Public vs physician views of liability for artificial intelligence in health care

Abstract

The growing use of artificial intelligence (AI) in health care has raised questions about who should be held liable for medical errors that result from care delivered jointly by physicians and algorithms. In this survey study comparing views of physicians and the U.S. public, we find that the public is significantly more likely to believe that physicians should be held responsible when an error occurs during care delivered with medical AI, though the majority of both physicians and the public hold this view (66.0% vs 57.3%; P = .020). Physicians are more likely than the public to believe that vendors (43.8% vs 32.9%; P = .004) and healthcare organizations should be liable for AI-related medical errors (29.2% vs 22.6%; P = .05). Views of medical liability did not differ by clinical specialty. Among the general public, younger people are more likely to hold nearly all parties liable.

INTRODUCTION

Use of artificial intelligence (AI) in health care has increased markedly in the past decade.¹ In recent years, AI has been shown capable of detecting atrial fibrillation, calculating left ventricular ejection fraction, and predicting sepsis and in-hospital mortality, among other functionalities,² and there are now more than 60 AI-based algorithms and medical devices approved by the U.S. Food and Drug Administration (FDA).³

The growing use of AI in clinical practice has raised questions about liability for medical errors that result from care jointly delivered by physicians and algorithms. Given the relatively recent emergence of AI in health care, there is little established case law about how liability issues should be adjudicated,⁴ and regulatory policy from the FDA is evolving.⁵ For example, FDA officials have recently stated that the agency has “many questions” regarding best practices for the design, development, and application of AI and is interested in stakeholder views.⁶

Liability and regulatory issues are further complicated by the fact that algorithms are often “black boxes” that can dynamically change when incorporating new data, and because of growing evidence that they can sometimes produce recommendations that are racially or otherwise biased.⁷ No studies have compared public and physician views of medical liability related to use of AI in health care. We conducted nationwide surveys to examine these differences.

MATERIALS AND METHODS

We conducted separate surveys of physicians and the public; both were administered by SSRS, an independent research firm.⁸ Each survey instrument was developed through literature review and expert input. Surveys were pilot tested with 10 individuals for clarity and length, and revised accordingly. The public survey used a hybrid probability-based online panel and a dual-frame random-digit-dial telephone sample. Online participants were drawn from the SSRS Opinion Panel, which is representative of the online adult U.S. population.⁹ The sample was supplemented with the SSRS Omnibus sample, a national, weekly telephone survey.¹⁰ The survey was conducted in English and Spanish between December 3 and December 18, 2019. No incentives were offered for participation. A total of 1007 respondents participated; the survey margin of error was ±3.8 percentage points. The full public and physician surveys are presented in the Supplementary Documents 1 and 2.

The physician survey was conducted between November 27, 2019, and April 12, 2020. A total of 750 physicians (internists, oncologists, and radiologists) were invited to participate. The sample was drawn from CarePrecise—a comprehensive national database of clinicians that includes National Provider Identifier numbers, specialties, and mailing addresses.¹¹ Surveys were mailed to the physicians’ medical offices; 192 physicians completed the questionnaire online or via hard copy. Physicians received a $50 incentive for participation. Compared with nonrespondents, physician respondents were more likely to be male and to be internists than to be oncologists or radiologists. There were no statistically significant differences by age or years in practice. Further details are available in the Supplementary Table 5.

The surveys asked respondents (question 13 in the physician survey, question 15 in the public survey) whether each of 4 parties should be held liable for errors when physicians use computerized algorithms to select diagnoses and treatments: the physician making the clinical decision, the vendor or company licensing the algorithm, the healthcare organization purchasing the algorithm, or the FDA or other governmental entity approving the algorithm for clinical use (participants were asked who they believe should be held liable, not who they believe is currently held liable). We used chi-square tests to examine differences between physician and public views, and to examine differences in physicians’ views by specialty and the public’s views by age. Analyses were conducted using R version 3.6.2 (R Foundation for Statistical Computing, Vienna, Austria). The study was approved by the Institutional Review Boards at Weill Cornell Medicine and the Yale School of Medicine. All individuals consented to participate in the study.

Data availability

These data were collected as part of a national study of public and physician views of AI in health care. Data are not publicly available but may be shared on reasonable request to the corresponding author.

