Abstract
General-purpose large language models (LLMs) like ChatGPT are increasingly used for medical advice despite lacking medical training and frequently producing incorrect or unsafe output. Older adults’ health information-seeking behaviors using LLMs remain poorly characterized. We conducted a cross-sectional survey of 574 US adults aged 50+ recruited via Prolific, balanced by sex and race. Participants reported health information sources, ChatGPT and PubMed use, demographics, and health literacy. Most participants (92%) searched online for health information. All had heard of ChatGPT, and 63% used it for medical information, compared to 44% who had heard of PubMed and 39% who used it. Those with inadequate health literacy had higher odds of ChatGPT use for medical advice (AOR 2.36, 95% CI 1.30–4.52) versus those with adequate health literacy. In conclusion, with more than half of older adults using LLMs for medical advice, the development of safer, purpose-trained medical LLMs is warranted.
Introduction
Health information-seeking behaviors, referring to ways of gathering health-related information, are associated with the ability to self-manage health and, ultimately, health outcomes.1 Adults in the U.S. may use general-purpose large language models (LLMs), such as ChatGPT, for medical advice.2 LLMs provide natural language responses to user queries that are easy to understand, but are not trained or tested for medical purposes. One review found 88% of included studies evaluating LLMs for medical purposes reported incorrect or incomplete output, and nearly half reported unsafe output.3 Chronic conditions affect 80–90% of adults age 35 and older,4 creating substantial medical information needs. However, older adults’ health information-seeking behaviors in the new era of broadly accessible LLMs have not been well characterized. This study investigated health information-seeking behaviors, including LLM use, among US adults aged 50 and older.
Methods
This cross-sectional survey study used Prolific to recruit participants balanced by sex and reflecting 2020 US Census racial distribution (58% White, 14% Black, 6% Asian, 20% other, and 2% mixed). Eligible participants could read English and were aged 50 and older.
Survey items were adapted from the Health Information National Trends Survey and a separate survey assessing online health information search habits and sources.5 We asked about familiarity with and frequency of using ChatGPT (selected because it is the most commonly used LLM with over 700 million users as of September 2025). To contrast this, we also asked about their use of PubMed (a federally supported biomedical library). Participants self-reported demographics and health literacy, completing surveys via Qualtrics. Quality was audited through attention checks and response times. Participants completing the questionnaire and passing attention checks received $16 per hour. The Columbia University Institutional Review Board approved this study.
We used descriptive statistics and logistic regression to examine demographic effects on ChatGPT use for health information (binarized as never versus any use). Analyses were performed using R version 4.1.2.
Results
Of 622 participants recruited, 48 were excluded (33 age ineligible, 9 incomplete surveys, 6 failed attention checks; 0 excluded for low response times), yielding 574 participants. Mean age was 58 (SD 6.8) years; 51% were female, 70% White, 16% Black or African American, and 10% Hispanic or Latino. Education levels were 30% high school or less, 38% Bachelor’s degree, and 32% graduate degree; 71% had adequate health literacy. Most (93%) used the Internet more than once daily.
Most participants (92%) used the Internet to search for health information (Table 1). All but two had heard of ChatGPT (100%), and 63% used it at least once for medical information. Fewer had heard of PubMed (44%) or used it (39%). Other common sources included licensed healthcare professionals (93%) and medical websites (75%). Fewer used government health agencies (40%), news websites (16%), and social media (15%).
Table 1.
