What to ask when patients discuss trust in healthcare AI

A healthcare AI team cannot act on a statement about trust until it knows what decision the statement describes.

A weighted U.S. survey found that 25.4% of respondents were comfortable or very comfortable with provider AI use, 43.0% were uncomfortable, and 31.5% were neutral. The study separately measured prior use, trust in AI versus clinicians, care-plan explanation preferences, perceived fairness, and demographic variation (source).

Those measurements give product teams several lines of inquiry, but they do not explain why respondents answered as they did. They also do not show that a disclosure, interface change, or governance control would change an answer.

Begin with the situation

Before asking about trust, describe the proposed use. Name the AI-supported action, the setting, the consequence, and the human role.

A patient may respond differently to AI used for an administrative task, a care-plan explanation, or an output that informs a clinician. The supplied research does not compare those specific product scenarios, so a team should not transfer an aggregate survey response directly to one of them.

Useful prompts include:

• What AI-supported action is being considered?
• Is the concern about the provider using AI, the output, or the person or system explaining it?
• What decision would remain with a clinician or staff member?
• What consequence would follow if the output were wrong or unclear?

These prompts are an SGL interpretation of the research, not a validated measurement instrument.

Follow the answer to its operating expectation

An answer about comfort should lead to a question about what the person expects the organization to do.

Five virtual deliberations involving 159 Michigan residents elicited requests for clear communication about AI use, patient and community input, oversight, accountability, and AI that augments rather than replaces human judgment and empathy (source). Those deliberations suggest topics for a product discussion:

• What disclosure would make the AI’s role clear?
• Where should patients or community members provide input?
• Who should oversee the system and be accountable for its use?
• Where is human judgment expected?

A stated preference does not prove that a particular control will produce trust, adoption, satisfaction, or safety. It identifies an expectation that a team can investigate and document.

Preserve the wording and population

The two studies should not be combined into a national trust index. One was a weighted U.S. survey. The other comprised deliberations with Michigan residents. The supplied evidence does not establish that the populations are comparable or that either group’s responses predict behavior in a deployed system.

For each finding, retain the population, setting, question wording, response distribution, and uncertainty. That context matters when a team decides whether evidence is relevant to a product surface or procurement requirement.

Fairness needs the same treatment. The U.S. survey measured perceived fairness and demographic variation, but the supplied excerpt does not provide subgroup results (source). A team can record the fairness question and response distribution. It cannot use this evidence to claim which groups differed or why.

Leave discovery with a decision record

The useful output of the conversation is a record tied to the proposed use. It should state:

• the AI-supported action and setting;
• the respondent population and exact question;
• the expressed comfort or intended reliance;
• the preferred source of explanation;
• the relevant fairness concern;
• the expected disclosure, participation, oversight, and accountability;
• the boundary of human judgment; and
• the questions the evidence did not answer.

This record does not convert an attitude into a product requirement by itself. It shows which expectation is under discussion, what evidence supports it, and what the team must still test.

When discovery moves into the build process, connect the recorded expectation to a product surface, an owner, and an evaluation plan. The decision should rest on the stated use and retained evidence, not on an unqualified trust label.