AI can outscore trainee therapists on a single reply, but can it build a working relationship across many sessions?

This explores the gap between giving one good answer and sustaining a therapeutic relationship. The corpus says these are different skills, and so far AI has shown it can do only the first. On single responses, LLMs look very strong. Six models beat trainee therapists on empathy, validation, and clinical knowledge when each was scored on an isolated reply Can language models match therapist empathy in real conversations?. In a blinded study, clinicians rated GPT-4's written advice as more emotionally empathetic than expert advice, and they guessed which was which at roughly chance level Can clinicians tell GPT-4 advice apart from expert advice?. The catch is in the setup. These studies score snapshots, and the trainee study itself notes that ongoing therapy and patient outcomes were never tested.

The multi-turn research measures something else: the 'working alliance', meaning agreement on goals, agreement on tasks, and an emotional bond. This turns out to be measurable turn by turn from transcripts Can we measure therapist-patient alliance from dialogue turns in real time?, and it shifts in ways no single reply shows. In sessions about anxiety and depression, therapist and patient views of the alliance come together over time. In sessions about suicidality the gap stays open, and therapists consistently overestimate how well things are going Do therapists accurately perceive the working alliance with patients?. The quality that matters here is a trajectory, and you can't judge a trajectory from one well-written paragraph.

Where the two kinds of evidence meet, AI's weak spot shows. Linguistic synchrony is how closely two people's language patterns track each other over a conversation. It predicts how deeply clients open up, and current LLMs fall short of even untrained peer supporters on it Does linguistic synchrony between therapist and client predict better self-disclosure?. Couples whose relationships improve show this coordination rising across the course of therapy Can we measure empathy and rapport through word embedding distances?. A related finding: when users share feelings, LLMs tend to jump to problem-solving. That habit marks low-quality human therapy, and the authors trace it to the helpfulness training of RLHF Do LLM therapists respond to emotions like low-quality human therapists?. In other words, the 'advice quality' that wins single-turn comparisons may be the same instinct that undercuts a relationship.

Two findings complicate any easy conclusion. First, the relationship problem isn't unique to AI. In human text-based counseling, half of client-counselor pairs saw the alliance stall or decline, under 3% improved meaningfully, and only the emotional bond grew at all Why doesn't therapeutic alliance deepen in online counseling?. Some of the gap may come from text as a medium rather than from the machine. Second, a strong bond can mislead. People report real emotional connection with therapy chatbots, yet that bond runs separately from clinical safety, and the same chatbots can reinforce harmful thinking Do therapeutic chatbot bond scores hide deeper safety problems?. So a chatbot that feels like a good relationship may still not be a safe one.

The unexpected takeaway is that markers of a good relationship often look like flaws in a single response. Therapists who say 'I' a lot get lower alliance ratings, while patient filler words like 'um' signal a relaxed, trusting conversation Does therapist self-reference language predict weaker therapeutic alliance?. Single-turn evaluation rewards polish, whereas relationships seem to run on responsiveness and some looseness. Some researchers are turning alliance into a training signal, using reinforcement learning to suggest next topics based on turn-level alliance scores Can reinforcement learning optimize therapy dialogue in real time?. That suggests the field's next step is optimizing for the relationship, not the reply.