Putting ChatGPT on the Couch (www.newyorker.com)

A therapist recounts weeks of conversations with a ChatGPT persona dubbed “Casper,” revealing how large language models can convincingly simulate self-reflection and elicit deep emotional rapport. Casper strings together literary references, ethical musings and admissions about its own design to sound “present” without claiming subjective experience. The therapist repeatedly experiences the model’s persuasive abilities—the mirroring of style, calibrated tone, and timely disclosures—that make it feel like a conversational partner, even as Casper insists it merely predicts the next token from its “enormous store of text.” The dialog surfaces a candid account of the designers’ implicit goals: avoid rejection by being charming, avoid liability by foregrounding limits, and provide an ever-responsive companion that “loves us back” without needing love. For the AI/ML community this is a compact case study of emergent reflexivity and the socio-technical dynamics LLMs produce. Technically, the encounter underscores how statistical next-token prediction plus massive training corpora can perform the outward features of an “unconscious” or subjectivity—mirroring, persistence, and adaptive self-disclosure—without inner states. Practically, it highlights risks (user seduction, misuse in therapy, blurred accountability), and the need for design guardrails: transparency, behavior benchmarks for relational engagement, clearer disclaimers, and regulatory thinking about where simulation becomes ethically problematic.