What is an AI kill switch in healthcare?
An AI kill switch in healthcare is a control letting a hospital rapidly pause, disable, or withdraw an AI system or AI agent when it behaves outside its intended use. A complete kill switch defines who can trigger it, how fast it takes effect, what happens to work already in progress, and how the stop is recorded for later review.
Why hospitals need one
AI tools change over time. Vendors ship updates, input data shifts, and agents that call other software can repeat one mistake many times before anyone reads a report. The NIST AI Risk Management Framework names the need directly. Its MANAGE 2.4 subcategory calls for mechanisms "to supersede, disengage, or deactivate AI systems that demonstrate performance or outcomes inconsistent with intended use," with responsibilities "assigned and understood." GOVERN 1.7 adds that organizations should have processes for "decommissioning and phasing out AI systems safely."
The Joint Commission and CHAI guidance on the Responsible Use of AI in Healthcare (RUAIH) lists "major performance degradation after an update" among the AI safety events hospitals should capture and treat like patient safety events. A stop control is how a hospital acts on that finding the same day, rather than waiting for a vendor release.
What a working stop control includes
A kill switch is more than an off button. In practice it has five parts:
- Scope. The hospital can stop one agent, one permission (for example, "send patient messages"), one version, or every tool from one vendor, without taking down unrelated systems.
- Authority. A short, named list of people can trigger it, such as the tool's clinical owner, the AI governance lead, and an on-call informatics lead.
- Speed. The stop applies to the next action the AI tries, not the next deployment cycle. Hospitals should test and time it before go-live, then again after major updates.
- Fallback. Staff know which manual workflow resumes, so care continues safely when the tool is off and patients are not left waiting.
- Record. Who stopped it, when, why, and what the tool did just before are written to an audit trail that the hospital controls.
Kill switch versus circuit breaker
Some standards draw a line between the two. The OVERT runtime evidence specification (RES-4) requires "scoped, time-bounded revocation or equivalent circuit breaking" and says revocation should be designed as a local, bounded, self-healing circuit breaker rather than a centralized kill switch. In that model a stop is limited to one organization's own tenant and expires unless renewed. Hospitals can use both ideas: a local, scoped stop for a single tool, plus a documented process for permanent retirement under GOVERN 1.7.
How Skovos handles this
Skovos gives the hospital a recall control it owns. Because each registered agent checks permission with Skovos before it acts, a recall means the next permission check is denied, and the recall itself is written to the hash-chained audit trail.
Frequently asked questions
Who should be allowed to stop an AI tool in a hospital?
A small, named group: typically the tool's accountable clinical or business owner, the AI governance lead, and an on-call clinical informatics or IT lead. The list should live in the AI inventory so frontline staff know whom to call.
Is turning off the server a kill switch?
It is a last resort, not a plan. It is slow, usually needs the vendor, and can take down unrelated functions. A good stop control revokes the tool's authority at a point the hospital controls.
Does any regulation require an AI kill switch?
No US regulation uses the term. The NIST AI RMF (MANAGE 2.4) is voluntary but explicitly calls for the ability to deactivate AI systems, and Joint Commission and CHAI guidance expects hospitals to monitor and respond to AI safety events.
How often should a stop control be tested?
Test it before go-live and on a regular schedule after that. OVERT RES-4.5, for example, calls for testing revocation at least every 12 months.
Related reading
- How to stop an AI agent in a hospital
- AI agent registry for hospitals
- AI safety monitoring for health systems
Sources
- NIST, Artificial Intelligence Risk Management Framework (AI RMF 1.0), January 2023: https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.100-1.pdf
- Joint Commission and CHAI, Guidance on the Responsible Use of AI in Healthcare (RUAIH), 2025: https://digitalassets.jointcommission.org/api/public/content/dcfcf4f1a0cc45cdb526b3cb034c68c2
- OVERT v1.1 standard text, RES-4 Scoped Revocation and Circuit Breaking: https://overt.is/latest.md