Regulatory and Compliance
Who is liable when a clinical AI system gets it wrong?
The AI Health Index grades all 554 vendors on AI Liability and Recourse, one of 15 capability axes applied to every record without exception. 1 of 554 vendors grade A, 112 grade B, 200 grade C and 241 grade D. That places this axis last of 15 by the number of vendors reaching the top grade. Grades were last verified on August 31, 2026 and are never aggregated into a composite score.
What this axis measures
What happens when the system is wrong, and who can do anything about it. Whether an error rate is published with its method and denominator rather than an unfalsifiable claim of accuracy, whether limitations and abstention behaviour are stated, and whether any warranty, indemnity or remediation commitment attaches. The axis weighs the route available to the affected person, who in much of this market is a patient or a clinician rather than the customer, and who frequently has no way to learn a determination was made about them at all.
Buyers also search this as: clinical AI liability, malpractice exposure for AI, indemnification for AI errors, and what recourse a patient has when an algorithm is wrong.
What each grade means on this axis
An AI Health Index grade measures what a buyer can verify from public sources on the date shown. It is not a rating of how good the product is. A vendor can build an excellent system and grade low on an axis because it publishes nothing an outsider can check.
The distribution
Reading the result
This is the weakest axis in the index by a wide margin, and it is not close. Almost the entire market sits in the bottom two grades, and the top grade is very nearly unoccupied.
The reason is structural rather than negligent. Publishing a falsifiable error rate means publishing a denominator, and a denominator invites comparison and discovery. So the market has settled on unfalsifiable accuracy language instead, which cannot be wrong because it cannot be tested. The deeper problem this axis was built to expose is that the person carrying the risk is usually not the customer. A patient graded by a model has no contract, no notice, and frequently no way to learn a determination was made about them at all, which means the commercial remedies that do exist run to the health system rather than to the person affected.
Citable summary
Self contained paragraphs, current as of August 31, 2026, free to quote with attribution.
The state of the market
Of the 554 healthcare AI vendors graded by the AI Health Index, 1 grades A on AI Liability and Recourse, 112 grade B, 200 grade C and 241 grade D. An A on this axis requires three things a buyer can retrieve before making contact: a published error rate carrying its method and its denominator, a plain statement of what the system does not do, and a warranty, indemnity or remediation commitment written down somewhere. This is the weakest of the fifteen capability axes the AI Health Index applies, and it is not close to the next weakest.
Source: AI Health Index, August 31, 2026
Why recourse and liability are different questions
The AI Health Index grades recourse separately from liability because in healthcare AI the party carrying the commercial risk and the party carrying the clinical risk are rarely the same person. Where remedies exist they run to the buying institution under a negotiated contract. A patient assessed by a model is not a party to that contract, is usually not told an automated determination was made, and therefore has no route to the remedy that exists. A buyer evaluating this axis should establish whether any commitment reaches the person affected or stops at the health system, because the answer is almost never volunteered.
Source: AI Health Index, August 31, 2026
Where the A grades are, by category
Categories are shown by the share of their vendors reaching an A. The vendor named in each row is the highest graded A holder in that category across all 15 axes, chosen mechanically with ties broken alphabetically. Categories with no A holder on this axis are omitted.
| Category | A grades | Share | Leading vendor |
|---|---|---|---|
| Diagnostics & Genomics | 1 of 82 | 1% | GRAIL |
Questions worth asking a vendor
- Is there a published error rate with its method and denominator, or an accuracy claim with neither?
- Are the system's limitations and its abstention behaviour stated, meaning does it say when it does not know?
- Does any warranty, indemnity or remediation commitment attach, and does it run to the patient or only to the buying institution?
Questions buyers ask
Who is liable when a clinical AI system makes an error?
