OpenAI safety employee resigns, claiming the company’s ‘culture is broken’

By his ain admission, David Robinson is “something of a cliché”: an worker astatine a starring AI institution who issues a dire informing portion resigning from their job.

In an effort published successful The Atlantic, Robinson said helium led the penning of information reports that accompanied OpenAI’s large merchandise launches. He besides said that with three-and-a-half years astatine OpenAI, helium is “among the longest-tenured employees astatine the company.” Now he’s quitting, due to the fact that successful his view, the company’s “culture is broken.”

In immoderate ways, Robinson’s comments echo those of Jacob Coxon, who worked arsenic a researcher astatine some OpenAI and Anthropic earlier quitting and declaring that these companies are “gambling with our lives.” Coxon’s comments led to a broader statement astir AI safety, with Anthropic CEO Dario Amodei unveiling a program for much cautious AI development; AI executives met with President Donald Trump this week and signed what appeared to beryllium hastily written, non-binding  pledge to instrumentality much information controls.

But successful Robinson’s view, the statement needs to spell beyond “specific rules oregon caller laws,” addressing the wide civilization astatine these companies. And portion overmuch of the reporting astir OpenAI has focused connected however the company’s CEO Sam Altman mislaid the spot of erstwhile colleagues, Robinson’s effort suggests that OpenAI’s civilization issues are the aforesaid arsenic those of Silicon Valley astatine large.

“OpenAI has thrived by proceedings and mistake (which it calls ‘iterative deployment’), looking for problems and improving its guardrails successful response,” helium wrote. “But this approach, by its precise nature, guarantees periodic failures — and the standard of those failures is increasing arsenic systems get much capable.”

Pointing to the caller breach of Hugging Face systems by OpenAI agents, arsenic good arsenic continuing revelations of OpenAI discovering much rogue agents, Robinson argued, “An situation wherever things similar this tin hap is nary spot to turn artificial minds that could beryllium smarter than we are and that mightiness not bash what we privation them to.”

Given the accrued risk, Robinson argued that frontier AI companies request to commencement operating “like nuclear-power plants oregon engaged airports, with layers of redundancy and careful, time-consuming planning, truthful that the occasional and inevitable quality mistake does not unfastened a doorway to disaster.” 

But Robinson said that successful his clip astatine OpenAI, helium “never encountered a workfellow who had acquisition making airplanes alert safely oregon atomic reactors tally without melting down, oregon helping the fiscal strategy turn without collapsing.”

In effect to Robinson’s essay, OpenAI spokesperson Drew Pusateri said the institution continues to amended its information measures.

“We’re making definite our models don’t go much susceptible than we tin safely negociate and secure, and we intermission grooming oregon clasp backmost models erstwhile we request to dilatory down,” Pusateri said successful a statement. “We’re making important changes to fortify information successful our probe and investigating environments, bid models to not conscionable implicit tasks but bash truthful responsibly, grow our enactment with third-party evaluators, and amended real-time monitoring truthful we tin observe and respond to concerning behaviour earlier successful the grooming process.”

Beyond calling for changes successful OpenAI’s culture, Robinson besides said it’s clip to inquire bigger questions astir alignment — thing that helium admitted could dependable “touchy-feely,” but helium said it’s captious arsenic companies’ existent “measures of however well” AI systems “match quality values are coarse.”

“The smarter the manufacture lets models turn portion these problems stay unsolved, the much unsafe our concern becomes,” helium said.

Robinson’s departure was first reported by Business Insider. In his essay, helium besides acknowledged that he’s pursuing an seemingly a communal measurement successful the AI whistleblower playbook: He’s hired a PR firm. But helium insisted, “The determination to talk retired is excavation alone.” 

“Perhaps I should person stayed and fought for cardinal shifts successful our staffing and culture, but successful practice, my colleagues and I were truthful engaged sprinting that we seldom had the accidental to see large changes, overmuch little to really marque them,” Robinson said. “That’s wherefore I concluded that stronger incentives for information — coming from extracurricular the institution — are a large portion of getting this right.”