Trust Turns Text into Action

For patient-facing health AI, a generated recommendation becomes clinically consequential through the person who receives it. The user may follow it, reject it, verify it elsewhere, delay action, or adopt only part of it. This chapter examines the trust and distrust that connect a specific piece of AI-generated health advice to those choices, and explains what the TAIGHA and TAIGHA-S scales can measure. The practical judgment is that a safe system should seek calibrated reliance, not maximum confidence in the machine.

Trust belongs to the advice instance

“Trustworthy” and “trusted” sound like two forms of the same idea, but they answer different questions. Trustworthiness describes a quality someone attributes to a system: perhaps it seems competent, reliable, transparent, or benevolent. Trust describes the user’s state of being willing to depend on a particular recommendation. A chatbot may appear imperfect and still be followed. Conversely, a system may have a good reputation while a user rejects one answer that feels implausible or unsafe.

That distinction is especially important in health. A user generally cannot inspect the training data, the model’s internal process, or the evidence behind a response. Their judgment forms during the interaction with the advice itself. The wording, explanation, apparent confidence, relevance to the symptom, and emotional tone can influence whether the user treats the recommendation as something to act on. If the advice is wrong or incomplete, misplaced trust can turn an information error into a behavioral risk. If the advice is sound but distrusted, the intended benefit may not reach the user.

The TAIGHA study—short for Trust in AI-Generated Health Advice—builds its measurement around this advice-level state. It treats trust as a willingness to follow health advice because the user expects it to help. That expectation has a cognitive side: the belief that following the recommendation will lead to a better health decision or outcome. It also has an affective side: feelings such as comfort, reassurance, or peace of mind when considering or relying on the recommendation.

Distrust is not merely a low score on the same scale. The study treats it as a protective stance: the user avoids following the advice because they expect it could cause harm. It too has cognitive and affective components. The user may judge that the recommendation is risky, while also feeling uneasy, tense, anxious, or distressed about acting on it. These reactions are related to trust but not reducible to its opposite. Someone can feel reassured by an answer and still recognize a reason to be cautious; someone else can distrust the source without having a clear factual objection.

The distinction changes the design goal. If a product team asks only how to increase trust, it may improve acceptance of both strong and weak advice. A more appropriate question is whether reliance tracks the quality, uncertainty, and urgency of the recommendation. A patient-facing system should not make a low-quality answer more persuasive simply because persuasion raises a trust score. Nor should it aim for blanket distrust that causes users to dismiss a useful emergency signal. Trust is valuable only when it is proportionate to what the system has established.