A new analysis from Nanyang Technological University, Singapore, reports that many professionals building artificial intelligence for clinical use are not well-versed in the rules meant to keep patients safe. Published in npj Digital Medicine, the survey of 122 medical AI developers across several regions found uneven awareness of regulatory frameworks and limited adoption of formal standards by their employers—signals that governance has not kept pace with rapid deployment in health care.
What the researchers found
The study—among the earliest multi‑regional surveys focused specifically on regulatory literacy in medical AI—polled developers in Singapore and elsewhere, including China, Hong Kong and the United Kingdom. As AI systems are increasingly used to interpret medical images and estimate disease risk, the authors warn that weak familiarity with guardrails such as the EU AI Act or Singapore’s Artificial Intelligence in Healthcare Guidelines could undercut safe implementation and public confidence.
| Measure | Result |
|---|---|
| Developers aware of at least one regulatory framework | 57% |
| Organizations without any adopted AI regulatory framework | 67% |
| Total respondents | 122 |
Experience and workplace setting mattered. Senior developers were more likely to recognize multiple frameworks and demonstrated greater familiarity than junior colleagues. Respondents working outside academia also tended to report higher awareness compared with those in academic settings. Importantly, developers employed by organizations that had formally adopted frameworks were notably more conversant with them than peers at non‑adopting institutions.
Accountability, but patchy governance
Participants largely agreed that those who build clinical AI systems should carry responsibility for their products. They also pointed to the need for oversight by government or institutional ethics bodies, and for collaboration with both end‑users and the data‑providing organizations that underpin model development. Still, the authors highlight a substantial implementation gap: even where individual awareness exists, most developers reported that their workplaces have yet to embed recognized frameworks into practice.
Why it matters for patients and providers
Clinical AI is moving from pilot projects into everyday use, from radiology triage to predictive tools that flag patients at risk of deterioration. Without clear, shared understanding of regulatory expectations—tied to concrete organizational adoption—systems may be deployed without adequate documentation, validation, bias assessment or post‑market monitoring. That raises the likelihood of performance drift, inequitable outcomes, and erosion of trust among clinicians and the public.
- Safety and effectiveness: Limited regulatory literacy can translate into gaps in testing, risk management and real‑world surveillance.
- Equity and transparency: Frameworks set expectations around dataset provenance, bias checks and explainability—crucial for fair access to care.
- Operational readiness: Clinicians and hospitals rely on robust documentation and accountability chains to integrate AI safely into workflows.
Bridging the knowledge-to-action gap
The authors recommend more direct engagement with developers to align day‑to‑day engineering choices with policy requirements. Because organizational uptake appears to drive familiarity, health systems and companies investing in AI may benefit from formal adoption of recognized standards and internal education tied to product lifecycles—from data curation and model training to deployment and monitoring.
The findings also suggest targeted support for junior developers and academic teams, who reported lower awareness on average. Practical measures could include training on applicable guidance, clear assignment of accountability within project teams, and partnerships that bring developers, end‑users and data stewards into earlier, structured dialogue about risk controls.
The bottom line
As countries scale up digital innovation in health, the study underscores a simple but consequential point: technical talent alone is not enough. Building and deploying medical AI safely requires shared fluency in the rules of the road—and workplaces that make those rules standard operating procedure. With 57% of respondents aware of at least one framework but 67% reporting no organizational adoption, the gap between knowledge and implementation remains a critical challenge.
