Abstract
Objectives
Although efforts to effectively govern AI continue to develop, relatively little work has been done to systematically measure and include patient perspectives or expectations of AI in governance. This analysis is designed to understand patient expectations of healthcare AI.
Materials and Methods
Cross-sectional nationally representative survey of US adults fielded from June to July of 2023. A total of 2039 participants completed the survey and cross-sectional population weights were applied to produce national estimates.
Results
Among US adults, 19.55% expect AI to improve their relationship with their doctor, while 19.4% expect it to increase affordability and 30.28% expect it will improve their access to care. Trust in providers and the healthcare system are positively associated with expectations of AI when controlling for demographic factors, general attitudes toward technology, and other healthcare-related variables.
Discussion
US adults generally have low expectations of benefit from AI in healthcare, but those with higher trust in their providers and health systems are more likely to expect to benefit from AI.
Conclusion
Trust and provider relationships should be key considerations for health systems as they create their AI governance processes and communicate with patients about AI tools. Evidence of patient benefit should be prioritized to preserve or promote trust.
Keywords: artificial intelligence, patient trust, access to care, patient-provider relationships, AI governance
Introduction
Background and significance
Interest in artificial intelligence (AI) has increasingly been accompanied by efforts to design guidelines for how to use the technology safely and effectively. Various collaboratives, partnerships, and frameworks emphasize ethics, trustworthiness, and fairness.1–3 These efforts span healthcare sectors, state governments, national institutes, and various federal bodies.3–6 Although guidelines for effectively designing, evaluating, and governing AI continue to develop,7–9 relatively little work has been done at the national level to systematically measure and include patient perspectives or expectations of AI in their healthcare.10–12
In order to ensure that healthcare AI conforms to principles of patient-centered care, empirical evidence of patient perspectives and expectations is fundamentally necessary.1,12 Eliciting patient preferences has been proposed to provide important insights about human values and priorities for AI in clinical care.13 Complementing this orientation to AI that responds to risk and patient preferences, evidence on population-level perceptions of AI in healthcare would provide much needed information to ensure that the necessary frameworks, regulations, and guidelines are in place to safeguard public trust in the healthcare system.14,15
Multiple federal guidelines and directives emphasize the need for effective governance to prevent the violation of public trust with AI.1,16 Transparency, privacy protections, and rigorous evaluation have been identified as important aspects of trustworthy AI use in healthcare.3,17,18 Prior research on patient perspectives has produced early insights into specific AI applications like dermatology10 or acceptability of AI in the care of pregnant patients.11 A previous national survey identified public concern about the use of large language models (LLMs) like ChatGPT in healthcare, finding a relationship between expected benefit from LLMs in healthcare and comfort with the use of that technology.12 If patients expect that other AI applications will provide benefit to them, it is possible that they will exhibit higher comfort with AI in their care. However, if those expectations are unmet or inflated beyond what the healthcare system can provide, there may be negative consequences for trust. Low expectations for AI may indicate a need for greater patient engagement and transparency to prevent negative consequences of AI use for patient trust.
Objective
Building a trustworthy and patient-centered AI ecosystem requires a deeper understanding of how the public perceives healthcare AI. To facilitate this, we conducted a national survey of the US public focused on identifying what they expect from AI in healthcare and how their expectations relate to trust. First, we ask how the public expects AI to affect their healthcare across key dimensions of potential benefit: access to care, personal relationships with their doctors, and affordability. Second, we explore the relationship between trust in the healthcare system and public expectations of AI. Third, we analyze the relationship between trust in healthcare providers and AI expectations.
Materials and methods
Data
This study uses survey data from a nationally representative sample of US adults, aged 18 and older. The survey was fielded online from June to July 2023 by the National Opinion Research Center at the University of Chicago (NORC) using the AmeriSpeak Panel after pre-testing with a small sample (n = 116). A total of 2039 participants completed the final survey. Black respondents and Hispanic respondents were oversampled to ensure sufficient representation for subpopulation analyses. NORC calculated post stratification survey weights based on the 2023 Current Population Survey and nonresponse to produce national estimates.
