More Than 800 Figures Urge a Halt to Superintelligence

More than 800 scientists, policymakers, business leaders and cultural figures had signed a call on October 22, 2025 to prohibit the development of superintelligence until there was scientific consensus about safety and control, together with strong public support. Signatories included Geoffrey Hinton, Yoshua Bengio and Steve Wozniak. The historical figure comes from a snapshot published that day: the online list remains open and its counter changes, so today’s total cannot reconstruct the launch on its own.

The statement is short; turning it into enforceable policy would not be. A signature expresses support for two conditions, but does not define which activities would be prohibited, how the superintelligence threshold would be recognized, who would measure consensus, what public support would mean or how compliance would be verified. Separating a political demand from its mechanism lets readers evaluate this and future AI governance proposals without dismissing them as vague or accepting them as though they were already law.

What the statement actually requests

The primary text calls for a prohibition on superintelligence development that would not be lifted before two requirements were met: broad scientific consensus that development could be safe and controllable, and strong public buy-in. It does not request a six-month pause or set a restart date. Nor does it claim that superintelligence already exists. Its introduction describes such a system as AI able to significantly outperform all humans on essentially all cognitive tasks.

Choosing a condition instead of a date changes the kind of proposal. A temporary pause ends even if the technical problem remains open. A conditional prohibition remains until evidence justifies lifting it. Yet that rhetorical strength creates a design obligation: words such as “safe,” “controllable,” “consensus” and “buy-in” must become observable criteria. If every lab can interpret them for itself, the condition stops nothing; if nobody can ever demonstrate them, the prohibition becomes permanent by construction.

First question: what is being regulated?

“Developing superintelligence” might mean training a model, scaling an existing one, researching algorithms, publishing weights, deploying an agent or connecting a system to tools. Each verb reaches different actors and moments. Prohibiting only deployment allows training and internal testing to continue; prohibiting basic research may capture work that also improves limited and useful systems. A rule must describe conduct, not only a feared outcome.

It must also distinguish a model from a system. A model in isolation may produce text; the surrounding system adds memory, browsing, code, permissions and infrastructure access. A dangerous capability can emerge from that combination even if the model does not change. A definition based only on a product name or exam score therefore becomes obsolete when the same component receives new tools.

Second question: where is the threshold?

Superintelligence is a hypothetical, comparative category: it requires a choice of which humans, which tasks and which performance level form the reference. A system may beat experts on a narrow test while failing at prolonged planning, physical interaction or novel situations. Averaging those capabilities hides very different profiles. A useful threshold combines capability, reliability and autonomy in specified scenarios with the severity of possible harm.

The 2025 International AI Safety Report, written to synthesize scientific evidence, focused on general-purpose AI and emphasized the substantial uncertainty surrounding the pace of future progress. It did not recommend a particular policy. That separation is instructive: scientific analysis can describe capabilities, risks and disagreements; deciding how much uncertainty justifies a prohibition is a public choice.

Existing rules show how proxies are used when capability cannot be measured directly. The European Union’s AI Act classifies general-purpose AI models with systemic risk through high-impact capabilities and presumes that level above a training-compute threshold, while allowing the Commission to designate other models using criteria such as autonomy, scale, tools or user numbers. Compute is verifiable, but it is a proxy rather than a definition of intelligence.

Third question: who verifies compliance, and with what access?

A prohibition works only if an authority can know what is being trained. It would require records of large computing runs, model documentation, independent evaluations and incident-reporting duties. Evaluators would need enough access to test dangerous behavior without receiving unnecessary trade secrets. Data centers, cloud providers and distributed systems would also have to be covered, because watching only famous developers moves the work to another entity.

Verification cannot rest on a demonstration prepared by the developer. It should include adversarial tests, unseen situations, evaluations of deception and autonomy, and monitoring after deployment. No single exam certifies lasting control: a system may pass one battery and behave differently after receiving tools, instructions or new contexts. Policy needs a sequence of evidence and the ability to suspend activity when new information appears.

Fourth question: how are consensus and public support measured?

Scientific consensus is not unanimity or the number of signatures on a letter. A policy must name an institution, a procedure for selecting experts, conflict-of-interest rules, published evidence and an agreement threshold. It should also separate two questions: whether a system is technically controllable and whether the remaining risk is acceptable. The first admits technical tests; the second includes values and the distribution of harms and benefits.

The call was accompanied by a survey commissioned by the Future of Life Institute of 2,000 U.S. adults between September 29 and October 5, 2025. Sixty-four percent said superhuman AI should not be developed until it was proved safe and controllable, or should never be developed; 5% supported moving as quickly as possible. Participants received short definitions before answering.

Those results measure a national sample under a particular wording, not “the opinion of humanity.” Strong public buy-in would require choices about territories, representation, prior information and the form of decision. A survey is evidence of preferences; a referendum, parliamentary law and deliberative consultation carry different forms of legitimacy. Quoting the percentage without the question, sample and geography would turn a useful measurement into a mandate broader than the evidence supports.

Fifth question: what triggers and lifts the prohibition?

The proposal needs a trigger that occurs before completed superintelligence. Waiting to confirm that a system exceeds all humans is too late to prevent its creation. Capability milestones, compute scales or combinations of autonomy and resource access could be used, subject to periodic review. Every indicator will produce false positives and false negatives; acknowledging them allows the regime to adjust without pretending that a perfect boundary exists.

Exit needs as much precision as entry. “Safe and controllable” could be translated into quantified limits, replicable audits, interruption mechanisms, cybersecurity, traceability and legal responsibility. A disagreement rule is also necessary: who resolves conflicting evidence and what happens while it is reviewed? If lifting the prohibition depends on a general feeling, the decision returns to those who already control development.

What a signature proves—and what governing still requires

The diversity of more than 800 signatories shows that concern crossed professional and ideological lines. It does not show that everyone shares the same probability of harm, definition of superintelligence or preferred legal instrument. A short statement creates a coalition precisely because it leaves those differences unresolved. Its function is to establish a principle; the next task is designing rules that can be audited.

The transferable skill is to subject any AI proposal to five tests: regulated object, threshold, verification, authority, and conditions for entry and exit. If one is missing, that does not make the risk imaginary; it means the document is still a political demand, not an executable mechanism. The distinction makes it possible to demand prudence about uncertain capabilities and precision from those who promise to govern them.