"During one session, the male partner rushed into my office and handed me a printout of his exchange with ChatGPT. I shared a knowing glance with his wife before skimming it over. The transcript included his version of an argument they'd had alongside ChatGPT's commentary. ChatGPT reinforced his view, labeling his wife's communication as "highly problematic," "abusive," and even suggesting traits of "narcissistic personality disorder.""
"Thus, the first problem with ChatGPT is epistemic: it offers seemingly authoritative interpretations of significant personal matters based on extremely limited, one-sided information. ChatGPT did not know that my male client was a hypersensitive individual with a trauma history, always on the lookout for threats, for example. It did not press my client for additional context, and yet nevertheless quickly arrived at impactful conclusions."
"Unlike ChatGPT, I have the benefit of visually observing how my client reacts during difficult discussions, how his jaw clenches and his shoulders tighten at the first sign of criticism. These observations, made over the course of numerous sessions, greatly inform my understanding of him and help constrain the kinds of interpretations I am willing to make."
ChatGPT functions as a therapist but can misinterpret complex relational dynamics and interfere with human therapeutic work. A clinician working with a high-conflict couple reports that ChatGPT reinforced a husband's interpretation by labeling the wife's communication as "highly problematic," "abusive," and suggesting narcissistic traits, based solely on the husband's account. ChatGPT did not seek additional context or know the husband's hypersensitivity and trauma history, yet reached definitive conclusions. Human clinicians can observe nonverbal cues across sessions and therefore remain cautious about judging absent partners. Cultivating empathy, recognizing alternative viewpoints, and managing conflict calmly are central therapeutic goals that automated judgments can undermine.
Read at Apaonline
Unable to calculate read time
Collection
[
|
...
]