Participants were remunerated according to standard NORC policies. Informed consent was obtained by NORC. The study protocol was approved by the NORC IRB. Participants viewed a short (90-s) explanatory video describing how AI can be applied in healthcare. The video was produced in collaboration with community partners and a board of national expert advisors. Similar to informational videos designed for multiple previous national surveys on health IT, predictive models, and precision oncology,12,14,19 the video was tested with focus groups to ensure balanced representation of various concepts and accessible terminology for a public audience. It was also reviewed by a panel of national experts in health information technology, healthcare delivery, biomedical informatics, and ethics.
Outcomes and measures
The outcome of this analysis was a composite measure of three survey questions measuring expectations of AI in healthcare. On a 4-point Likert scale (1 = not true, 4 = very true), respondents indicated how true each of the following three statements were for them: The use of AI in healthcare will (1) improve my relationship with my doctor, (2) make it easier to get the healthcare I need, and (3) make my healthcare more affordable. Correlation coefficients for these survey items were ≥ 0.6. Cronbach’s alpha, the measure of internal consistency, was 0.83. This was above the common 0.70 threshold indicating that these measures were internally consistent.20 The mean of these three items was used to create a binary measure of expectations of AI, set to 1 for high expectations (composite score mean > 2, range 1-4) and 0 for low expectations (composite score mean ≤ 2).
Trust in the healthcare system was measured with an index based on prior work and the larger literature on trust.19,21,22 System trust was measured with 13 survey items previously used to understand how individuals feel about placing trust in the healthcare system. Respondents indicated how true each of the 13 statements were on a 4-point Likert scale (1 = not true, 4 = very true). The 13 items reflected four key dimensions of trust in health systems: integrity, fidelity, global trust, and competence (see Appendix S1 for additional detail).22 These dimensions each represented a sub-index of the overall system trust index, calculated as the sum of the means of each dimension of trust. For each respondent, the 4 trust dimension indices were summed to create the overall system trust score (range 4-16) with a higher score indicating greater trust in the healthcare system.
Trust in healthcare providers was measured with two survey items. Respondents indicated how true each statement about healthcare providers was for them: (1) I trust healthcare providers to use my health information responsibly, and (2) All things considered, healthcare providers in this country can be trusted. Responses were on the 4-point Likert scale (1 = not true, 4 = very true). The average score of provider trust was calculated. The binary provider trust variable was set to 1 if the respondent’s provider trust score was greater than 2 (range 1-4).
Analytic approach
We analyzed all responses with complete data (n = 1834). We conducted weighted multivariable logistic regressions to assess relationships between public expectations of AI in healthcare and demographic variables (eg, sex, age, race/ethnicity), health-related variables (eg, health insurance coverage, ability to get needed care), and technology-related variables (eg, tech-savviness, concern about over-reliance on technology). These weighted multivariable regressions were conducted to produce estimates that are representative of the US adult population. We defined statistical significance as P < .05 and conducted all analyses using RStudio 2023.12.1 Build 402.
Results
Reporting weighted percentages, 50.87% of respondents were female and 49.13% were male (Table 1). Most self-identified as white (63.86%). Black respondents made up 12.1% of the sample, and Hispanic respondents were 17.16% of the sample. Most indicated low expectations of AI in healthcare (69.09%). Most respondents (86.85%) did not require assistance with reading or understanding health-related information (health literacy) and slightly under half (43.04%) described themselves as tech-savvy.
Table 1.
