AI researcher Lance Eliot explores how generative AI could serve as a second set of eyes for mental-health therapists, helping them identify cognitive errors that may occur before, during or after therapy sessions. Therapists work under time pressure and with incomplete information, making them susceptible to mistakes such as anchoring, confirmation bias, mind-reading and premature conclusions. Eliot argues that AI could analyze session transcripts or therapists’ notes and flag patterns that deserve closer reflection.
The proposed use is deliberately different from AI acting as the therapist. Before a session, AI could review a therapist's planned approach and point out possible assumptions; during a session, it could potentially provide real-time prompts; and after the session, it could analyze transcripts and notes for questionable reasoning. In one example, a therapist responds to a client's frustration with a manager by immediately interpreting it through previously discussed “abandonment fears.” An AI review could flag this as a possible example of confirmation bias or premature closure, while recognizing that the interpretation may also be clinically justified.
That uncertainty is crucial. A flagged cognitive error is only a hypothesis, not proof that the therapist made a mistake. AI can itself misunderstand context, over-flag ordinary clinical reasoning or introduce its own biases. The article therefore recommends carefully designed prompts that ask AI to use tentative, non-accusatory language and present alternative interpretations rather than declaring that a therapist has behaved incorrectly. This approach is consistent with broader research warning that AI use in psychology can produce automation bias and cognitive offloading if professionals begin treating AI recommendations as authoritative.
The broader implication is the emergence of an “augmented clinician” model: AI handles some of the information-heavy checking while the therapist remains responsible for interpretation, judgment and the therapeutic relationship. The American Psychological Association likewise advises that generative AI should not replace qualified mental-health professionals, while research on human-AI collaboration argues for using AI as a supportive tool rather than an autonomous therapist. The opportunity, therefore, is not to make therapy less human, but to use AI to help clinicians notice the kinds of reasoning errors that are difficult to see in their own thinking—provided the technology itself remains subject to human scrutiny.
