The AI revolution is inherently human – AI regulation should be too

We’ve all seen the pictures – blue, glowing brains, white robots and disembodied motherboards – that usually accompany pieces on how AI is going to transform health care. And we’ve read the stories – about keeping up with the ‘AI revolution’, getting ahead of the curve and ‘harnessing its potential’.

While the evidence base for these claims is growing, what AI advocates often miss is that the AI ‘revolution’ is not driven by technology alone; it is inherently human. From the data about people that AI is built on to the data labellers and engineers who build it, the commissioners who buy it, the clinicians who test it and the patients who experience it, AI is not a floating brain but a tool situated in a deeply complex social system that is shaped by people.

And so, the future of AI is not inevitable. It is something we can collectively steer, and in which we all have a say. Nowhere is this more important than in health, which is why the MHRA established a national commission last year to consider how the regulatory framework needs to evolve. As research partner to the commission, we, in partnership with Ipsos, have worked with the UK public to hear and amplify their voices and understand what it will take to build and retain trust in a new regulatory approach. 

We conducted deliberative research with 78 people from across England and Wales. Across a series of workshops, they were given balanced information about AI and asked to grapple with some of the difficult trade-offs its use creates. We heard their excitement for AI’s potential – such as reducing the NHS’ admin burden or detecting subtle abnormalities in scan images – but also their fears. While individuals often held conflicting views – frequently being nervous or scared while also impressed or enthusiastic – as conversations progressed, a clear set of principles and priorities emerged.

Perhaps unsurprisingly, we heard just how important it was for regulators, and regulation, to prioritise the accuracy of AI tools, which, given enough time in the real world, many expected to be better than a relevant clinician. While many accepted that AI – like humans – may never be perfect, they felt the bar needs to be high to justify its introduction into the NHS.

Those we spoke to also placed a premium on humans remaining involved in AI-informed decisions and were even willing to sacrifice many of the efficiency gains offered by AI if it meant a human could stay ‘in the loop’. But the value of human oversight wasn’t just to reassure the public about the safety of an AI tool. For many, human involvement was also critical to ensuring accountability for when things don’t go as planned. Because while human error may be an accepted facet of what makes us human, findings suggested that machine error could be harder to accept, and people wanted to be sure that if there was a problem, a human would ultimately be held accountable. 

The same concerns arose when participants considered whether AI will work equally well for everyone. They understood that a system might perform well overall while still being less accurate for some groups. But participants wanted to be sure that no one would receive worse care as a result, and that any inequalities in performance should be a priority for improvement. They wanted regulation to ask who benefits, who might be disadvantaged and what happens when problems emerge in practice.

This is not to say their views were not pragmatic. From the information provided, many recognised that we may only really understand whether an AI tool works – and for whom – once it is being used in the real world. Because of this, participants were generally supportive of a risk-based approach to regulation: more stringent checks for higher-risk tools, while allowing less onerous upfront checks for lower-risk uses, provided there was careful monitoring and safeguards in place if problems emerged. 

But, of course, this approach will only work if the NHS has the infrastructure and capability to test and monitor AI tools in real-world use, something we are currently exploring. For those we spoke to, regulation was not about stopping innovation but about making sure the level of scrutiny matched the potential consequences if things went wrong.

Ultimately, we heard a call for regulation that could keep pace with the technology, learn from what happens in practice and give people confidence that the benefits of AI were being realised without losing sight of what matters to them. That means grounding regulation in the experiences, priorities and expectations of the people who will ultimately live with the consequences of these decisions.

AI may change what is possible, but that does not mean it can tell us what is acceptable: that is a human question. 

This article was originally published by HSJ on 10 September 2026.