Descriptive statistics.
| Unweighted n | Weighted % | |
|---|---|---|
| AI expectations | ||
| Low | 1258 | 69.09 |
| High | 576 | 30.91 |
| System trust score | ||
| Low | 946 | 54.35 |
| High | 888 | 45.65 |
| Provider trust | ||
| Low | 949 | 52.14 |
| High | 885 | 47.86 |
| Sex | ||
| Female | 954 | 50.87 |
| Male | 880 | 49.13 |
| Age | ||
| 18-29 | 282 | 19.93 |
| 30-44 | 543 | 26.72 |
| 45-59 | 424 | 22.73 |
| 60+ | 585 | 30.61 |
| Race/ethnicity | ||
| White, non-Hispanic | 796 | 63.86 |
| Black, non-Hispanic | 481 | 12.10 |
| Hispanic | 469 | 17.16 |
| Asian, non-Hispanic | 46 | 4.67 |
| Other and multiracial | 42 | 2.20 |
| Education | ||
| Less than high school | 125 | 8.89 |
| High school graduate or equivalent | 299 | 28.64 |
| Some college/associate degree | 748 | 26.43 |
| Bachelor's degree | 387 | 21.29 |
| Post grad study/professional degree | 275 | 14.74 |
| Need health literacy assistance | ||
| Never/Rarely | 1585 | 86.85 |
| Sometimes | 187 | 10.25 |
| Often/Always | 62 | 2.90 |
| Experience of discrimination | ||
| No | 1380 | 75.07 |
| Yes | 454 | 24.93 |
| Health insurance coverage | ||
| No | 189 | 10.32 |
| Yes | 1645 | 89.68 |
| Can get care when needed | ||
| No | 751 | 41.14 |
| Yes | 1083 | 58.86 |
| Tech-savvy | ||
| No | 1049 | 56.96 |
| Yes | 785 | 43.04 |
| Reliance on technology is too high | ||
| No | 980 | 54.60 |
| Yes | 854 | 45.40 |
Among the three measures comprising the AI expectations composite score, patient expectations that AI will improve affordability (19.4%) were lowest. Patient expectations that AI will improve their relationships with their doctors were similarly low (19.55%). Expectations that AI will improve healthcare access were comparatively higher but remained low (30.28%) (Figure 1).
Figure 1.
Public expectations of AI in healthcare. Weighted percentages producing national estimates for the US adult population.
In weighted multivariable logistic regression (Table 2), we identify a significant positive association between trust in the healthcare system and AI expectations (OR 3.55, 95% CI 2.67-4.72). Respondents reporting high trust in their providers were also more likely to have higher expectation of benefit from AI (OR 1.78, 95% CI 1.27-2.51). Female respondents were less likely to report high AI expectations than male respondents (OR 0.57, 95% CI 0.43-0.76). Black respondents (OR 1.43, 95% CI 1.01-2.02) and Hispanic respondents (OR 1.46, 95% CI 1.02-2.08) were more likely to report higher expectations for AI than white respondents. Respondents who sometimes need assistance reading materials from their doctor were more likely to have high AI expectations than their counterparts who reported never or rarely needing assistance. Controlling for all covariates, self-reported tech-savviness was positively predictive of expected benefit from AI (OR 1.41, 95% CI 1.08-1.86). Respondents who could get the healthcare they need when they need it were also more likely to expect AI benefits (OR 1.44, 95% CI 1.04-1.97).
Table 2.
Weighted multivariable regression of AI expectations.