RESULTS

The physician respondents included 40.6% internists, 26.6% radiologists, and 32.8% oncologists; the mean number of years in practice across specialties was 26.6. Physicians from practices of different sizes, ownership types, and regions were represented. Public respondents were 60.9% non-Hispanic White and 51.3% female. The adjusted response rate was 32.2% for the physician survey and 49.0% for the public survey. Further details are available in Supplementary Tables 3 and 4 and Supplementary Methods.

Physicians were the most frequently cited liable party by both physicians and the public; however, the public was more likely than physicians to believe that physicians should be held responsible (66.0% vs 57.3%; P = .020) (Figure 1, Supplementary Table 2). Compared with the public, physicians were more likely to believe that the company or vendor selling the algorithm (43.8% vs 32.9%; P = .004) and the healthcare organization purchasing it (29.2% vs 22.6%; P = .05) should be liable. There was no difference in the degree to which physicians and the public believed that the FDA or other governmental entity should be held responsible.

Figure 1.

Physician vs public perceptions of liable parties when medical errors result from use of artificial intelligence in health care. The chi-square test was performed to determine whether physicians’ and patients’ attitudes differed significantly regarding the liable parties for medical errors occurring when physicians and the computer program work together to treat patients. The x-axis includes the 4 parties that physicians or patients think should be liable for the medical error occurred when physicians and the computer program work together to treat patients. The 4 potentially liable parties are (1) the physician making the decision, (2) the company or vendor that provides the algorithm, (3) the healthcare organization purchasing the algorithm, and (4) the Food and Drug Administration (FDA) or governmental entity approving the algorithm for clinical use.

Among the public, younger individuals (<50 years of age) were more likely to believe that the physician (70.9% vs 60.6%; P < .001), company or vendor (37.4% vs 27.7%; P = .004), and FDA (28.1% vs 18.1%; P < .001) were to blame when a medical error occurs involving AI (Table 1). There were no significant differences in perceptions of liable parties among physicians by specialty or years in practice (Supplementary Table 1).

Table 1.

Physician perceptions (by specialty) and public perceptions (by age) of liable parties related to use of artificial intelligence in health care

Liable Party	Physician Perceptions by Specialty				Public Perceptions by Age
	Internists (n = 78)	Radiologists (n = 51)	Oncologists (n = 63)	P Valuea	18-49 y (n = 537)	≥50 y (n = 469)	P Valueb
Physician making the decision				.39			<.001
Yes	49 (62.8)	35 (55.6)	26 (51.0)		381 (70.9)	284 (60.6)
No	29 (37.2)	28 (44.4)	25 (49.0)		156 (29.1)	185 (39.4)
Company/vendor that provides the algorithm				.13			.004
Yes	38 (48.7)	21 (33.3)	25 (49.0)		201 (37.4)	130 (27.7)
No	40 (51.3)	42 (66.7)	26 (51.0)		336 (62.6)	339 (72.3)
Healthcare organization purchasing the algorithm				.18			.084
Yes	25 (32.1)	13 (20.6)	18 (35.3)		136 (25.3)	92 (19.6)
No	53 (67.9)	50 (79.4)	33 (64.7)		401 (74.7)	377 (80.4)
FDA or governmental entity approving the algorithm for clinical use				.11			<.001
Yes	21 (26.9)	9 (14.3)	15 (29.4)		151 (28.1)	85 (18.1)
No	57 (73.1)	54 (85.7)	36 (70.6)		386 (71.9)	384 (81.9)

Values are n (%). Percentages may not total 100 because of rounding.

FDA: Food and Drug Administration.

^aThe chi-square test was performed to determine whether physicians’ attitudes differed significantly by specialty regarding the liable parties for medical errors occurring when physicians and the computer program work together to treat patients.

^bThe chi-square test was performed to determine whether patients’ attitudes differed significantly by age regarding the liable parties for medical errors occurring when physicians and the computer program work together to treat patients.

DISCUSSION

The public is significantly more likely to believe that physicians should be held responsible when an error occurs during care jointly delivered by physicians and algorithms, though the majority of both groups hold this view. Physicians are more likely than the public to believe that vendors and healthcare organizations should be liable; physicians’ views of medical liability do not seem to differ by specialty or number of years in practice. Among the public, younger individuals were significantly more likely to hold physicians, vendors, and the FDA responsible for medical errors resulting from the use of AI compared with older individuals.

