AI needs governance, not hype
Experts argue that clinical AI performs better when constrained by validated terminology and explicit context instead of free‑text generation. A technical piece recommends grounding models in FHIR terminology services to reduce hallucinations, and broader reporting warns that rapid AI deployment raises regulatory, privacy and responsibility questions for healthcare organisations ( ).
A hospital note is written in messy human language, but billing, quality reporting, and many clinical systems run on fixed code lists. When an artificial intelligence model jumps straight from free text to a code, it can sound fluent and still invent a code that was never supported by the record. (better.care) The fix is less like giving the model a bigger brain and more like giving it a locked menu. The Health Level Seven Fast Healthcare Interoperability Resources standard, usually called FHIR, includes terminology services that let software look up valid code systems, expand allowed value sets, and validate whether a code is actually permitted in a given context. (hl7.org) A code system is the master dictionary, like the official list of airport codes. A value set is the shorter list you are allowed to use for one task, like “only airports in California” instead of every airport on Earth. (hl7.org) The Better.care piece published on April 9, 2026 says its approach is to ground clinical coding agents in FHIR terminology services and pass that structure through the Model Context Protocol, which is a way for a model to call outside tools instead of guessing from memory. The result it describes is “precise, verifiable answers” in place of confident hallucinations. (better.care) That changes the model’s job. Instead of asking it to improvise a diagnosis code from a paragraph, you ask it to find the right concept, check the allowed list, and return something a terminology service can verify against a CodeSystem or ValueSet resource. (hl7.org) This is not just a technical cleanup for coders. The United States Food and Drug Administration says artificial intelligence and machine learning in medical devices need careful management across the full product life cycle, and its January 6, 2025 draft guidance asks developers to address transparency, bias, design, development, and documentation. (fda.gov, fda.gov) The pressure to move fast is coming from the market at the same time. Mashable’s April 9, 2026 overview says healthcare groups are being pushed to adopt artificial intelligence tools while patients and providers are still sorting out privacy, regulation, and who is responsible when a system gets something wrong. (mashable.com) Those responsibility questions are already turning into rules. Illinois signed a law on August 1, 2025 banning artificial intelligence from acting as a standalone therapist, which shows regulators are willing to draw lines when software starts looking too much like unsupervised care. (mashable.com) Federal regulators have drawn a similar line in insurance. The Centers for Medicare and Medicaid Services said insurers cannot rely entirely on an algorithm or artificial intelligence system to make coverage decisions, which puts a human-accountability backstop under automated tools. (mashable.com) So the real split in healthcare artificial intelligence is not between “uses AI” and “doesn’t use AI.” It is between systems that are tied to validated medical vocabularies, explicit context, and reviewable outputs, and systems that are asked to freestyle in one of the most regulated sectors in the economy. (better.care, hl7.org, fda.gov)