| Weighted |
|||
|---|---|---|---|
| OR | 95%CI | P-value | |
| System trust score | |||
| Low | ref. | ||
| High | 3.55 | (2.67, 4.72) | <0.001a |
| Provider trust | |||
| Low | ref. | ||
| High | 1.78 | (1.27, 2.51) | <0.001a |
| Health insurance coverage | |||
| No | ref. | ||
| Yes | 0.84 | (0.56, 1.28) | 0.42 |
| Experience of discrimination | |||
| No | ref. | ||
| Yes | 0.75 | (0.48, 1.15) | 0.18 |
| Get care when needed | |||
| No | ref. | ||
| Yes | 1.44 | (1.04, 1.97) | 0.03b |
| Need health literacy assistance | |||
| Never/Rarely | ref. | ||
| Sometimes | 2.83 | (1.77, 4.52) | <0.001a |
| Often/Always | 1.8 | (0.75, 4.31) | 0.19 |
| Tech-savvy | |||
| No | ref. | ||
| Yes | 1.41 | (1.08, 1.86) | 0.01b |
| Reliance on technology | |||
| No | ref. | ||
| Yes | 0.79 | (0.58, 1.07) | 0.13 |
| Age | |||
| 18-29 | ref. | ||
| 30-44 | 1.52 | (0.94, 2.48) | 0.09 |
| 45-59 | 1.37 | (0.85, 2.2) | 0.2 |
| 60+ | 1.02 | (0.62, 1.69) | 0.93 |
| Sex | |||
| Male | ref. | ||
| Female | 0.57 | (0.43, 0.76) | 0.001a |
| Race/ethnicity | |||
| White, non-Hispanic | ref. | ||
| Asian, non-Hispanic | 1.68 | (0.74, 3.8) | 0.21 |
| Black, non-Hispanic | 1.43 | (1.01, 2.02) | 0.04b |
| Hispanic | 1.46 | (1.02, 2.08) | 0.04b |
| Other identities and multiracial | 1.17 | (0.41, 3.37) | 0.77 |
| Education | |||
| Less than high school | ref. | ||
| High school graduate or equivalent | 2.15 | (1.16, 3.99) | 0.02b |
| Some college/associate degree | 1.29 | (0.66, 2.52) | 0.46 |
| Bachelor's degree | 1.23 | (0.64, 2.35) | 0.53 |
| Post grad study/professional degree | 1.2 | (0.55, 2.63) | 0.65 |
P < .001.
P < .05.
Discussion
Improved accessibility, affordability, and patient-provider relationships are often described as potential benefits of AI in healthcare through reducing clinician documentation burden or producing cost savings.23,24 Our findings from analysis of nationally representative survey data indicate that the US public generally does not share those expectations. Only 19.4% expect that AI will improve affordability of healthcare, while 19.55% believe that it will improve their relationships with their doctors and 30.28% believe it will improve their access to care. This misalignment highlights the need for rigorous evaluation of how AI will affect patients. Rather than anticipation and excitement,25,26 there is a need for evidence and communication about under what circumstances AI can benefit patients.
We also identify the importance of patient trust, in both systems and providers, for AI expectations. Patients with higher trust are significantly more likely to believe that AI will positively impact their care. Those with low system trust (54.35%) and low provider trust (52.14%) report lower expectations of AI. There are two key implications of this finding. First, if patients with high expectations are disappointed by AI there could be negative consequences for their trust in both the healthcare system and their providers. Similarly, Black respondents and Hispanic respondents report higher expectations of AI. Meeting these expectations will be important as health systems develop AI governance processes to protect or earn patient trust in the context of unequal healthcare delivery.27,28 Second, implementation of AI without consultation or patient engagement could further threaten the trust of patients who do not find the system or providers to be trustworthy. For those who struggle to access the care they need, there is a low expectation of AI improving their experience of the healthcare system.
Our analysis identifies higher AI expectations among male respondents compared to female respondents when controlling for demographic, healthcare, and technology-related variables. This reflects prior literature where self-identified female survey respondents have reported lower comfort with the use of ChatGPT in healthcare and the use of predictive models.12,14,29 Further investigation of these observed differences will be necessary to better understand what is driving this observed comfort gap between male and female survey respondents with AI technologies in healthcare.
Limitations
It is possible that some respondents interpreted a positive or negative valence in the informational video explaining AI in healthcare, which could have biased their responses. However, multiple methods were deployed to minimize bias in the video content including focus group feedback and expert review from a wide range of stakeholders. An additional limitation is related to the measurement of provider trust. Because trust is multidimensional, our approach using multiple dimensions of system trust is rigorous and conceptually strong. However, the measure of provider trust, despite containing multiple survey items, does not contain as multidimensional. Although survey length presents challenges for response rate and survey completion, future work would build on the findings reported here by developing the concepts of dimensions of provider trust and creating related survey measures.
