AI in the Clinic: The Slow Path to Adoption
Medicine is where these systems could do the most good and where the evidence bar is highest. The bar is the slow part, and it is also the part that produces trust.
Healthcare is the domain where the case for artificial intelligence is most compelling and where deployment has been slowest. Both facts have the same cause: the cost of being wrong is measured in harm to a person.
Imaging, scribing, coding, queue order
- Medical imaging, where a model flags findings for a radiologist to confirm.
- Documentation, where ambient scribing removes a substantial clerical burden.
- Coding and billing, where errors are costly and verification is straightforward.
- Triage support, where the model prioritises a queue rather than making a decision.
A model used in care is a medical device
A model used in clinical care is a medical device in most jurisdictions, which means prospective validation, documented performance across populations, monitoring after deployment, and clear allocation of liability when something goes wrong.
These requirements are expensive and slow, and they are the reason the sector has not moved at the pace of consumer software. They are also the reason clinicians trust the tools that do get through.
Three debates that have not resolved
Whether a model that performs well on average but poorly on a subgroup should be permitted. What happens when a clinician disagrees with a recommendation they cannot explain. And who is accountable for a decision made with a recommendation nobody fully understands.
None of these are technical questions, and all of them determine the pace of adoption more than capability does.
The clinician keeps the legal responsibility
When a model recommends and a clinician accepts, who is responsible for the outcome? The clinician retains legal responsibility in most jurisdictions, which creates an awkward incentive: adopting a tool that improves average outcomes while exposing the individual to blame for its errors.
Health systems have addressed this through governance rather than law, specifying when a recommendation may be relied upon and documenting the reasoning. That works where the system’s behaviour is understood, and it becomes difficult when a model’s basis for a recommendation cannot be explained.
Validation, monitoring, and timelines in years
Approval regimes require evidence of performance across the populations where the tool will be used, monitoring after deployment, and a process for reporting problems. Those requirements are appropriate and they produce timelines measured in years.
The practical consequence is that the tools reaching clinics are the ones attached to organisations with the resources to navigate that process: established device manufacturers and large health systems. General-purpose assistants enter through documentation and administration rather than clinical decision-making.
Image credit and licence details for every photograph on this site are listed on the credits page. This article is editorial content; it carries no sponsored material.