AI-written therapy feedback performed similarly in one trial; that does not establish equivalence

This is a rare useful AI trial because it asks not just whether a language model can sound therapeutic, but whether people improve when it supplies supervised feedback during a structured therapy program. It also manipulated what participants believed the source to be.

The participants improved in active treatment, and no statistically detected difference emerged between AI and human feedback. However, the investigators did not establish equivalence or noninferiority. The feedback was also psychologist-reviewed. These are important boundaries for any proposal to replace a qualified therapist with an unsupervised companion.

The thoughtful next experiment is an appropriately powered noninferiority design with harms, treatment completion, alliance, human-review burden and longer-term remission.

The evidence: 263 adults with social anxiety disorder randomized among psychologist-reviewed AI-generated feedback, human feedback, unguided self-help and waitlist, with experimental manipulation of stated feedback source. Active guided therapy reduced social anxiety relative to waitlist (reported d -1.23). AI versus human feedback did not differ statistically, but the trial explicitly did not establish noninferiority. Initial expectations favored a human source, whereas actual source/belief did not detectably change symptom improvement or therapeutic alliance.

unsupervised autonomous psychotherapy.