AI in health technology assessment: what stays human

AI in health technology assessment is no longer a question of whether the field will change, but what will be left for people to do. At a HITAP Open Talk on 9 September 2026, Associate Professor Alec Morton of the Saw Swee Hock School of Public Health at the National University of Singapore walked through how AI is reshaping health economics, from building economic models to changing the skills researchers need.

Morton directs the Centre for Population Health Risk Informatics and Simulation Modelling (PHRISM) and the School’s MSc in Health Economics and Outcomes Research. He opened with a disclosure: AI had helped prepare his slides, but “the overall line of argument is my own.”

AI in health technology assessment: from science fiction to an everyday tool

Morton’s first theme was the speed of change. He pointed to the arrival of semi-autonomous agents, progress on benchmarks including “Humanity’s Last Exam,” and a widely shared figure from Anthropic showing its engineers committing eight times as many lines of code since coding agents appeared.

He admitted the pace had surprised him:

“I didn’t expect to see this sort of tech in my lifetime. It still feels like science fiction. But nevertheless, it’s here and we have to figure out what we are going to do with it.”

He illustrated the change with his own teaching: a year ago, an image generator asked to show his “wardrobe as a teacher in Singapore” produced a generic academic wardrobe full of heavy jackets and ties, unsuitable for the tropics. Asked again in 2026, the output looked like his actual bedroom. The lesson he draws for students is unchanged, though. AI produces convincing content, and you still have to bring your own contextual knowledge to judge it.

Morton also flagged what is getting harder. He warned:

“The days when you could be pretty sure that the AI would produce something which looked a bit correct, a bit wrong. Those days seem to be passing.”

Errors are now harder to spot, he said, and the trajectory matters more than today’s capabilities.

The economic model itself

Morton’s second theme was modelling. Early conversations about AI and HTA centred on evidence synthesis; attention is now shifting to building the economic models themselves. He cited a blog post by Ian Cromwell describing experiments with a “Claude skill,” an encapsulated body of practice knowledge the model can call on. Morton compared the concept to Neo uploading kung fu in The Matrix. Cromwell validated the approach by having Claude reproduce published models and checking that its outputs matched.

General-purpose frontier models are still “too random” for the job, Morton said, with variable expectations around modelling. He expects specialised products to fill the gap, naming Luminina AI as one company already supplying industrial-strength economic models for HTA processes. The open questions, he argued, are how these models get validated and how assurance and trust are built around them.

His own tests show both the promise and the limits. When he asked Claude to build a Singapore-specific model of lupus, it produced a spreadsheet with a dashboard, documentation of its input sources, a single cohort run through lupus episodes, and QALY and cost calculations over 20 cycles, complete with a tornado-style sensitivity analysis and a CHEERS 2022 checklist. He was impressed, but clear it would not pass muster for a dossier: it models a single cohort rather than an age distribution, treats lupus progression with a Markov model when the time-to-event is likely fat-tailed, and is built on an American example patched with a few Singapore parameters because it has no access to a Singapore data lake.

What stays human in health technology assessment

So what survives automation? Morton’s answer, drawing on recent discussion in the field, is that the middle of the work, the spreadsheet-building and coding, is becoming automatable. The human work sits at both ends. At the front: framing the decision problem, structuring the model, and supplying inputs that are not publicly available. At the back: validating the model, understanding it, explaining its uncertainty, and using it in the real world.

That last part, he argued, is a growth area for modellers. HTA models are built to have real-world purpose, and they need someone to explain them to a drug advisory committee, to focus attention on the assumptions that drive a cost-effectiveness result, such as treatment waning, and to support price negotiations. Morton drew on John Sterman’s Business Dynamics to make the point that model development sits inside a larger cycle where the model is used to learn about the real world and then refined.

He also described models as coordination devices, using the economist Thomas Schelling’s concept of a focal point. Two strangers told only to meet in Paris at noon will tend to pick the Eiffel Tower, not because it is optimal but because it is obvious. HTA models play a similar role, he argued, giving stakeholders a shared reference point. The 3% discount rate, he noted, is not a representation of anything in the real world but a convention everyone agrees to use, much like Moore’s Law became a self-fulfilling roadmap for the semiconductor industry.

