Three OpenAI researchers were dismissed this week after speaking out about the company’s approach to AI safety, with the fired employees publicly stating they were let go for “prioritising safety” over the company’s commercial and deployment priorities. The BBC reported that the researchers said OpenAI had punished them for being outspoken about safety concerns, while the company disputed the characterisation and said they were dismissed for mishandling sensitive information. NPR confirmed that the three fired workers maintained OpenAI’s stated commitment to safety was increasingly subordinated to competitive pressure from Google’s Gemini and Anthropic’s Claude model lines. The firings come at a moment when OpenAI is preparing for what is expected to be one of the largest initial public offerings in technology history, placing intense commercial pressure on the company’s leadership to demonstrate revenue growth and deployment scale rather than precautionary restraint.
The received wisdom
The progressive and tech-critical response frames these firings as a confirmation of long-held fears about the AI industry’s governance structure. OpenAI began as a non-profit with an explicit safety-first mission; it has progressively restructured itself toward a for-profit model under commercial pressure, and each restructuring has diluted the original governance safeguards. The firing of safety researchers is, in this reading, the predictable endpoint of that trajectory: when commercial imperative and safety caution conflict, commercial imperative wins. The solution proposed by this school of thought is external regulation — a federal AI safety board, mandatory pre-deployment testing requirements, whistleblower protections for AI researchers, and potentially mandatory licensing for large frontier models. The European Union’s AI Act, already in partial implementation, is held up as a model. The argument concludes that the AI industry cannot self-regulate its way to safety any more than the financial industry could self-regulate its way out of the 2008 crisis.
This is a coherent and in many respects empirically well-grounded critique. The revolving door between OpenAI’s safety team and its deployment teams has been documented. The pattern of safety concerns being raised internally, noted, and then overridden is not new.
A different read
But the regulatory remedy being proposed deserves rigorous scrutiny before we embrace it. The call for a federal AI safety board sounds sensible until you ask what, specifically, it would be empowered to do — and who would staff it. The EU AI Act, held up as the global gold standard, has already been criticised by AI researchers across the political spectrum for being written primarily by regulators with limited technical expertise, creating compliance burdens that fall most heavily on smaller entrants while entrenching the incumbents (including OpenAI, Google, and Anthropic) who can afford dedicated regulatory teams. The paradox of poorly designed AI regulation is that it produces the opposite of its intended effect: it consolidates power among the largest and least safety-conscious actors while suppressing the diverse ecosystem of smaller, more transparent models.
There is also a deeper question about what “AI safety” actually means in practice. The fired researchers’ concerns — at least as reported — appear to relate to deployment velocity and internal culture, not to any specific identified catastrophic risk. The BBC reported that OpenAI disputes that safety was the actual cause of the firings. The he-said-she-said quality of the dispute points to a structural problem: there is no independent body with the technical expertise and legal standing to adjudicate whether a company’s safety practices meet a meaningful standard. The regulatory gap is real. But the solution is not obviously a federal board staffed by political appointees; it might be something more like the model of independent auditing that applies to public companies’ financial statements — rigorous, expert, independent, and with genuine legal consequences for misrepresentation.
Meanwhile, OpenAI is not the only actor in this space. The BBC reported separately that Anthropic has introduced usage policies banning users from being “cruel” to its AI systems — a development that raises its own questions about where the line between genuine safety research and corporate virtue signalling lies. The AI safety discourse has a tendency to conflate several distinct concerns: catastrophic long-term risk from superhuman systems, near-term harm from biased or manipulative outputs, corporate governance failures in deployment decisions, and symbolic anxieties about the moral status of AI itself. Conflating these produces policy responses that are confused about what they are actually trying to prevent.
The governance vacuum is real. The correct response is to build oversight structures that are technically literate, politically insulated, and narrowly focused on measurable harms — not to create the next iteration of security theatre that makes regulators feel responsible without actually making AI systems safer.
What to watch
Watch OpenAI’s IPO prospectus when it is filed — the risk disclosures around safety culture and employee relations will be legally significant and will reveal how the company characterises these firings to investors. Watch whether any congressional hearing addresses these dismissals specifically; several senators have been building AI oversight agendas. Watch Anthropic’s and Google’s recruitment from OpenAI’s safety team — the departure of credentialed safety researchers from the market leader benefits competitors and may accelerate the inter-company dynamics that produce genuinely problematic deployment decisions. And watch whether the fired researchers pursue legal action; whistleblower law in the AI context is an entirely untested area.
--- J