Health information seeking among survey participants (n=574)
| Tried looking for information about your health from any source | 553 (96.3%) |
| Used the Internet for these health-related needs in the last 12 months | |
| Look for health or medical information | 525 (91.5%) |
| View medical test results | 349 (60.8%) |
| Made an appointment with a health care provider | 331 (57.7%) |
| Send a message to a healthcare provider | 325 (56.6%) |
| Heard of ChatGPT | 572 (100%) |
| If yes, frequency of ChatGPT use for medical information | |
| Often | 99 (17.2%) |
| Sometimes | 154 (26.8%) |
| Once or twice | 111 (19.3%) |
| Never | 208 (36.%) |
| N/A- have not used ChatGPT | 2 (0.3%) |
| Heard of the online biomedical library, PubMed | 254 (44.3%) |
| If yes, frequency of PubMed use for medical information | |
| Often | 27 (4.7%) |
| Sometimes | 116 (20.2%) |
| Once or twice | 81 (14.1%) |
| Never | 30 (5.2%) |
| N/A- have not used PubMed | 320 (55.7%) |
| Sources relied upon for information about personal health | |
| A licensed healthcare professional (i.e., doctor or nurse) | 531 (92.5%) |
| Medical website (i.e., WebMD) | 428 (74.6%) |
| Scientific researchers | 336 (58.5%) |
| Government health agencies | 232 (40.4%) |
| Other people with the same health condition | 193 (33.6%) |
| Family, friend, or community member | 177 (30.8%) |
| News website | 93 (16.2%) |
| Social media (i.e., Facebook or Instagram) | 87 (15.2%) |
| Religious organizations and leaders | 15 (2.6%) |
In adjusted models (Table 2), Asian participants had lower odds of using ChatGPT for medical information versus White participants (adjusted odds ratio [AOR] 0.36, 95% confidence interval [CI] 0.16–0.78). Participants with graduate degrees (AOR 2.57, 95% CI 1.51–4.43) and Bachelor’s degrees (AOR 1.66, 95% CI 1.01–2.72) had higher odds of using ChatGPT versus those with high school education or less. Those with inadequate health literacy had higher odds of ChatGPT use (AOR 2.36, 95% CI 1.30–4.52) versus those with adequate health literacy.
Table 2:
Differences in frequency of ChatGPT use for medical advice by sociodemographic characteristics, n=574
| Never (n=208) | Any use (n=364) | Unadjusted OR (95% CI) | p value | Adjusted OR (95% CI) | p value | |
|---|---|---|---|---|---|---|
| Age | ||||||
| Reference: 50–59 | 122 (33%) | 246 (67%) | — | — | — | — |
| 60–69 | 67 (41%) | 95 (59%) | 0.70 (0.48–1.03) | 0.070 | 0.76 (0.48–1.21) | 0.247 |
| 70+ | 19 (45%) | 23 (55%) | 0.60 (0.32–1.15) | 0.121 | 0.76 (0.36–1.63) | 0.476 |
| Sex | ||||||
| Reference: Female | 113 (39%) | 179 (61%) | — | — | — | — |
| Male | 95 (34%) | 185 (66%) | 1.23 (0.87–1.73) | 0.236 | 1.04 (0.69–1.57) | 0.845 |
| Race | ||||||
| Reference: White | 148 (37%) | 254 (63%) | — | — | — | — |
| Black or African American | 19 (21%) | 72 (79%) | 2.21 (1.30–3.90) | 0.004 | 1.83 (0.98–3.55) | 0.063 |
| Asian | 20 (53%) | 18 (47%) | 0.52 (0.27–1.02) | 0.058 | 0.36 (0.16–0.78) | 0.010 |
| American Indian or Alaska Native | 7 (47%) | 8 (53%) | 0.67 (0.23–1.93) | 0.441 | 0.53 (0.15–1.83) | 0.316 |
| Native Hawaiian or Other Pacific Islander* | 0 (0%) | 1 (100%) | — | — | — | — |
| Prefer not to answer* | 14 (56%) | 11 (44%) | — | — | — | — |
| Ethnicity | ||||||
| Reference: not Hispanic/Latino | 191 (38%) | 314 (62%) | — | — | — | — |
| Hispanic/Latino | 16 (29%) | 40 (71%) | 1.52 (0.84–2.86) | 0.176 | 1.67 (0.78–3.86) | 0.206 |
| Prefer not to answer* | 1 (9%) | 10 (91%) | — | — | — | — |
| Education | ||||||
| Reference: High school or less | 85 (50%) | 86 (50%) | — | — | — | — |
| Bachelor’s degree | 78 (36%) | 139 (64%) | 1.76 (1.17–2.65) | 0.007 | 1.66 (1.01–2.72) | 0.045 |
| Graduate degree (Master’s or higher) | 44 (24%) | 139 (76%) | 3.12 (2.00–4.94) | <0.001 | 2.57 (1.51–4.43) | <0.001 |
| Prefer not to answer* | 1 (100%) | 0 (0%) | — | — | — | — |
| Finances | ||||||
| Reference: More than enough | 22 (39%) | 35 (61%) | — | — | — | — |
| Enough | 95 (28%) | 245 (72%) | 1.62 (0.89–2.89) | 0.105 | 1.78 (0.92–3.43) | 0.086 |
| Not enough | 83 (52%) | 77 (48%) | 0.58 (0.31–1.07) | 0.087 | 0.65 (0.31–1.34) | 0.244 |
| Prefer not to answer* | 8 (53%) | 7 (47%) | — | — | — | — |
| Health Literacy | ||||||
| Reference: Adequate | 185 (40%) | 275 (60%) | — | — | — | — |
| Inadequate | 23 (21%) | 89 (79%) | 3.28 (1.93–5.89) | <0.001 | 2.36 (1.30–4.52) | 0.007 |
| Health insurance coverage in the last year | ||||||
| Reference: Yes, for full year | 185 (39%) | 294 (61%) | — | — | — | — |
| No/ Part of the year | 21 (24%) | 65 (76%) | 1.95 (1.17–3.36) | 0.013 | 2.03 (0.93–4.87) | 0.091 |
| Prefer not to answer* | 2 (29%) | 5 (71%) | — | — | — | — |
| Type of health insurance of in the past year | ||||||
| Reference: Medicaid | 58 (34.3%) | 111 (65.6%) | — | — | — | — |
| Other than Medicaid | 150 (37.2%) | 253 (62.8%) | 0.87 (0.59–1.27) | 0.483 | 0.73 (0.46–1.17) | 0.198 |
Models did not converge due to too few observations.
