What should AI have to demonstrate before public release?

Watching a documentary about artificial intelligence recently left me thinking about responsibility.

I use AI regularly. I build tools with it, and I have just started an MSc in AI and UX. I can see why people are excited about its potential, and I share that excitement.

I also want to know who decides when a powerful system is ready to be released—and what evidence they need before making that decision.

I do not support a blanket ban on superintelligence. I support continued development with independent scrutiny, enforceable safeguards and deployment decisions that can change when new evidence emerges.

That position comes with a responsibility of its own: being willing to accept “not yet” when the risks cannot be adequately controlled.

Testing is already happening

Independent AI evaluation is not a new idea.

The UK AI Security Institute has described its work testing models before and after deployment. METR, an independent research organisation, evaluates frontier AI capabilities and develops scientific methods for assessing risks associated with autonomous systems. These organisations are already working on parts of the problem.

But testing needs careful interpretation. AISI has cautioned that independent evaluations cannot yet provide confident certification that a system is safe. A test provides evidence about particular capabilities and vulnerabilities, under particular conditions. It cannot answer every question about what will happen afterwards.

For me, that makes the decisions surrounding testing just as important as the tests themselves.

What happens when a concerning result appears? Who sees it? Who can require changes? Who can say that release must wait?

What I would like to see

These are principles I would support, rather than a description of a complete system already in place.

• Independent scrutiny proportionate to risk. The strongest requirements should apply where capabilities, access or intended uses could cause serious harm. Evaluators need enough access, time and independence to challenge a developer’s claims. A small, low-risk tool should not automatically face the same requirements as a highly autonomous system.
• Clear conditions for deployment. Before release, there should be an explicit account of the intended use, foreseeable risks, safeguards and remaining uncertainties. Decision-makers should explain why the evidence supports that particular deployment.
• Staged release where appropriate. Limited access and monitored trials can help gather evidence before wider use. However, a small trial is not automatically safe: some risks may be unacceptable even with few users, and some releases cannot easily be reversed.
• Continuing responsibility after launch. Monitoring, incident reporting and reassessment should continue as systems change. Organisations should have a practical response when safeguards fail, including restricting access or withdrawing a system where necessary.
• Public accountability. People affected by AI should have ways to raise concerns and challenge consequential decisions. Transparency should make scrutiny possible without publishing sensitive information that enables misuse.

None of this requires a promise of zero risk. It requires justified decisions, clear responsibility and a willingness to act on uncomfortable evidence.

The human part needs testing too

My psychology background makes me particularly interested in the phrase “human oversight”.

Imagine someone reviewing an AI-generated recommendation. They see a polished explanation and an approval button. There is technically a person involved.

But what would we need to know before calling that an effective safeguard?

Can they check the evidence? Do they understand the system’s limitations? Do they have enough time to review it? Can they reject its recommendation without being penalised? Can they stop an action before it causes harm?

These are testable questions. Research could help us understand whether particular interface designs make oversight more useful, while still recognising that success in a controlled study would not establish safety in healthcare, public services or other consequential settings.

There is room to contribute

I am at the beginning of my formal AI studies. I do not have a finished regulatory blueprint, and I do not think enthusiasm alone qualifies anyone to certify a powerful system as safe.

What I can do is ask specific questions, learn from existing research and look for a manageable contribution. I have contacted my MP about safeguards and the Ada Lovelace Institute about opportunities to participate.

I want AI’s benefits to be available widely. I also want the people developing, deploying and governing it to be accountable for the risks they ask others to bear.

Those questions deserve clear answers.