TL;DR

Researchers at Nanyang Technological University surveyed 122 people building AI for healthcare across Singapore, China, Hong Kong and Britain, and found most accept they should answer for what they ship — while barely more than half could name a single regulatory framework. Two-thirds work for employers that have adopted none. The findings appear in npj Digital Medicine.

The mismatch is the story. A workforce willing to shoulder liability without knowing the standards it is liable against is a peculiar kind of exposure, and it sits directly underneath tools now used to read scans and estimate disease risk.

Awareness of frameworks such as the EU AI Act or Singapore’s healthcare AI guidelines reached 57% of those asked. Seniority helped, and so did working outside academia: senior developers and those in commercial settings knew of more frameworks, and knew them better, than junior or academic colleagues. Employer behaviour mattered most of all. Where an organisation had formally adopted a framework, its developers understood the rules; where it had not — the majority — they did not.

Wilson Goh, who co-led the work as an assistant professor at NTU’s medical school and serves as chief data scientist at its Centre of AI in Medicine, made the case that developers are the people best placed to judge data quality, model limitations, bias and hallucination risk. His concern is that the willingness to take responsibility is not matched by knowledge of what responsibility entails.

For UK readers, Britain’s inclusion in the sample is the useful part — this is not solely an Asia-Pacific finding. It also arrives at an awkward moment. Separate polling published this week put AI use among NHS clinicians at 90%, with two-thirds of those staff saying they worked out their own practice before any employer guidance existed. Uptake at the bedside and rule-awareness in the development shop are moving at different speeds, in the same direction, with nothing joining them up.

The researchers’ recommendations are structural rather than punitive: build regulatory training into developer education, use senior-to-junior mentoring inside companies to navigate a genuinely complicated landscape, and have national regulators work towards harmonising rules across borders instead of leaving developers to reconcile them. Professor Joseph Sung, who led the study, framed the fix as collective ownership across every group involved rather than a burden dropped on any one of them.

Looking Forward

The MHRA’s healthcare AI reform consultation reported strong consensus last week. This study suggests the harder problem is not writing rules but reaching the people who build the software — particularly at the two-thirds of organisations currently working to no framework at all.