How AI catastrophe warnings can divert attention from corporate accountability

That is not evidence that the reviews are worthless. It is a reason to resist treating the word external as a substitute for an answer. A review commissioned by a company could still uncover serious problems. Whether outsiders can verify its findings or require a response is a separate question.

Trump praised the AI industry and stressed maintaining a U.S. lead in the technology. The voluntary accord lets his administration and participating companies announce safety commitments together. But a commitment, by itself, does not tell the public who decides when a system is safe enough or who can challenge that decision.

Gebru’s concern is about that division of power. If executives define the risks, select the tests and decide what to disclose, they retain considerable control over the terms of scrutiny. That is a risk to examine, not a finding about how every part of the accord operates. Until its review and disclosure arrangements are clear, its accountability value cannot be judged from the announcement alone.

Can catastrophe warnings crowd out present-day accountability?

In a Democracy Now! interview published Thursday, Gebru described discussion of hypothetical superintelligent machines as “a ploy for regulatory capture.” That is her argument about who gets to shape oversight, not an established finding that companies that signed the accord have captured regulators.

The strongest objection to her view is straightforward: a powerful future system could pose risks that deserve serious attention before they arrive. Testing advanced AI and bringing in reviewers may help identify dangers that a company misses. Gebru’s critique does not settle the broader debate over catastrophic AI risk, and present-day accountability is not a reason to ignore it.

The problem arises if a forecast of extraordinary future harm becomes the main yardstick for safety. A company can warn about what a later generation of AI might do while leaving less dramatic questions about current systems unanswered. Reliability in consequential uses, checks before a machine-assisted finding informs a decision and responsibility for failures all warrant scrutiny.

Gebru argues that existing laws could provide routes for holding companies responsible and that cybersecurity practices warrant scrutiny. The interview identifies no court ruling establishing liability for a particular AI failure, and the accord is not shown to displace those laws. Her narrower point survives those limits: promises to test a product should not be confused with a public determination of responsibility when something goes wrong.

Company assessments could identify problems early, and outside assessments might expose what internal tests miss. Their potential value, however, is not the same as a public right to verify or challenge their conclusions. People who depend on, work with or face decisions informed by AI systems have an interest in whether anyone can contest the judgments companies make about safety.

What does an AI-assisted intelligence error show?

The accountability question becomes less abstract in CNN’s account of a military intelligence assessment prepared with chatbot assistance. The assessment misidentified the cargo of a Chinese vessel and prompted preparations to intercept it. The chatbot was not identified; the account does not establish that its developer was responsible for the error.

CNN did not report that an interception occurred or that a chatbot acted independently. It described a mistaken assessment entering a consequential decision process. Responsibility in such a case cannot be assigned simply by pointing at a chatbot. The questions run through the chain of use: who chose the tool, what information it supplied, how the report was checked and who had authority to act on it.

Those distinctions matter because an error can expose weaknesses at more than one point. A developer’s testing, an organization’s decision to use a tool and a human reviewer’s checks are different matters. Public scrutiny would require enough information to distinguish among them rather than assuming the technology or its maker explains every failure.

The incident also gives substance to Gebru’s challenge. A pledge to review AI systems may be relevant, but it does not answer whether a particular deployment was appropriate or whether people caught a bad result before it shaped a decision. Oversight has to reach the use of a system, not just the promises made about it.

Who gets a voice beyond company leadership?

The Right Livelihood Foundation named Gebru a 2026 award recipient on Wednesday, citing her work challenging concentrated power in AI. Her response to that concentration is not merely to ask CEOs for better answers. She describes the Distributed Artificial Intelligence Research Institute, which she founded, as bringing technologists together with refugee advocates, labor organizers and artists to help shape technology and build grassroots power.

Her distinction gets at something an external review alone may not provide: a role for people outside an executive circle in deciding which risks deserve attention in the first place.

That does not mean every participant would decide every technical question. It means workers and affected communities could help identify consequences that a company’s chosen tests overlook, while independent researchers could assess claims the public is otherwise asked to accept on trust. For those contributions to matter, they would need access to relevant information and a way to press for action.

The AP account does not establish whether the accord provides for such participation. The unresolved questions are practical: who can see review findings, who can dispute them and who can demand a response. If those powers stay with company leaders, the leaders gain the benefit of a public safety commitment while retaining control over its meaning.