AI in the Medical Field + Accountability for its Mistakes
AI is everywhere. The majority of professions and career fields today are affected - in some shape or form - by the usage of AI, and medicine is no exception. Such nonhuman involvement raises crucial questions in terms of liability. Who is to blame when AI makes a mistake? Who is to blame for a misdiagnosis or an inefficient treatment plan? Despite the increasing use of artificial intelligence in hospitals and clinics, there are many unexplored complexities regarding medical liability when technology errs or shows bias. The legal system today is equipped to handle human medical malpractice; AI blunders are uncharted territory and have yet to be confronted by the law.
Currently, AI is slowly being incorporated into healthcare, with prominent usage in many areas already. Firstly, a particular AI software is being used to interpret brain scans. According to tests conducted in the UK, this software can even identify the timeframe in which the patient’s stroke occurred! Similarly, many studies have noted that AI can sometimes be used to analyze x-rays to ultimately detect bone fractures that care providers may have initially missed. This particular feature appears to be especially beneficial as x-ray technicians are in demand. Additionally, some healthcare institutes have experimented with using AI to determine ambulance requirements and, in turn, which patients actually required hospital transfer. Finally, it is also observed that AI can be used to predict and detect diseases. However, while all of these abilities seem efficient and promising, it cannot be discounted that technology can make mistakes; errors that the law has yet to specifically regulate.
Of course, there are human-based liability standards built into the law, especially regarding expert testimony and the admissibility of it. These pre-existing standards have been shaped by two important cases: Frye vs. United States (1923) and Daubert vs Merrell Dow Pharmaceuticals (1993). In Frye, a man accused of second-degree murder wished to introduce a new type of lie detector in order to prove his innocence. Interestingly, The Court of Appeals for the District of Columbia ruled that the expert testimony (by the doctor who had administered the test prior to the trial) regarding the validity of the test was inadmissible. The reasoning behind this decision lay in the fact that the test was not yet, at the time, widely accepted by the psychological community and thus its use as evidence would be unjustifiable. The Frye ruling thus came to be known as the “general acceptance” test, which was used for 70 years! Eventually, however, the Frye standard was superseded by the Daubert standard, established in Daubert vs Merrell Row Pharmaceuticals. In Daubert, two children and their respective parents sued Merrell Row with the claim that the children were born with severe birth defects because their mothers ingested Bendectin. The lower court excluded the plaintiff's expert evidence on the basis that it was not “generally accepted”, a necessity that had been outlined by Frye. Nonetheless, the Supreme Court overturned the lower court’s decision and ruled that the Federal Rules of Evidence (FRE) superseded the Frye standard. The Supreme Court ruling emphasized the expert’s methodology over simply scientific acceptance and held that trial judges must now act as gatekeepers, ensuring that the testimony is relevant and reliable. Factors (set by Daubert) that a judge should consider when determining the validity of the theory or technique in question include whether the theory can be/has been tested, whether it has been peer reviewed, the known/theoretical rate of error, the existence/maintenance of standards regarding the operation of the technique, and, of course, whether or not it is "generally accepted” in the scientific community. It becomes visible that AI-generated evidence poses critical challenges to these clear-cut standards, contradicting almost all the necessary factors required to be considered admissible. For example, AI evolves incredibly fast and is updated surprisingly often, leaving no time for peer review. Similarly, AI’s complex and generative nature makes it inexplicable, sometimes, even by experts. Consequently, a rate of error is also difficult to derive. Finally, if an AI company refuses to disclose their source codes, it will become even more challenging to evaluate the methodology of the tool used. Clearly, the generative nature of artificial intelligence poses significant challenges to the standards regarding admissibility that are currently in place.
More questions regarding liability arise when authoritative approval is involved; who is to blame if the product/technology in question has been approved federally? To examine this concept, it is imperative to take a close look at Riegel v. Medtronic (2008) and Wyeth v. Levine (2009). Riegel was about a man named Charles Riegel who suffered severe injuries when a medical device - approved by the strict FDA premarket approval (commonly referred to as PMA) - ruptured during surgery. Surprisingly, the Supreme Court ruled that FDA-approved devices cannot be subject to state lawsuits over the safety or design of the product because that would conflict with federal regulation. Naturally, there are exceptions to this principle: for example, if a product doesn’t follow the FDA rules or wasn’t manufactured according to the approved style. Nevertheless, Riegel served as a powerful shield for medical device manufacturers. The decision undoubtedly causes concern by blatantly raising the question: what will happen if the FDA approves AI technology? Interestingly, the answer to such a question may be found within Wyeth v. Levine. In another state lawsuit regarding a medical accident and improper labeling of a product, Wyeth (a pharmaceutical company) argued that because the FDA approved their label, federal law preempted Levine’s claims, which were based on state law. The Supreme Court eventually ruled against Wyeth, reasoning that FDA approval was not a complete shield against lawsuits; as there was no “impossible” conflict between the federal and state law, the Court found that Wyeth could simply comply with both and strengthen their labeling. Wyeth is a landmark case that reveals the importance of transparency from AI developers and the necessity of understanding that AI technology is not yet perfected. Companies must understand that, in the case of - as an example - bias from technology that relies on generative machine learning, they are still in danger of a lawsuit.
After examining the laws currently in place to handle medical liability, the challenges that AI poses to this existing legal frame become very clear. To begin, the “black box” problem is the most prominent concern within this matter. It refers to the general inability to understand AI’s complex generative process from start to finish, making it very unclear how the algorithm came to an incorrect conclusion: perhaps an inaccurate diagnosis or flawed treatment plan. Consequently, it becomes very difficult to identify who is to blame. Is it the physician, who utilized AI? Is it the hospital or clinic, who sanctioned/implemented the usage of the AI? Maybe it is the developer, for providing flawed technology. Furthermore, the validity of AI becomes incredibly problematic to solidify, especially in terms of the standards established in Daubert vs Merrell Dow Pharmaceuticals. Artificial intelligence is difficult to peer review, operates without a known rate of error, often cannot be traditionally “tested” (because outputs vary significantly based on input), and does not always have/follow clear-cut standards in terms of usage.
As the field of medicine expands in unforeseen ways due to the addition of artificial intelligence, the law must expand with it to adequately account for AI liability. The medical liability laws that exist today - whether it be regarding expert testimony as seen in Daubert vs Merrell Dow Pharmaceuticals or the products themselves as referenced in Riegel v. Medtronic and Wyeth v. Levine - are not equipped to fairly rule upon the mistakes made by AI. While these cases set undeniably crucial precedents, rules and boundaries must be set, especially for technological companies, before they are crossed.
Bibliography
“Daubert v. Merrell Dow Pharmaceuticals, Inc. | 509 U.S. 579 (1993) | Justia U.S. Supreme Court Center.” Justia. Accessed March 16, 2026.
https://supreme.justia.com/cases/federal/us/509/579/.
North, Madeleine. “7 Ways Ai Is Transforming Healthcare.” World Economic Forum, August 13, 2025. https://www.weforum.org/stories/2025/08/ai-transforming-global-health/.
“Wyeth v. Levine | 555 U.S. 555 (2009) | Justia U.S. Supreme Court Center.” Justia. Accessed March 16, 2026. https://supreme.justia.com/cases/federal/us/555/555/.
“Riegel v. Medtronic, Inc. | 552 U.S. 312 (2008) | Justia U.S. Supreme Court Center.” Justia. Accessed March 16, 2026. https://supreme.justia.com/cases/federal/us/552/312/.