The International AI Safety Report 2026
found something most institutions
would rather not sit with.
AI models gave harmful answers
to 19 percent of medical questions.
Not edge cases.
Nearly 1 in 5.
That number does not mean
the technology is broken.
It means the system it landed in
was not ready to own what it produced.
AI amplifies whatever system it enters.
Strong governance becomes more precise.
Weak governance becomes more visible.
Most institutions are having
the wrong conversation.
They are asking who owns the data.
That conversation is necessary.
It is not sufficient.
Because in clinical AI,
the data problem is masquerading
as a technology problem.
And the governance problem is masquerading
as a data problem.
That is the layer
the data conversation keeps missing.
You can clean every pipeline.
Govern every input.
Audit every source.
If no one owns the clinical call at the end,
clean data produces confident outputs
with no one to defend them.
The institutions closing this gap
are not buying better technology.
They are having a different conversation.
↳ What did we authorize this agent to do?
↳ Who owns the decision it shapes?
↳ Where does the Handoff live in the chart?
The Accountability Gap™ (TAG™)
does not live at the data layer.
It lives at the Handoff.
Clean data is not the same as a named owner.
You already know which deployment
has one without the other.
Mo Johnson, MD MBA is a cardiothoracic surgeon and the founder of GPe Research. Field Notes are short dispatches from the clinical AI accountability frontier, published alongside the MedicoVigilance™ newsletter at medicovigilance.org.
Follow the work on LinkedIn: linkedin.com/in/mo-johnson