experiments on Anne-Kathrin Kleine

Does it matter how your doctor uses AI? Patient trust in AI-supported medical decisions

Fri, 21 Aug 2026 00:00:00 +0000

More and more physicians make decisions together with AI-based clinical decision support. But how do patients feel about such hybrid decisions? In this project led by Insa Schaffernak, we examined whether the type of AI support and its timing in the decision process affect patients’ trust.

What we did

In two preregistered vignette experiments, 489 and 570 participants from Germany imagined four eye-care consultations, co-developed with ophthalmologists. The physician used no AI, descriptive AI (which highlights features in retinal images), or diagnostic AI (which additionally suggests a preliminary diagnosis). Study 2 also varied whether the physician looked at the AI output after an independent own assessment (sequential) or at the same time as the medical information (concurrent). We also analyzed more than 2,600 open-ended responses.

What we found

Diagnostic AI support led to lower trust in the medical decision and in the physician, and to a stronger feeling that the patient’s individual situation was neglected, compared with descriptive AI support. Participants’ willingness to follow the advice did not differ.
Study 2 replicated these results – and showed that timing matters: the disadvantage of diagnostic AI disappeared when the physician first assessed the case independently before consulting the AI.
In their open answers, participants worried that erroneous AI output could bias physicians and wished for AI as an independent second opinion.

Effects of the type (descriptive vs. diagnostic) and timing (concurrent vs. sequential) of AI support on trust in the decision, perceived trustworthiness, uniqueness neglect, and willingness to follow the advice. Figure 5 from Schaffernak et al. (2026), Journal of Medical Internet Research, licensed under CC BY 4.0.

What this means

Patients accept AI support more readily when it is used for analysis rather than for decisions – or when physicians visibly retain their independence. How AI is integrated into the consultation can therefore matter as much as whether it is used at all.

Read the paper

Schaffernak, I., Cecil, J., Kokje, E., Kleine, A.-K., Saad, A., Zemo, F., & Lermer, E. (2026). Effects of type and timing of clinician-facing AI support on patient trust in medical consultations: 2 vignette experiments. Journal of Medical Internet Research, 28, e93172. https://doi.org/10.2196/93172

Training future psychotherapists with AI-enabled virtual patients: A randomized study

Wed, 15 Jul 2026 00:00:00 +0000

Preparing psychotherapy trainees for the complexity of real clinical encounters is a persistent challenge. AI-enabled virtual patients – simulated patients with whom trainees can conduct conversations and receive automated feedback – promise hands-on practice without risks for real patients. In this preregistered study led by Julia Cecil, we tested whether such simulations improve training outcomes.

What we did

87 psychology students and psychotherapy trainees from Germany and the UK were randomly assigned to either four online sessions with virtual patients (with automated feedback) or four sessions with pre-recorded role-play videos. Each session focused on one disorder – depression, panic disorder, generalized anxiety disorder, or adjustment disorder. Before and after the training, we measured perceived psychotherapeutic competence, clinical self-efficacy, knowledge, and insecurity.

What we found

Virtual patient training improved perceived competence and self-efficacy and reduced insecurity – but these effects did not exceed those of the video training.
Neither training increased knowledge about mental disorders.
More clinical experience was associated with higher competence and lower insecurity, but did not change how much participants gained from the training.
Insecurity decreased earlier with the video training than with the virtual patients.

Mean (a) competence, (b) self-efficacy, (c) knowledge, and (d) insecurity before and after the training. Figure 3 from Cecil et al. (2026), BMC Medical Education, licensed under CC BY 4.0.

What this means

AI-enabled virtual patients are a scalable and engaging complement to psychotherapy training, not (yet) a superior alternative. Good pre-briefings and more authentic virtual patients could increase their benefits.

Read the paper

Cecil, J., Kokje, E., Kleine, A.-K., Schaffernak, I., Vogel, L., Angerer, S., Lermer, E., & Gaube, S. (2026). The effect of AI-enabled virtual patient simulation on training outcomes and insecurities in psychotherapy education. BMC Medical Education, 26, 1156. https://doi.org/10.1186/s12909-026-09917-x

Can the timing of AI advice reduce overreliance? Evidence from face matching

Thu, 05 Feb 2026 00:00:00 +0000

Human–AI teams should ideally perform better than either humans or AI alone. A key obstacle is overreliance: people follow AI advice even when it is wrong. In this project led by Eesha Kokje, we tested whether changing when AI advice is presented can help – using face matching, a task in which people decide whether two photos show the same person.

What we did

In three preregistered online experiments (N = 80, 77, and 86), participants judged pairs of unfamiliar faces with the help of an AI system that was correct in 92.5% of cases. We compared the usual concurrent presentation (advice shown together with the faces) with three non-concurrent alternatives: binary advice on demand, similarity ratings on demand, and conditional advice that was only shown when a participant’s initial decision contradicted the AI.

What we found

Overall accuracy did not differ between concurrent and non-concurrent advice in any experiment.
When participants requested binary advice, they followed it more – although they requested it on only about a quarter of the trials.
When they requested similarity ratings, they followed the advice less: overreliance decreased, but so did reliance on correct advice. Similarity ratings were not more helpful than simple advice.
With conditional advice, participants followed the AI less often. They were more confident when rejecting incorrect advice, but less confident when accepting correct advice.

(a) Agreement with AI advice by advice accuracy and (b) types of reliance (AA = appropriate acceptance, AR = appropriate rejection, OR = overreliance, UR = underreliance) for concurrent versus conditional advice. Figure 6 from Kokje et al. (2026), Cognitive Research: Principles and Implications, licensed under CC BY 4.0.

What this means

Non-concurrent advice can reduce overreliance on AI – but it is no free lunch, because people then also reject more correct advice. The goal should be calibrated reliance, which may require combining presentation formats with information about when the AI is likely to be right.

Read the paper

Kokje, E., Lermer, E., Kleine, A.-K., & Gaube, S. (2026). AI-augmented decision-making in face matching: Comparing concurrent and non-concurrent advice presentation. Cognitive Research: Principles and Implications, 11, 11. https://doi.org/10.1186/s41235-026-00707-z