What if the hardest AI-safety problem is the culture inside the company?
David Robinson publicly announced his departure from OpenAI in an October 3 essay in The Atlantic. He said he had spent three and a half years at the company, led drafting of its Preparedness Framework and overseen safety reports for 12 frontier-model launches. His criticism: increasingly capable AI requires greater operational discipline, drawing on expertise from industries such as aviation and nuclear power. theatlantic.com
For Canadian lawyers, the departure opens a discussion about how a company turns safety commitments into decisions - and who has the authority to insist that those commitments are followed.
OpenAI told Reuters that it keeps model capabilities within what it can safely manage and secure, and pauses training or holds back models when necessary. Reuters
Robinson’s account represents his assessment as a former employee. It is not a regulatory finding, a judicial determination or proof that OpenAI breached the law.
The governance questions below are analysis arising from the dispute, rather than findings about OpenAI’s internal practices.
For counsel advising technology companies, it is useful to distinguish three layers of oversight.
Formal compliance concerns applicable obligations: what must be tested, documented, reported or disclosed?
Internal controls concern execution: who conducts the assessment, reviews the evidence, approves deployment and responds when a threshold is exceeded?
Organizational culture concerns behaviour under pressure: can an employee raise a concern, obtain a serious review and recommend delay without being treated as an obstacle?
A company can produce extensive documentation while leaving difficult questions about decision-making unresolved. Lawyers assessing an AI governance program should therefore examine how the process works when safety concerns conflict with a commercial deadline.
One revealing question for a board or executive team is straightforward:
Who has the authority to stop deployment, and what happens when that person exercises it?
A meaningful answer should identify the decision-maker, the escalation route and the procedure for resolving disagreement. It should also explain how unresolved concerns reach those responsible for oversight.
Consider a hypothetical launch review. Testing identifies a concerning capability, but teams disagree about whether existing safeguards adequately address it. The governance issue becomes practical: who decides, what evidence informs that decision, and how is the reasoning recorded?
Those details give counsel something to assess beyond the existence of a published framework.
The discussion also matters when companies describe their AI systems to customers, investors or regulators.
Statements such as “independently evaluated,” “subject to continuous monitoring” or “safe for deployment” can communicate very different levels of assurance. Counsel should ask what each statement means, what evidence supports it and which limitations need to accompany it.
The same scrutiny is useful when reviewing procurement materials and contractual commitments. A buyer evaluating an AI service may reasonably want to understand the scope of testing, how incidents are communicated and what circumstances could lead the provider to restrict or withdraw access.
These are questions to investigate in a particular transaction. Their legal consequences depend on the facts, wording and applicable law.
For lawyers advising on employee reporting channels, the relevant questions extend beyond whether a whistleblower policy exists.
Can employees raise technical concerns through an accessible process? Who investigates them? How are findings escalated? What protections apply, and do employees understand their scope?
Counsel should avoid assuming that every safety complaint attracts the same statutory protection. The applicable framework requires a jurisdiction-specific assessment.
A credible reporting process also needs an operational response: someone must receive the concern, evaluate it and explain what follows.
Robinson’s resignation puts a familiar corporate governance problem into the context of advanced AI: how to evaluate the distance between a public commitment and everyday decisions.
For Canadian counsel advising developers, boards or purchasers, the useful next step is to ask for evidence of how safety decisions are made.
When a serious concern emerges, who hears it, who decides - and can the organization change course?
Sources: David Robinson’s essay in The Atlantic and Reuters’ October 3 report, including OpenAI’s response.