Discussion
Over 60% of adults aged 50 and older have used ChatGPT for medical information, with use varying by race, education, and health literacy. By contrast, our recent Prolific study of younger participants (mean age 39) reported lower rates (33%) of ChatGPT use for medical information.6 Findings align with our prior work showing that adults with lower healthy literacy are more open to using AI-based tools in their care.7 A limitation of this study includes that we recruited participants via an online sample, who may have been predisposed to using AI-based tools and may differ from other adults aged 50 and older. Future work should develop purpose-trained medical LLMs that prioritize safety and usability for older adults with low health literacy.
Funding:
Dr. Turchioe and Ms. Shamnath report funding from NIH under award number R00NR019124. Dr. Benda reports funding from NIH under award number R00MD015781.
Sponsor’s Role:
The sponsor had no role in the design or conduct of the study, including data collection, management, analysis, or interpretation of the data. The sponsor did not participate in the preparation, review, or approval of the manuscript, and had no influence on the decision to submit the manuscript for publication.
Footnotes
Conflicts of interest: Dr. Turchioe reports equity in Iris OB Health LLC and consulting fees from Boston Scientific. Dr. Xu reports salary support from Google. The remaining authors have no conflicts to declare.
Human ethics and consent to participate declarations: The Columbia University Institutional Review Board approved this study (IRB AAAU2028). Participants read an information sheet and consented to participate in this specific study before completing the survey.
Data sharing:
Data, statistical code, and the informed consent document will be made available upon reasonable request after completing a data use agreement.
References
- 1.Nutakor JA, Zhou L, Larnyo E, Addai-Dansoh S, Cui Y. Impact of health information seeking behavior and digital health literacy on self-perceived health and depression symptoms among older adults in the United States. Sci Rep. 2024;14(1):31080. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 2.Ipsos. Can we rely on generative AI for healthcare information? 2024. Accessed August 5, 2025. https://www.ipsos.com/en-us/can-we-rely-generative-ai-healthcare-information [Google Scholar]
- 3.Busch F, Hoffmann L, Rueger C, et al. Current applications and challenges in large language models for patient care: a systematic review. Commun Med (Lond). 2025;5(1):26. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 4.Watson KB, Wiltz JL, Nhim K, Kaufmann RB, Thomas CW, Greenlund KJ. Trends in multiple chronic conditions among US adults, by life stage, Behavioral Risk Factor Surveillance System, 2013–2023. Prev Chronic Dis. 2025;22:E15. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 5.Korshakova E, Marsh JK, Kleinberg S. Health information sourcing and health knowledge quality: Repeated cross-sectional survey. JMIR Form Res. 2022;6(9):e39274. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 6.Mendel T, Singh N, Mann DM, Wiesenfeld B, Nov O. Laypeople’s use of and attitudes toward large language models and search engines for health queries: Survey study. J Med Internet Res. 2025;27:e64290. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 7.Benda N, Desai P, Reza Z, et al. Patient perspectives on AI for mental health care: Cross-sectional survey study. JMIR Ment Health. 2024;11:e58462. [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.
Data Availability Statement
Data, statistical code, and the informed consent document will be made available upon reasonable request after completing a data use agreement.