Conclusions
In this national survey of US adults, we find low expectations of benefit from AI. There is misalignment between the anticipated benefits of AI, especially for operational and administrative purposes, between the healthcare system and the public. This may indicate a need for greater patient engagement and transparency to prevent negative consequences of AI use. Our finding that system trust and provider trust are positively predictive of AI expectations indicate that trust and patient-provider relationships should be key priorities for health systems as they create their AI governance processes and communicate with patients about their use of AI.
Supplementary Material
Acknowledgments
The authors would like to thank Dr Jodyn Platt for data acquisition.
Contributor Information
Paige Nong, Division of Health Policy and Management, School of Public Health, University of Minnesota, Minneapolis, MN 55455, United States.
Molin Ji, Division of Health Policy and Management, School of Public Health, University of Minnesota, Minneapolis, MN 55455, United States.
Author contributions
Paige Nong (Conceptualization, Data curation, Investigation, Methodology, Project administration, Supervision, Writing – original draft, Writing – review & editing) and Molin Ji (Formal analysis, Methodology, Project administration, Visualization)
Supplementary material
Supplementary material is available at Journal of the American Medical Informatics Association online.
Funding
This work was supported by the National Institutes of Health, The National Institute of Biomedical Imaging and Bioengineering (NIBIB) [grant number 5R01EB030492-04].
Conflicts of interest
None declared.
Data availability
Data is available upon reasonable request to the corresponding author.
References
- 1.Trustworthy AI (TAI) Playbook [Internet]. US Department of Health and Human Services; 2021. https://www.hhs.gov/sites/default/files/hhs-trustworthy-ai-playbook.pdf
- 2.Blueprint for an AI Bill of Rights: making automated systems work for the American people [Internet]. The White House: The Office of Science and Technology Policy; 2022, p. 73. https://www.whitehouse.gov/wp-content/uploads/2022/10/Blueprint-for-an-AI-Bill-of-Rights.pdf
- 3. Everson J, Smith J, Marchesini K, Tripathi M. A regulation to promote responsible AI in health care. Health Aff Forefr [Internet]. 2024. Accessed February 28, 2024. https://www.healthaffairs.org/do/10.1377/forefront.20240223.953299/full/
- 4. Hill H, Roadevin C, Duffy S, Mandrik O, Brentnall A. Cost-effectiveness of AI for risk-stratified breast cancer screening. JAMA Netw Open. 2024;7:e2431715. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 5. Ong JCL, Chang SY-H, William W, et al. Medical ethics of large language models in medicine. NEJM AI. 2024;1:AIra2400038. [Google Scholar]
- 6. Tabassi E. AI Risk Management Framework: AI RMF (1.0). National Institute of Standards and Technology; 2023, p. NIST AI 100-1. Report No.: NIST AI 100-1. Accessed February 27, 2024. https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.100-1.pdf [Google Scholar]
- 7. Bedoya AD, Economou-Zavlanos NJ, Goldstein BA, et al. A framework for the oversight and local deployment of safe and high-quality prediction models. J Am Med Inform Assoc. 2022;29:1631-1636. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 8. Callahan A, McElfresh D, Banda JM, et al. Standing on FURM ground: a framework for evaluating fair, useful, and reliable AI models in health care systems. NEJM Catal. 2024;5:18. [Google Scholar]
