Primary Care Patients Report Cautious Perceptions Toward Portal Artificial Intelligence Chatbots

Patients using an academic primary care portal evaluated a hypothetical artificial intelligence chatbot with cautious ratings regarding safety and accuracy before implementation, according to findings from a cross-sectional survey. Health systems have increasingly considered integrating conversational tools directly into patient portals to handle patient communications. However, little was previously known about how active portal users perceive these artificial intelligence systems when connected directly to their medical records, particularly in light of documented concerns regarding large language models answering clinical questions.

To evaluate patient attitudes toward portal-embedded artificial intelligence prior to any active rollout and identify associated patient characteristics, researchers surveyed active primary care patient portal users across an academic health system. The investigation sought to establish baseline acceptance levels and determine how demographic characteristics, previous digital experience, and attitudes toward technology influence willingness to interact with conversational artificial intelligence in a healthcare portal environment.

Study participants received a formal description of a hypothetical portal-embedded artificial intelligence chatbot based on Epic’s Emmie system. After reviewing the operational description, participants completed evaluations rating four core domains: general acceptability, trust in receiving accurate health information, perceived patient safety, and personal willingness to talk to a chatbot. Each evaluated domain was measured using standardized 5-point Likert scales to capture the full spectrum of patient response.

To assess independent relationships between patient traits and evaluation scores, the authors employed multivariable linear regression and ordinal logistic regression models. These statistical models systematically adjusted for sociodemographic factors, including age, gender, race, and retirement status. The analyses also examined technological behavior and attitudes by incorporating portal task counts, overall portal use frequency, and measured levels of technophobia.

The study collected responses from 2,955 portal users, achieving a response rate of 16.1 percent, with 2,678 completed responses qualifying for the final statistical analysis. Across the 5-point evaluation scale, the overall mean score for general acceptability reached 3.30. In contrast, measures examining trust and clinical safety scored lower, with mean ratings of 2.59 for trust in accurate health information, 2.51 for perceived patient safety, and 2.79 for personal willingness to talk to the tool.

Multivariable regression revealed distinct sociodemographic associations across all four evaluated domains. Older age was associated with higher acceptability, registering a positive coefficient of 0.013 per year of age, while Nonwhite race was also associated with higher acceptability with a coefficient of 0.218. Conversely, female gender was associated with lower acceptability ratings, registering a negative coefficient of -0.115. These demographic patterns remained consistent across trust in accuracy, patient safety, and willingness to converse, with all observed associations reaching statistical significance at p values of 0.044 or lower.

Patient experience with the digital portal and general comfort with technology also demonstrated significant associations with chatbot perceptions. A greater portal task count was positively associated with higher acceptability, displaying a coefficient of 0.055 with a p value of 0.001, as well as an increased willingness to talk to the chatbot, with an odds ratio of 1.100 and a p value of 0.006. In contrast, technophobia showed a pronounced negative relationship with acceptability; participants with mild technophobia reported lower acceptability scores with a coefficient of -0.212, while moderate or high technophobia yielded a coefficient of -0.429, both with p values below 0.001. Additionally, retired status was associated with lower acceptability, registering a coefficient of -0.119 with a p value of 0.029, alongside lower perceived safety, with an odds ratio of 0.742 and a p value of 0.006.

The study authors concluded that patient perceptions varied substantially across specific outcomes and individual patient characteristics, showing higher openness among certain demographics but widespread caution regarding clinical safety and factual accuracy. The researchers noted that these findings support a cautious approach to deploying portal-embedded artificial intelligence chatbots in healthcare systems, emphasizing that implementation strategies must consider patient perceptions alongside the known operational limitations of large language models.

Subject of Research: Patient perceptions and acceptability of artificial intelligence chatbots in primary care patient portals

Article Title: Patient Acceptability and Perceptions of an Artificial Intelligence Chatbot in a Primary Care Patient Portal: Cross-Sectional Survey

Article References: Haessner, P., Romero, C. L., Ike-Okpe, O., Hammer, N. C., & Gregory, M. E. (2026). Patient Acceptability and Perceptions of an Artificial Intelligence Chatbot in a Primary Care Patient Portal: Cross-Sectional Survey. Journal of General Internal Medicine. https://doi.org/10.1007/s11606-026-10965-8

Image Credits: AI Generated

DOI: 10.1007/s11606-026-10965-8

Keywords: artificial intelligence, patient portals, primary care, chatbots, large language models, patient acceptability, health information accuracy, perceived patient safety, academic health systems, Epic Emmie, health informatics, digital health, technophobia, cross-sectional survey, survey research, patient trust