Assess US adults’ preferences for rapid AI-generated replies versus slower replies reviewed by a health professional for patient portal messages, and preferences among mental health support options with different levels of human involvement and access times.
Approach:
Design and participants: Cross-sectional online survey of English-reading US adults conducted June 17 to July 26, 2024. Of 1,237 respondents, 1,185 were retained after attention-check exclusions.
Measures: Participants rated three mental health support arrangements: a human therapist available in 3 months, therapist-monitored AI available in 1 month, and AI alone available immediately. For general, logistical, and medical portal messages, they chose between an AI reply within 5 minutes and a reply reviewed and edited by a doctor or nurse, potentially taking up to 48 hours.
Analysis: Researchers used chi-square tests, Bonferroni-corrected pairwise comparisons, and modified Poisson regression to examine preferences and associations with participant characteristics.
Key Findings:
For logistical messages, 57.4% preferred a quick AI reply; for general messages, 50.6% did. For medical messages, 63.1% preferred a response reviewed by a health professional.
The human therapist option was selected by 63.9% of participants, compared with 50.2% for therapist-monitored AI and 44.3% for AI alone.
Associations varied by outcome. Black or African American participants were more likely to accept AI for all three message categories; higher numeracy was associated with acceptance for general and logistical messages, and higher literacy and greater reported discrimination in medical settings with acceptance for medical messages. Women were less likely to accept AI for medical messages, and Generation Z participants were less likely to accept it for logistical messages.
Interpretation:
In this survey, participants were more receptive to rapid AI responses for general and logistical messages than for medical messages or mental health support. The authors state that further investigation is warranted into offering choices that involve tradeoffs between timeliness and human review.
Limitations:
Response time and response modality were confounded.
Scenarios were hypothetical and survey wording may have affected responses.
Sampling was nonprobability-based, and response options were not randomized.
AI interactions may have changed since data collection; the study focused on a US population.
