OpenAI safety chief quits, warns of broken industry culture

David Robinson, who wrote safety reports that accompanied every major model release at OpenAI, resigned last week and published an editorial in The Atlantic warning that the AI industry’s culture is broken, arguing that the sector’s relentless optimism ignores potential dangers.

Key points

• David Robinson resigned from OpenAI after writing safety reports for all major releases.
• He calls the sector’s culture “broken” and warns of unchecked optimism.
• Robinson urges nuclear‑grade safeguards, likening labs to power plants or airports.
• Other safety experts, including Jacob Coxon and Robert O’Callahan, have also left leading firms.
• The exodus signals growing unease among researchers about AI risk.

What happened

Robinson’s resignation came after months of internal frustration. He had been responsible for the safety briefings that OpenAI released alongside each new model, a role that grew in visibility as the company’s outputs expanded.

In a piece in The Atlantic, Robinson says the culture of Silicon Valley—characterized by extreme confidence, perpetual sprints, and unimpeded optimism—has created a blind spot toward potential harm. He writes that frontier labs should operate like nuclear power plants or busy airports, with layers of redundancy and careful, time‑consuming planning, so that the occasional and inevitable human error does not open a door to disaster. The report quotes David Robinson saying, "Given today’s risks, frontier labs need to run like nuclear power plants or busy airports, with layers of redundancy and careful, time‑consuming planning, so that the occasional and inevitable human error does not open a door to disaster."

Robinson is not alone. Jacob Coxon left Anthropic and warned that AI could kill us all by the end of the decade. He was followed by Robert O’Callahan, Bilal Chughtai, Josh Engels at Google DeepMind, and Joe Benton at Anthropic, all citing safety concerns as their reason for departure.

The backstory

Safety reporting at OpenAI began with the release of GPT‑3 in 2020, when the company began publishing risk assessments alongside model documentation. Over the years, the role expanded as new releases demanded more rigorous scrutiny, but the organization’s internal culture remained focused on rapid iteration and market advantage. This tension between speed and safety has fueled a growing chorus of voices calling for stricter oversight.

The exodus of safety professionals has accelerated in 2026, reflecting a broader trend of researchers leaving high‑profile AI labs. Their departures highlight the gap between corporate risk appetite and the technical realities of scaling advanced models.

Causes and cultural drivers

Robinson attributes the problem to a systemic overconfidence that rewards speed over safety. The current incentive model in tech rewards new features and larger models, not the laborious process of threat analysis. When safety concerns are treated as optional add‑ons, they become invisible to decision‑makers.

The culture of perpetual sprints creates a feedback loop: rapid releases generate revenue, which fuels further investment in faster iteration, leaving little room for long‑term risk mitigation. This dynamic also discourages whistleblowers, as the cost of speaking up can outweigh perceived benefits.

Trade‑offs and future risks

The trade‑off is clear: investing in redundant safety layers slows deployment and increases cost, but the alternative is catastrophic failure or regulatory crackdowns. Companies that choose speed may win short‑term market share, but risk long‑term damage to trust and legal standing.

Second‑order effects include erosion of public trust, tighter regulations, and potential blacklisting of firms that ignore safety. Competitors that adopt a more cautious approach could capture markets that value reliability over novelty.

What builders and businesses should do

• Conduct independent safety audits for each new model and publish findings publicly.
• Adopt a safety‑first governance framework that includes redundancy, time‑consuming planning, and human‑in‑the‑loop oversight.
• Create internal channels for engineers to raise risk concerns without retaliation.
• Benchmark safety practices against industry standards such as ISO 27001 or NIST SP 800‑53.
• Allocate a dedicated budget for long‑term risk mitigation rather than short‑term feature development.

The bottom line

Watch for regulatory announcements in the next six months, as lawmakers may codify the safeguards Robinson calls for. Companies that ignore the warning risk losing market share and facing costly remediation.