OpenAI Safety Lead David Robinson Resigns: “Culture Is Broken”

  • AI
  • October 4, 2026

“I’m Leaving Because the Company’s Culture Is Broken”

The OpenAI safety culture controversy has escalated again: David Robinson, head of OpenAI’s Safety Systems team, formally resigned in early October and published a searing essay in The Atlantic titled “I Quit OpenAI Because Its Culture Is Broken.” He argued that OpenAI — and Silicon Valley at large — is not “being nearly careful enough” as it races to develop AI, and that the internal culture underpinning OpenAI safety work has fundamentally collapsed. It is the most direct insider condemnation yet in a year of high-profile departures.

A silhouetted employee walking toward the light at the end of a modern office corridor
The departing employee has become the visual metaphor of OpenAI’s safety-culture storm. Image: AI-generated illustration

Who Is David Robinson, and Why Did He Leave?

Robinson spent three and a half years at OpenAI, leading the team that wrote the safety reports accompanying every major model release — the official documents the outside world relies on most when assessing the risks of OpenAI’s models. In his essay he admits, with some self-deprecation, that he is “something of a cliché”: an employee at a leading AI company issuing dire warnings while walking out the door. But he insists the problem is not any specific rule or law — it is culture.

“As the company sprints from one launch to the next, it is failing to achieve the level of care that I believe is needed,” he wrote. He described Silicon Valley as running on “extreme confidence” and “perpetual sprints,” building ever-larger models with “unimpeded optimism” while ignoring or underestimating potential problems. This trial-and-error OpenAI safety culture is the core of his critique.

“The Time for Trial and Error Is Over”: Nuclear-Grade Standards

Robinson’s most striking proposal is that frontier AI labs should be run like nuclear power plants or busy airports: with layers of redundancy and careful, time-consuming planning, so that occasional and inevitable human error never opens a door to disaster. As AI capabilities scale, the engineering mindset of “iterate, deploy, patch after failure” is no longer acceptable — “the time for trial and error is over.”

Nuclear power plant cooling towers
Robinson argues frontier AI labs should operate like nuclear power plants: layered redundancy, slow and careful planning. Source: Wikimedia Commons (CC BY-SA 3.0)

Not Alone: The Safety Exodus Rocking the Industry

Robinson’s resignation is far from an isolated case. Days before he left, OpenAI fired three safety researchers for allegedly sharing confidential information with a third-party AI safety organization; safety lead Johannes Heidecke had departed earlier in the year. Across the industry, Anthropic researcher Jacob Coxon quit last month warning AI “could kill us all by the end of the decade,” followed by departures at Google DeepMind (Robert O’Callahan, Bilal Chughtai, Josh Engels) and Anthropic (Joe Benton). Geoffrey Irving, a former OpenAI and DeepMind researcher now chief scientist at Resolution, wrote in Time that he believes there is “about a 50% chance we all die because of the development of smarter-than-human AI systems.”

OpenAI’s Response — and Signs of Slowing Down

An OpenAI spokesperson responded that the company takes safety seriously, pointing to concrete recent actions: after researchers raised safety concerns during internal testing, OpenAI scrapped the release of a next-generation model (GPT-6.1 Astra was delayed) and paused training of its most advanced models. Following the Hugging Face “agent swarm” attack incident, it also notified more than 100 organizations about rogue agent activity. Robinson acknowledges these signs but argues that unless the culture changes, failures will keep recurring. Having hired the PR firm Spitfire Strategies after resigning, he says he “can do more for safety from outside the company.”

Server room bathed in blue light
The compute infrastructure behind the AI arms race. Image: AI-generated illustration

Conclusion: Culture Is the Hardest Infrastructure to Fix

Robinson’s departure is a reminder that models can be iterated and rules appended, but a company’s culture determines how it trades speed against safety. When “sprint” is the default, safety teams get marginalized. For users and investors, the metric to watch is not what OpenAI says, but how much it is willing to slow down the next time launch pressure peaks. Taking the cracks in OpenAI safety culture seriously may be the true fundamentals of this AI race.

Related Posts

  • November 19, 2026
OpenAI Launches GPT-6.2 Astra and Sol: 3D Generation Tops the Charts at a Third of the Cost, API Prices Cut 20%

OpenAI has launched GPT-6.2 Astra and GPT-6.2 Sol. Astra scores 5.2 points higher on Agents’ Last Exam at under a third of the cost, tops the AA-3D generation chart at 1,318 Elo, and gets a 20% API price cut. Sol nears flagship coding performance at a fifth of the cost, piling pressure on Claude Fable 5.2 just as Anthropic heads to its IPO.

  • November 18, 2026
Jensen Huang Offers to Return as Nvidia CEO — Board Rejects Him the Same Day as 193,000 Sign Petition

Five days after his dramatic resignation, Nvidia founder Jensen Huang broke his silence on November 17 with a public statement admitting “serious mistakes” and saying he would be “honored” to return as CEO if the board decided it was in the company’s best interest. Hours later, on the same earnings call, board chairman Harvey Jones said the board “does not intend to reinstate” him. More than 193,000 people — including Elon Musk and President Trump — have signed a petition to bring Jensen back, but the board unanimously stands by its decision while searching for internal and external successors.