JMIR Publications

ABSTRACT

Background:

Generative artificial intelligence (AI) chatbots are increasingly used for health questions. A search engine returns a list of sources; a conversational system selects among them and returns one fluent, personalized answer. Detail, coherence, personalization, and agreement with what a user already believes can raise how credible an answer seems without raising how accurate it is. Surveys describe who uses these systems and how far they trust them, but not whether reliance is aligned with what users understand about how the answer was produced.

Objective:

This study aimed to examine whether understanding of how chatbot answers are produced is associated with reliance on credibility-enhancing conversational features and with self-reported adherence to health advice without independent checking, among adults living in Poland.

Methods:

This was a preregistered cross-sectional questionnaire study of adults aged 18 years or older living in Poland, recruited nonprobabilistically through an open link and poststratified to Statistics Poland margins for sex, age band, education, settlement size, and region. A mechanism-awareness index (range 0-3) was formed from 3 statements about how answers are produced, with "don't know" scored as not aware. One survey-weighted regression was fitted per registered hypothesis, each adjusted for 7 prespecified covariates including frequency of chatbot use. Confirmatory models were fitted in the subsample reporting health-related chatbot use.

Results:

Of 3360 submissions, 3348 were valid, of whom 2590 (77.4%) reported health-related chatbot use. The 2 primary awareness associations were approximately null before adjustment and became negative only after conditioning on chatbot-use frequency. Unadjusted estimates were close to null for credibility-cue reliance (β=−0.06, 95% CI −0.16 to 0.03) and for unverified adherence (odds ratio [OR] 0.98, 95% CI 0.81-1.18); after adjustment for the 7 registered covariates, both were negative (β=−0.30, 95% CI −0.38 to −0.22; OR 0.61, 95% CI 0.49-0.76). Replacing general chatbot-use frequency with health-specific use frequency produced a smaller shift (OR 0.81, 95% CI 0.68-0.96). In exploratory analyses, the conditional awareness-adherence association differed across levels of general chatbot use (interaction P<.001). Credibility-cue reliance was strongly associated with epistemic closure—stopping further search after a convincing answer—before and after adjustment (fully adjusted OR 3.75, 95% CI 2.99-4.70) and at every threshold of the outcome (threshold-specific ORs 3.06-9.34).

Conclusions:

Mechanism knowledge and reliance behavior are not simply aligned. Mechanism awareness showed essentially no marginal association with acting on chatbot advice unchecked, and the registered negative association emerged only once intensity of use was held constant, and not uniformly even then. Reliance on conversational credibility cues, by contrast, was strongly associated with ending further information search. Because the sample overrepresents health-AI users, prevalence estimates are not offered as population parameters. Mechanism education alone should not be treated as a sufficient safety strategy for generative AI in health.