AI Risk Atlas Prototype/DemoUnofficial independent experiment. Not an official xAI product. Scores can be wrong.

Back to mitigations
Proposed

Regulate high-stakes health modes as medical devices

If it behaves like a device, it is licensed like one.

No compiled flow or public database names this control yet.

Who should own it
FDA / MHRA-class regulators
Regulators
How quickly it can land
Quarters
Cross-org work; expect most of a year.
Expedited implementation
6 months
180 calendar days with a crash team
Normal implementation
1.5 years
540 calendar days as a planned program

Risk this mitigates

20
De facto clinical AI without clinical controls

Given that a general chatbot is already a weekly health advisor for hundreds of millions of people, there is a possibility of medical decisions being made on model output that is not a licensed device, not a record, and not under a clinician’s duty resulting in population-scale diagnostic and dosing errors, and delayed care for people who trusted the chat more than a clinic.

Residual composite 20 · still above the threshold

Effect if implemented

Applied to every failure scenario on that risk, then re-ranked. Axes are clamped at 1.

Likelihood
1
Consequence
1
Urgency
0