Geoffrey Hinton requests that models of artificial intelligence be evaluated before they are made available, under the eyes of an independent authority. Resumed in the Lebanese press of 10 November, his speech raises a concrete question: who decides that a model can be disseminated, on what evidence and with what monitoring then? AI controls already exist in several forms, without constituting the international system which it calls for.
The researcher expresses this position in an interview with the British press published on 3 November. He criticizes the trade race between large technology companies and fears of critical infrastructure incidents. It nevertheless retains the hope of significant benefits. His speech advocates a framework for development, rather than a general renunciation of technology.
In the report published in Lebanon, Hinton proposes an international authority inspired by drug control. Companies should provide evidence to compare expected benefits and possible damages prior to marketing. This is a proposal from the researcher, not the announcement of an institution already created.
AI controls: what should be tested before access?
A successful demonstration of what a model can accomplish in a given situation. A safety assessment also seeks its failures, misuse and the conditions under which its protections cease to function. These objectives require different tests.
The risk management framework of the U.S. National Institute of Standards and Technology, AI RMF 1.0, includes pre-deployment testing and regular use. It organizes work around governance, risk identification, measurement and management. This framework is a management tool, not an international marketing authorisation.
Its logic requires clarification of the context. An error in a draft text and an erroneous action in an operational system do not have the same consequences. The question therefore goes beyond the score obtained by a model on an exercise: it also focuses on the people exposed, the tools available and the decisions that the system can make.
In concrete terms, a useful test should be used to determine which version was examined and under what conditions. A model limited to text production does not have the same possibilities as a software agent. The perimeter of access is thus part of Hinton’s problem.
Simulated Attacks to Search for Failures
The European Artificial Intelligence Regulation provides for a series of obligations for suppliers of general-purpose models with systemic risk. Article 55 requires assessments using protocols and tools corresponding to the state of art. It explicitly mentions adversarial tests, which seek to put the model in default.
The text also requires assessing and reducing systemic risks, documenting and reporting serious incidents without undue delay. Cybersecurity must cover the model and its physical infrastructure. These requirements concern a specific category of models: they do not mean that any application has an identical safety certificate.
The European approach thus gives concrete translation to several topics of the debate. The supplier must produce information on risks and corrective actions. Evaluation is not only a public assessment of the impressive nature of a system.
The European Code of Practice complements this scheme. Published in July 2025, it is presented by the Commission as a voluntary way to demonstrate compliance with legal obligations. Its chapters cover transparency, copyright, safety and security. The latter relates to the models covered by the systemic risk obligations.
Independence is not just on behalf of the controller
Hinton’s proposal places a separate authority from manufacturers at the centre of the decision. Its interest lies in the possible conflict between the speed of marketing and the time required for testing. A company can know its model better than anyone else while having an economic interest in broadcasting it quickly.
In order for external control to provide additional information, several questions need to be resolved: what version do evaluators access? What results can they examine? Can they replicate the tests? Who decides what happens next when a major failure appears? These questions describe the practical conditions of an examination, without presuming an answer for a particular company.
A comparison with drugs also has an obvious limit. A general model can be used for very different tasks and receive new tools after dissemination. Its assessment file must therefore specify the uses and conditions examined. A favourable appreciation within a given perimeter does not meet all possible uses.
Hinton’s personal criticism of some leaders is part of his judgment. They do not replace test results or evidence of particular misconduct. The debate on responsibility becomes more precise when it comes to documented decisions, access and incidents.
After launch, maintain the ability to act
The NIST Practical Guide requires monitoring of risks and benefits throughout the life of the system. It plans to examine whether the deployment should continue and adapt responses to the observed risks. Controls may include access, use limits, documentation and organisation of human work around the system.
This continuity responds to a concrete difficulty: actual uses may reveal problems absent from the initial tests. An update, a change in integration or a new user population changes the context. The monitoring must be able to link an incident to the version used and to the conditions of its operation.
Operational responsibility then begins with identifiable decisions. Who can limit a function? Who gets the alerts? Who checks the effectiveness of a correction? An evaluation report loses some of its usefulness if no one is responsible for acting on its results.
Hinton refers to possible disruptions of electricity networks, banking services or health systems. These scenarios express its concern; They do not constitute the announcement of further established attacks on these infrastructures. However, his appeal moves the debate towards a verifiable requirement: documenting capabilities and risks before opening, and then maintaining the means to correct, restrict or stop problematic use.
Comments