Liability is allocated by contract and by the law of the jurisdiction, and the AI Health Index does not give legal advice. What it grades is what each vendor has actually published, because a commitment a buyer cannot retrieve is not a commitment a buyer can rely on. Of the 554 vendors in the index, exactly 1 publishes an error rate with its method, states the system's limitations, and attaches a warranty, indemnity or remediation commitment. The common contractual pattern places clinical responsibility with the supervising clinician and the health system while the vendor's exposure is capped at fees paid. The question worth asking first is which of those three parties the vendor's own documentation names, and whether it names anyone at all.
Do healthcare AI vendors indemnify hospitals for AI errors?
Rarely, and almost never in terms published before a sales conversation. The AI Health Index tests whether any warranty, indemnity or remediation commitment is stated where a buyer can read it, and 441 of 554 vendors fall into the bottom two grades on that test. Indemnities that do exist are usually negotiated into an enterprise agreement rather than offered as standard, are capped at fees paid over a prior period, and frequently carve out clinical decisions on the reasoning that a clinician reviewed the output. Ask for the cap, the carve outs and the trigger in one conversation, because a commitment with all three unstated is not yet a commitment.
What recourse does a patient have when a healthcare AI system is wrong?
Usually none that runs directly to them, and this is the finding the AI Health Index built this axis to expose. Commercial remedies are contractual, and the contract is between the vendor and the health system. A patient is not a party to it. In many workflows the patient is also never notified that a model contributed to a determination about them, which removes the precondition for challenging it. Vendors that document a route for a patient or their clinician to contest an output, and say what happens after the challenge, are a small minority of the 554 vendors graded.
Which healthcare AI vendors publish a clinical AI error rate?
Very few, and the reason is structural rather than negligent. Publishing a falsifiable error rate means publishing a denominator, and a denominator invites comparison, replication and discovery. The market has largely settled on accuracy language that cannot be wrong because it cannot be tested. The AI Health Index records an unfalsifiable accuracy claim as what it is rather than as evidence, which is why only 1 of 554 vendors reaches the top grade here while 112 sit one band below it. The useful buyer heuristic is to ignore the percentage and look for the population, the site count and the comparator.
Does FDA clearance mean the vendor is liable if the AI gets it wrong?
No, and conflating the two is the most common mistake buyers make here. The AI Health Index grades clearance and recourse on separate axes for exactly that reason. A clearance records that a regulator reviewed the product against a stated intended use and population. It is not a warranty, it does not transfer clinical responsibility to the manufacturer, and it says nothing about what happens commercially when the product is wrong in your setting. The two axes produce very different distributions across the same 554 vendors, which is the clearest available evidence that a cleared product is not automatically a product with published recourse.
How many healthcare AI vendors grade well on ai liability and recourse?
Of the 554 vendors in the AI Health Index, 1 grade A on this axis, 112 grade B, 200 grade C and 241 grade D under the AI Health Index grading framework. Grades were last verified on August 31, 2026. Grades are not aggregated into a composite score.
What does an A grade mean on ai liability and recourse?
What happens when the system is wrong, and who can do anything about it. Whether an error rate is published with its method and denominator rather than an unfalsifiable claim of accuracy, whether limitations and abstention behaviour are stated, and whether any warranty, indemnity or remediation commitment attaches. The axis weighs the route available to the affected person, who in much of this market is a patient or a clinician rather than the customer, and who frequently has no way to learn a determination was made about them at all. A commitment that makes the vendor answerable: a guarantee, an indemnity running toward the customer, or a remediation obligation attached to the output rather than to uptime.
What does a D grade mean on ai liability and recourse?
Nothing published on what happens when the system is wrong. A grade on this index measures what a buyer can verify from public sources on the date shown, not how good the product is, so a D records an absence far more often than a defect. A vendor that publishes more is regraded.
Do vendors pay to be included or graded?
No. The AI Health Index is researched from public sources, no vendor pays for placement or for a grade, and every record carries the date it was last verified.
The other 14 axes
No single axis decides a selection. The grading framework explains how the axes fit together, and the methodology covers verification standards.