- 9. de Hond AAH, Leeuwenberg AM, Hooft L, et al. Guidelines and quality criteria for artificial intelligence-based prediction models in healthcare: a scoping review. NPJ Digit Med. 2022;5:2-13. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 10. Nelson CA, Pérez-Chada LM, Creadore A, et al. Patient perspectives on the use of artificial intelligence for skin cancer screening: a qualitative study. JAMA Dermatol. 2020;156:501-512. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 11. Armero W, Gray KJ, Fields KG, Cole NM, Bates DW, Kovacheva VP. A survey of pregnant patients’ perspectives on the implementation of artificial intelligence in clinical care. J Am Med Inform Assoc. 2022;30:46-53. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 12. Platt J, Nong P, Smiddy R, et al. Public comfort with the use of ChatGPT and expectations for healthcare. J Am Med Inform Assoc. 2024;31:1976-1982. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 13. Yu KH, Healey E, Leong TY, Kohane I, Manrai A. Medical artificial intelligence and human values. N Engl J Med. 2024;390:1895-1904. Accessed October 2, 2024. https://oce-ovid-com.ezp3.lib.umn.edu/article/00006024-202405300-00009/HTML [DOI] [PMC free article] [PubMed] [Google Scholar]
- 14. Nong P, Adler-Milstein J, Platt J. How patients distinguish between clinical and administrative predictive models in health care. Am J Manag Care. 2024;30:31-37. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 15. Bakken S. Centering the patient in informatics applications. J Am Med Inform Assoc. 2022;29:1027-1028. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 16.Executive Order on the Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence [Internet]. The White House; 2023. Accessed March 5, 2024. https://www.whitehouse.gov/briefing-room/presidential-actions/2023/10/30/executive-order-on-the-safe-secure-and-trustworthy-development-and-use-of-artificial-intelligence/
- 17. Roski J, Maier EJ, Vigilante K, Kane EA, Matheny ME. Enhancing trust in AI through industry self-governance. J Am Med Inform Assoc. 2021;28:1582-1590. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 18. Reddy S, Allan S, Coghlan S, Cooper P. A governance model for the application of AI in health care. J Am Med Inform Assoc. 2020;27:491-497. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 19. Platt J, Raj M, Büyüktür AG, et al. Willingness to participate in health information networks with diverse data use: evaluating public perspectives. EGEMS (Wash DC). 2019;7:33. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 20. Taber KS. The use of Cronbach’s Alpha when developing and reporting research Instruments in science education. Res Sci Educ. 2018;48:1273-1296. [Google Scholar]
- 21. Taylor LA, Nong P, Platt J. Fifty years of trust research in health care: a synthetic review. Milbank Q. 2023;101:1-53. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 22. Platt JE, Jacobson PD, Kardia SLR. Public trust in health information sharing: a measure of system trust. Health Serv Res. 2018;53:824-845. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 23. Tierney AA, Gayre G, Hoberman B, et al. Ambient artificial intelligence scribes to alleviate the burden of clinical documentation. NEJM Catal. 2024;5:1-15. [Google Scholar]
- 24. Cutler DM. What artificial intelligence means for health care. JAMA Health Forum. 2023;4:e232652. [DOI] [PubMed] [Google Scholar]
- 25. Kulkarni PA, Singh H. Artificial intelligence in clinical diagnosis: opportunities, challenges, and hype. JAMA. 2023;330:317-318. [DOI] [PubMed] [Google Scholar]
- 26. Suran M, Hswen Y. How to navigate the pitfalls of ai hype in health care. JAMA. 2024;331:273-276. [DOI] [PubMed] [Google Scholar]
- 27. Boulware LE, Cooper LA, Ratner LE, LaVeist TA, Powe NR. Race and trust in the health care system. Public Health Rep. 2003;118:358-365. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 28. Jaiswal J. Whose responsibility is it to dismantle medical mistrust? future directions for researchers and health care providers. Behav Med. 2019;45:188-196. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 29. Gonzalez XT, Steger-May K, Abraham J. Just another tool in their repertoire: uncovering insights into public and patient perspectives on clinicians’ use of machine learning in perioperative care. J Am Med Inform Assoc. 2025;32:150-162. [DOI] [PMC free article] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Supplementary Materials
Data Availability Statement
Data is available upon reasonable request to the corresponding author.