The number of US medical malpractice verdicts reaching $10 million or more grew by roughly 67% between 2013 and 2023, according to The Doctors Company. Clinical AI is now entering that environment at scale, with tools embedded in diagnostics, prescribing, and treatment planning across American health systems. How that technology affects who bears liability - and whether it increases or reduces the risk of harm - is a question carriers, risk managers, and hospital legal teams are only beginning to work through.
Jared Kaplan (pictured), CEO of Michigan-based insurtech Indigo, a medical malpractice carrier built on individual physician risk data, shares where the early evidence is pointing and what it means for coverage, documentation, and litigation exposure.
Risk managers waiting for AI-specific malpractice data to sharpen their strategy may be waiting longer than they expect.
When asked about what the claims record shows right now, Kaplan’s answer is blunt. "The honest answer is that the data doesn't exist yet, and I would be skeptical of anyone who suggests otherwise," he said. "Medical malpractice has a long reporting tail: claims arising from care delivered in 2025 and 2026 are largely still years away from resolution, while clinical AI has been in widespread use for only about two years."
What is appearing in the record now, Kaplan noted, is AI as a factor in care, not as the alleged cause of harm. The tools turning up in case files include ambient documentation systems, imaging triage software, and clinical decision support platforms.
"The emerging failure mode is not necessarily that the model was wrong," Kaplan said. "It is the gap between what the tool recommended and what the clinician documented about why they agreed or disagreed. That is fundamentally a documentation issue before it is a technology issue - and it is one that organizations can address today."
Most risk managers treat nuclear verdicts as a market-wide pressure that healthcare happens to share. Kaplan's starting point is different.
"Healthcare is not merely following the nuclear verdict trend; it is one of the areas driving it," he said. "Claim frequency has declined substantially over the past decade, while severity has moved in the opposite direction. The largest verdicts now average around $50 million."
The verdict environment is compounding existing strain across the medical malpractice market's underwriting environment. The natural instinct is to treat AI as a new source of malpractice exposure. Kaplan, however, tested that instinct against published safety data from outside the healthcare sector.
"Consider autonomous driving: the initial concern was that automation would create a new category of accidents, but published data indicates the opposite," Kaplan said. "Waymo reports 82% fewer injury-causing crashes and 94% fewer serious-injury-or-worse crashes than the human benchmark across the cities where it operates – approximately 0.7 injury crashes per million miles, compared with a human rate near 3.9."
Waymo's published safety data confirms those figures. It covers 220.6 million rider-only miles through March 2026 across five US cities. A Swiss Re-led peer-reviewed study of the same platform found it eliminated bodily injury claims entirely across 3.8 million autonomous miles.
The structural argument extends directly to medicine, Kaplan explained. "The reason automation can work is that most crashes stem from lapses in attention and fatigue - limitations machines don't share," he said.
"Medicine has a comparable structure: a significant share of malpractice claims can be traced to diagnostic misses, missed follow-up, and communication breakdowns, precisely the kind of failures that become more likely at 2:00am during a double shift. We, therefore, expect clinical AI to reduce the frequency of medical error rather than increase it."
Kaplan pointed to a second, more immediate variable: what AI leaves behind in the record.
"The genuinely new factor is the audit trail," he said. "AI can create a permanent, discoverable record of decision-making that manual practice did not generate."
That record cuts in both directions, Kaplan explained. A plaintiff can argue the tool flagged an issue the physician ignored. The defense can point to documented clinical reasoning the physician would not otherwise have produced.
Individual case defensibility gets most of the attention in discussions about AI and malpractice. The structural question, Kaplan argued, is a different one entirely.
"What may shift is not necessarily the amount of harm, but who ultimately bears the liability," Kaplan said. "As automated systems assumed more of the driving decision, liability began moving toward the manufacturers and away from operators.
"If medicine follows a similar path, physician exposure could narrow over time while a new product liability exposure develops for AI vendors. That transition is what carriers and risk managers should be watching. It is a fundamentally different question from whether AI makes an individual malpractice case harder to defend."
How AI involvement affects jury dynamics is another concern, according to Kaplan.
"Complexity doesn't automatically favor plaintiffs; it favors the side with a clearer, more intuitive story," he said. "The risk is that the phrase 'black box algorithm' offers plaintiffs a simple, emotionally compelling narrative, while the defense may need extensive expert testimony to explain and contextualize the technology."
That liability gap is not unique to medical malpractice. Across other lines, AI liability exposure is already outpacing the coverage organizations think they have. Kaplan drew the distinction at the level of the medical record, rather than the technology itself. Two scenarios produce very different cases.
"If the tool was confirmatory and the physician's independent clinical reasoning is documented alongside it, the defense is centered on the physician's judgement," he said. "If the tool was decisive and that reasoning is absent from the record, the defense is effectively being asked to defend an algorithm – a far more difficult case to present to a jury.
"Risk managers should, therefore, focus less on whether to use AI at all and more on how it is implemented. The answer to litigation exposure isn't slower adoption; it's stronger documentation discipline around the technology already in use."
Much of the market conversation around AI and medical malpractice treats pricing as an open question carriers are actively working through. Kaplan draws a sharper line.
"No carrier is pricing AI risk with precision - including us - because there's not yet credible loss data against which to price it," Kaplan said. "Any insurer applying an AI-specific load today is making an estimate rather than relying on mature actuarial experience."
The AI pricing gap, however, is not Kaplan's primary concern. He argued that a more immediate mispricing is already present in how the market underwrites individual physicians. Most carriers rely on three variables alone: specialty, geography, and claims history.
"That approach overlooks significant variation in individual physician risk, creating a live mispricing today - not merely a hypothetical issue three to five years from now," he said. Over the same window, Kaplan noted, claim-severity trends and social inflation are likely to drive loss costs far more than any AI-related claims.
His longer-term view runs against the direction of current market thinking. Where most of the industry treats AI as a surcharge risk, Kaplan anticipated the opposite outcome.
"Our view is that proven clinical safety tools may eventually function much like home security systems in homeowners insurance, where adoption earns a premium credit rather than a surcharge," he said. "That may still be a decade away, but it is the likely direction of travel."
Kaplan's advice for risk managers does not require the AI medical malpractice liability picture to clarify before it becomes useful. Each of the five steps he outlined can be acted on now.
The first is to build a complete inventory of every AI tool in clinical use. Kaplan noted that the shadow inventory of tools clinicians have adopted outside formal procurement channels is almost always larger than leadership realizes. That gap needs to close before any meaningful risk assessment can begin.
The second step follows directly: the medical record must document clinical reasoning alongside the tool's output, not just the tool's recommendation. Kaplan described this as the single highest-leverage action on the list.
Vendor contracts are the third area, and one of the two most commonly overlooked. "Many vendor agreements disclaim clinical responsibility entirely," Kaplan said, "and organizations should know that before a claim arises, not after."
The fourth step is disclosure. Undisclosed AI tool exposure is more problematic at renewal than disclosed exposure, Kaplan noted. Organizations should tell their carrier and broker which tools are in use. Carriers that understand a health system's technology stack are better positioned to price the risk appropriately. A written AI risk management policy is the starting point for making that conversation productive.
The fifth and most commonly neglected step concerns what AI systems retain. "Model outputs and clinician override logs may be discoverable," Kaplan said. "No organization should first learn what its electronic health record preserves during a deposition."
