Nobody gets sued but the doctor: The legal vacuum at the center of the AI physician revolution
Abstract
Who absorbs liability when an AI-assisted clinical decision causes patient harm, and what does the current legal ambiguity mean for founders, investors, and health systems deploying these tools at scale?
Key findings and data points:
- FDA has cleared over 1,300 AI-enabled medical devices as of late 2025, with 295 cleared in 2025 alone; 62% are software as a medical device (SaMD)
- 66% of U.S. physicians used AI in clinical practice in 2024, up from 38% in 2023 – a 78% single-year jump
- Malpractice claims involving AI tools increased 14% between 2022 and 2024; most involved diagnostic AI in radiology, cardiology, and oncology
- No existing federal law assigns liability to AI developers when an AI tool contributes to patient harm; current courts default to the treating physician
- The Federation of State Medical Boards recommended in April 2024 that clinicians, not vendors, be held liable for AI-generated errors
- GPT-4 outperformed physicians using GPT-4 in complex diagnostic cases in a 2024 study, suggesting the human-AI hybrid may actually underperform pure-AI in some contexts
- AI healthcare market estimated at $39.25B in 2025, projected to reach $504B by 2032 at ~44% CAGR
- AI-focused digital health deals in H1 2025 ran 83% larger than non-AI deals; $3.95B of the $6.4B raised went to AI companies
- Only 2% of U.S. radiology practices had integrated AI reading tools by 2024 despite hundreds of FDA-cleared options
The argument: The AI diagnostics revolution is outrunning the legal infrastructure designed to govern it. Physicians bear all the liability, vendors bear almost none, and the incentive structures this creates are deeply misaligned with both patient safety and long-term enterprise value for companies in the space. The founders who figure this out early – and build accountability into their products rather than contract it away – will have a significant structural advantage.
Table of Contents
The Setup: 950 Cleared Devices, 2% Adoption, and a Liability Cliff Nobody’s Talking About
What the FDA Has and Hasn’t Done
The Physician Gets the Bill: How Malpractice Law Currently Assigns Blame
Deskilling, Overdependence, and the Colonoscopy Problem
The Black Box Defense and Why It Won’t Hold
What the EU Is Doing That the U.S. Isn’t
Founder Implications: Build the Accountability Layer Now
Investment Thesis: The Liability Arbitrage Window Is Closing
The Setup: 950 Cleared Devices, 2% Adoption, and a Liability Cliff Nobody’s Talking About
Here’s a number that should make every health tech founder pause: as of late 2025, the FDA has cleared more than 1,300 AI-enabled medical devices. Radiology alone accounts for the overwhelming majority. And yet, a 2024 Associated Press survey found that only about 2% of U.S. radiology practices had actually integrated AI reading tools into their workflows. Think about that gap for a second. It’s not a regulatory problem. It’s not a technical problem. The tools exist, they’re cleared, and in many cases the clinical evidence behind them is legitimately strong. A large Swedish randomized trial found 17.6% higher cancer detection rates with AI-assisted mammography screening. Viz.ai’s stroke detection algorithm hits AUC above 0.90 on retrospective datasets. Aidoc’s intracranial hemorrhage tool reports sensitivity above 90% with low false-positive rates. The performance is real.
So what’s the holdup? Some of it is workflow friction, integration headaches, and the usual institutional inertia that makes healthcare adoption timelines look like geological epochs. But a big, underappreciated chunk of it is liability terror. Clinicians are watching a legal landscape where they get to absorb all the downside of an AI error, while the vendor takes the upside in the form of a recurring SaaS contract. That is a structurally bad deal, and the smarter physicians are figuring that out fast.
The broader picture is even more interesting. U.S. digital health startups raised $6.4B in H1 2025, with AI-focused companies capturing roughly $3.95B of that, and those deals ran 83% larger than non-AI deals on average. Eight healthcare AI unicorns were minted in 2025. The money is flowing in fast and hard. Meanwhile, the legal infrastructure governing what happens when one of these tools contributes to a missed diagnosis or a treatment error is essentially non-existent in any coherent, codified sense. There is no federal statute. There is almost no AI-specific case law. There are strongly worded recommendations from medical boards telling physicians they’re on the hook, and there are vendor contracts full of indemnification language that makes the physician’s counsel wince. That combination – massive capital inflows, real clinical deployment, and a liability vacuum – is the setup for a genuinely consequential legal and market reckoning.
What the FDA Has and Hasn’t Done