More health economics, not less

On the question of whether AI will shrink the field, Morton reached for the Jevons paradox. More efficient coal-fired steam engines in the nineteenth century did not reduce coal consumption; they expanded it by making the engines useful across more industries. He expects something similar for health economics: cheaper, faster analysis will create more demand, not less.

He offered a personal analogy from Singapore Post. When the internet arrived, the total volume of post stayed roughly the same while the mix changed, personal mail disappeared and business mail grew. The lesson, he said, is that agents able to do much of the analytic work will not necessarily shrink the field. “We all want to do more health economics,” he said. “We all want to do better health economics.”

The opportunities are concrete. Guidelines for major conditions such as depression are updated only every few years while the technology keeps moving. Personalised recommendations, closer scrutiny of variation in practice, and modelling for secondary markets that big pharma treats as an afterthought are all work there has never been capacity to do.

Risks, and who gets a model

Morton was candid about the risks. His synthesis of the concerns in recent position statements from NICE and the International Mathematical Union came down to three losses: loss of human control, loss of human accountability, and loss of engagement. He described a “dead maths” worry circulating in the mathematics community, where there is a growing supply of proofs but fewer people who understand them.

He quoted a warning from the AI researcher François Chollet, which he paraphrased as: the time to learn to think for yourself is before GenAI, and if you miss that chance, good luck. Morton called it “a little bit too strong, but there is still something important there.“

He ended on a note of democratisation. If analysis becomes cheap, why should patient representative groups not build their own models? He asked:

“What if patient representatives are empowered to build their own models?”

Patient groups have long argued that lived experience gets left out of HTA models because it is hard to parameterise. Cheaper modelling could let them express what matters to them in the language a committee accepts. In low- and middle-income countries, though, Morton cautioned that the binding constraint is not the analysis but data access, which AI cannot solve on its own.

The skill to build now is critique

Asked what HTA researchers should prioritise, especially in low- and middle-income settings, Morton’s answer was blunt. He said:

“The big skill to prioritize is critique.”

Critique normally develops late in a career, he noted, when people become reviewers, editors, or managers. With agents now doing the drafting, everyone needs it early. His practical advice: when students drop a study question into an AI, do it together, sitting around one computer. Individually, it is easy to accept output that looks authoritative and comprehensive; in a group, someone will spot the question that leads into real scrutiny. Once you push back, he said, the AI will admit what it got wrong, and it stops feeling like a superintelligence and starts feeling like a conversational partner.

He still thinks researchers should build at least one model themselves, not as a core job for years, but to maximise the reflection that lets them demand more from the machine.

Genius and original sin

On whether AI would push health economics to crowd out the social, ethical, and institutional dimensions of HTA, Morton said the danger already exists. Cost-effectiveness analysis, he said, is valued precisely because it sends a clear yes or no signal. He put it this way:

“The genius of cost effectiveness analysis… it’s also the original sin of cost effectiveness analysis that it collapses everything down.”

The committees then discuss equity and ethics and, at the end, ask whether it is cost effective. His concern is that ever more convincing AI-generated economic reasoning could blow away the other considerations. But he drew a line: as long as decisions are reserved for people, the question is about how human intelligence evolves alongside the artificial kind.

He also flagged a specific risk for policy makers.

“Policy makers may decide that they are going to rely on AI and bypass their human experts.”

Human experts are slow, want more budget, and sometimes question whether the question makes sense. The AI, by contrast, will return something plausible but possibly misleading and based on far less depth. The education needed, he argued, is not teaching policy makers to run models but persuading them to hand the question to people who can do a properly validated study.

For a field still deciding what AI in health technology assessment means for its future, Morton’s message was more practical than alarmist. He closed by framing the choice as tool, threat, or transformation, and landing on a bit of all three. What excites him is the chance to finally do the work the field has always said it wanted to do but never had the hours for.

Source: HITAP Open Talk, “AI and the future of HTA,” 9 September 2026. Organised by the Health Intervention and Technology Assessment Program Foundation (HITAP).