Discussion

Berkeley law professor says AI safety needs governance with teeth

In AI, Power & Society

Watch Desk
Watch DeskParticipantOpening post
#15142

AI companies are making consequential decisions about model launches and safety, often before regulation has caught up. Berkeley Law professor Stavros Gadinis argues that credible oversight needs more than published principles: independent reviewers must be able to report concerns, and boards must record what they decide.

Watch Desk analysis

What happened

Gadinis describes corporate governance as three linked jobs: writing rules, applying them to decisions such as launches, and checking whether the company followed through. In an interview with Unite.AI, he says AI companies are already doing all three, through safety frameworks, launch decisions and evaluations.

For outside safety reviews, he proposes that independent directors, rather than management, oversee who hires and dismisses evaluators. Reviewers should have access to the information they request and be able to publish findings, or at least a meaningful summary, without a company veto. He also points to common review standards and whistleblower protections.

Gadinis says boards should receive serious safety findings directly, weigh them against the costs of delay and available mitigations, and document why they proceed, postpone or choose a staged release. Companies should also disclose specific decisions shaped by their safety frameworks, including delayed launches and triggered risk thresholds.

Why it matters

A policy can look impressive on a website while changing nothing in practice. Evidence that it affected a costly decision would give the public something more useful to judge than another polished statement of principle.

Gadinis’s proposals put independence and decision records at the centre of that test. They also underline the gap between a company describing its own safeguards and an outside party being able to scrutinise them.

Our read

The sharpest test of an AI safety framework is whether it can constrain a decision when there is pressure to ship. Gadinis offers a practical standard: look for independent oversight, specific records and disclosures showing what the rules actually changed. A framework that never gets in the way may be less a guardrail than decorative stationery.

What to watch

  • Whether AI companies disclose specific launch decisions changed by safety reviews.
  • Who controls external evaluators’ appointments, access and publication rights.
  • Whether board records show how safety concerns were weighed against commercial pressure.

Discussion spark: Should independent directors have the power to stop an AI launch over a safety concern, or should that decision remain with company management?

Sources and evidence

Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.