includes a statement from an OpenAI spokesperson. [ 1 ]
- CLS reports the departing employee’s criticism of OpenAI’s risk awareness. [ 2 ]
- Yicai reports the story, attributing it to Xinhua. [ 3 ]
- Jiemian reports the story, attributing it to CCTV News. [ 4 ]