To our knowledge, this is the first study comparing physician and public views of medical liability around the use of AI in health care. Older studies on medical errors in general have found that both physicians and the public are concerned about them, but that the public is more likely to believe that they are a major problem.¹² Prior work has also found that physicians and patients have different views on how and when medical errors should be disclosed, but these issues have not been examined as they relate to AI.¹³

Several recent studies have examined physicians’ and the public’s views on other issues related to medical AI. For example, a recent study in the United Kingdom found skepticism among general practitioners that AI could replace many clinical functions but found optimism that it could reduce administrative burden.¹⁴ Recent studies using focus groups in Canada have found that people are generally supportive of medical AI, though they are concerned about privacy and security issues, as well as about commercial motives and loss of human connection with clinicians.^15,16 However, none of these studies specifically addressed liability concerns.

Our study focused on the normative question of who should be liable for AI-related medical errors, not on how the medico-legal system currently approaches these issues. Indeed, there is little case law on liability for errors that occur when physicians use AI to deliver care, and it is not yet clear how these issues will be adjudicated, as no major cases related to the use of medical AI have been decided by the courts. But several factors are likely to be relevant. First, tort law generally favors the accepted standard of care: physicians are less likely to be held liable in cases in which AI recommends—and they pursue—the standard of care, regardless of clinical outcome.¹⁷ Such a system incentivizes conservative applications of AI and may reduce the impact of algorithms that deviate from standard practice, even when their recommendations could produce better outcomes.⁴ Second, determinations of liability will likely be influenced by who created the algorithm used.¹⁸ For example, if a physician uses AI developed by his or her own healthcare organization, the organization may face a higher risk of liability through “enterprise liability,” which holds hospitals and medical groups, instead of individual physicians, accountable for errors.¹⁹ However, if the algorithm was purchased from an external vendor, legal liability may shift to that company. In some cases, FDA approval shields a company from liability; in other cases, however, it does not, especially if a less rigorous, expedited approval pathway is used, as is currently the case for many AI algorithms.^20–22 Going forward, physicians may have several avenues to influence AI use and policy through interactions with their healthcare organizations, professional societies, and malpractice insurers, but it is less clear how patient views will be incorporated.

Our study has limitations. First, our study focused on 3 physician specialties likely to be affected by AI in health care, but views of other specialties may differ. Second, our response rate for physicians was relatively low, which may limit generalizability, and respondents were more likely to be male and internists compared with nonrespondents. Third, given that medical AI is relatively nascent, stakeholder views may evolve as various parties learn more about its applications and have personal experience with medical errors resulting from its use.

CONCLUSION

Legal and malpractice considerations related to AI in health care are evolving, and the FDA and the National Academy of Medicine recently highlighted understanding stakeholder views as a key aspect of developing medical AI policy.^6,23 The data presented in this study may help inform clinical practice, regulatory policy, and healthcare organizations’ use of AI. Healthcare leaders should be aware of differences in perceptions of liability that exist between physicians and patients, and consider avenues—through education, communication, and design and use of AI—to bridge this disconnect. Future work should examine the reasons for observed differences in perceptions of liability and track physician and public attitudes over time.

FUNDING

This work was supported by the Physicians Foundation Center for the Study of Physician Practice and Leadership at Weill Cornell Medicine and the Agency for Healthcare Research and Quality (5K12HS023000 [to SA]). The Physicians Foundation and Agency for Healthcare Research and Quality had no role in the design and conduct of the study; collection, management, analysis, and interpretation of the data; preparation, review, or approval of the manuscript; or the decision to submit the manuscript for publication.

AUTHOR CONTRIBUTIONS

DK, LPC, and SA conceived of the study, supervised its conduct, and oversaw data collection; DK and YQ analyzed the data; DK wrote the first draft of the article. All the authors participated in important discussions about the project, contributed to the revision of the manuscript, and approved the article for submission.

SUPPLEMNTARY MATERIAL

Supplementary material is available at Journal of the American Medical Informatics Association online.

CONFLICT OF INTEREST STATEMENT

DK has received grants from the American Medical Association, outside this work. LPC has received grants from the American Medical Association, Arnold Ventures, and the American Board of Family Medicine, outside this work. SA has received grants from the National Cancer Institute, American Cancer Society, American Society for Clinical Oncology, and National Science Foundation, outside this work. No other disclosures were reported.