Independent evaluation of frontier models

You must now ask your AI providers directly about pre-deployment testing, who conducted independent evaluation, and how they will report serious incidents — because a political declaration of support does not alter your contractual position or operational obligations.

An analysis of 2 October 2026 of mandatory pre-deployment tests and independent evaluation of frontier AI models argues that decision-makers should not wait for international enforcement mechanisms to emerge, but should instead establish their own verification practices now. Finland and Norway presented the call on 21 September 2026, and by the end of that month leaders from 29 countries had joined. In our assessment, this political coalition draws international attention to frontier AI control but leaves open how that control is enforced in practice — and the gap between political commitment and operational certainty is where your responsibility lies.

What does the declaration actually require?

The call is a political declaration, not a treaty or binding law. It sets out three governance layers for frontier models — generally understood as highly capable AI systems — but establishes no enforcement regime and no sanctions for non-compliance. The declaration itself does not change what your providers are contractually obliged to do, nor does it create a new supervisory body with immediate authority. Support remains open and voluntary; countries can join later, and absence from the published list of signatories says nothing about a nation's AI policy.

Which concrete controls do you need to establish now?

Because the call places independent evaluation at its centre but no international verification body yet exists, you cannot wait for external oversight to materialise. You should now establish the following:

1. Document which model each workflow uses and the lawful basis for the data it processes — map your sensitive AI deployments against existing standards and reporting channels that already apply to them.
2. Require your provider to name the evaluator and demonstrate their independence — set this out in your processor agreements and procurement terms before deployment.
3. Establish internal escalation rules for capability or safety thresholds — designate a named responsible person and a fixed reporting route for serious incidents.
4. Obtain unresolved evaluation findings before deploying the model for sensitive work — do not treat a provider's assurance of testing as sufficient; require sight of what the independent evaluator actually found.
5. Define per workflow which human approval is required — document the decision points where a person, not the model, must take responsibility.

Why does evaluator independence matter most?

The declaration identifies independent evaluation as the mechanism through which frontier models remain under human control. This means the independence of the evaluator is not a procedural detail but the operational point that matters for procurement officers and legal teams. If the evaluator is chosen, funded or directed by the provider, independence is compromised. You should therefore require your provider to name the evaluator, explain how that evaluator's independence is assured, and share findings — including unresolved concerns — before you deploy the model for work that touches sensitive capabilities or safety thresholds.

What happens if you treat the declaration as a guarantee?

The risk of confusing a political declaration with an operational guarantee is that you defer responsibility that remains yours. No international body currently exists to verify that providers have carried out adequate testing or that evaluations were genuinely independent. The declaration asks states to report serious AI incidents, but no channel yet exists for that reporting. This creates a gap between what the declaration aspires to and what is enforceable today. You cannot close that gap by waiting; you must close it by setting your own verification practices now.

Tooling can help you track which models are deployed where, which standards apply, and which evaluations have been completed. What tooling cannot do is determine whether an evaluator is truly independent, whether the tests were adequate for your use case, or whether the person approving deployment has understood the risks. That judgement remains yours, regardless of which countries have signed the declaration.