Britain’s AI Security Institute says it is not seeking mandatory requirements for frontier AI developers, while polling cited by Resultsense found 82% of UK adults surveyed backed mandatory safety testing. That gap puts a practical question at the centre of AI policy: who gets to decide whether a powerful model is ready to launch?
UK AI Security Institute Watch analysis
What happened
Resultsense’s analysis compares the UK’s voluntary approach with a US executive order that also invites developers to submit models for review rather than requiring government clearance before release. The article says the US order allows voluntary reviews of up to 30 days before launch and rules out interpreting it as authority for mandatory licensing or pre-clearance.
The polling comes from the Ada Lovelace Institute’s Great Expectations briefing, based on 1,928 UK adults surveyed in September 2025. Resultsense says 89% agreed AI products should be held back until proven safe, even if progress slowed, and 82% backed mandatory testing, compared with 33% comfortable with voluntary measures. The AISI, meanwhile, says it has independently evaluated more than 55 models before release, while describing its aim as best practice that could inform formal standards later.
Why it matters
The difference is not simply between testing and no testing. It is between a system where developers choose whether to take part and one where a required check could delay release. Voluntary reviews may build useful practice, but they leave the timetable and participation largely in developers’ hands.
The polling records public preference, not a legal mandate or proof that mandatory testing would make models safer. Still, the mismatch matters for anyone deciding how much weight to put on a safety process that companies can choose to join.
Our read
The key issue is not whether pre-release testing happens. The AISI says it has evaluated dozens of models. It is whether participation should remain voluntary when the public appears to want a firmer gate. That is a policy choice, not a technical detail to be tidied away in the footnotes.
What to watch
- Whether the AISI turns its best-practice work into formal standards, and whether those standards require participation.
- Whether the US review framework attracts broad developer participation and publishes enough detail to judge its reach.
- Whether public support for mandatory testing becomes a concrete UK policy proposal.
Discussion spark: Should frontier AI developers be required to pass safety checks before release, or is voluntary testing the better way to improve practice without slowing progress?
Sources and evidence
- Gary Marcus's AI 'Vietnam' moment ended in a voluntary test. Britain's public wants a mandatory one – Resultsense (26 October 2026, 08:56 UTC)
not affiliated with or endorsed by UK AI Security Institute