Move by Move: Dissecting How LLMs Conduct Psychotherapy

> *"We should not only ask what AI can do, but what it is actually doing."*

Opening: A Late-Night Confession

Imagine it is 2 a.m. You cannot sleep. Instead of calling a friend, you open a chat window and talk to an entity that never tires and never judges. "I've been feeling terrible lately..." It replies almost instantly: "That sounds like a difficult time. Can you tell me more?" It absorbs your emotions like a perfect sponge—yet rarely gives anything substantive back. This scenario is becoming a daily reality for millions. But we rarely ask: is this AI actually "treating" you, or just "collecting" you?

The paper reviewed here, by an interdisciplinary team of psychologists and NLP researchers, builds a precise scalpel to dissect the process of AI-delivered psychotherapy.

Chapter 1: Psychotherapy Is Not Small Talk

Real psychotherapy—whether CBT, psychoanalysis, or person-centered therapy—is a highly structured technical activity. A trained therapist makes hundreds of deliberate therapeutic moves in a 50-minute session.

The paper's core contribution is an Ontology of Therapeutic Moves, distilled from the MULTI-60 inventory (60 therapeutic behavior codes used in psychotherapy research) into 10 compact, functional categories:

| Move | Plain meaning | Function | |------|---------------|----------| | Inquiry | "Can you say more?" | Gather information, guide self-exploration | | Reflection | "It sounds like you're frustrated" | Confirm emotions, build empathy | | Interpretation | "Is that anger perhaps rooted in fear?" | Offer new perspectives, promote insight | | Psychoeducation | "Anxiety is a survival mechanism..." | Teach psychological knowledge, empower | | Validation | "Your feelings are entirely reasonable" | Normalize the client's experience | | Affirmation | "You're doing well" | Reinforce positive behavior | | Confrontation | "That contradicts what you said earlier" | Gently point out inconsistencies | | Self-disclosure | "I'd feel confused too in that situation" | Build the therapeutic alliance | | Immediacy | "I notice you're withdrawing right now" | Attend to the here-and-now interaction | | Therapeutic Silence | (deliberately saying nothing) | Give space for emotions to settle |

Chapter 2: Validation by Five Psychologists

The authors did not stop at building the taxonomy. Five licensed psychologists independently annotated real counseling transcripts, and inter-annotator agreement was measured. The paper's automated judge-based annotation approach achieved agreement comparable to the human experts. The taxonomy is both human-interpretable and machine-scalable—making it possible to quantify, for example:

• Human therapist A: Inquiry 25%, Reflection 30%, Psychoeducation 15%...
• AI model B: Inquiry 75%, Reflection 10%, Psychoeducation 2%...

Chapter 3: What the Data Reveals About AI Therapists

The "over-inquiry" problem

AI models use Inquiry about 3× more often than human therapists. The result feels less like therapy and more like a gentle interrogation. The authors describe this as context-anchored behavior: AI inherits strategies initiated by human therapists, but when leading a conversation itself, it defaults to questioning—because asking is the safest strategy and requires no substantive content.

A psychoeducation desert

AI almost never provides psychoeducation. Knowledge like "anxiety is the result of an overactive amygdala" transforms "something is wrong with me" into "my brain is working in a particular way," converting shame into understanding. AI rarely offers it—possibly due to training data, its tuning as a "listener," or conservative defaults. The result: AI therapy becomes an emotional dumping ground with few moments of insight.

The "tool exposure" experiment

The paper's most striking intervention: explicitly exposing the 10 therapeutic moves to the model as tools, telling it to choose the right one for the situation. Results:

• The AI's move distribution deviation from humans dropped by half
• Turn-level alignment with human therapists improved by 7–9 percentage points
• No fine-tuning required—prompting alone

This shows AI is not "unable" to behave like a therapist; it simply "doesn't know" what it is doing—like a talented guitarist handed a chord chart.

Chapter 4: Deeper Implications

Companionship vs. therapy

If millions confide in AI at night, the feeling of being heard has value. But if they mistake it for treatment, they may miss moments requiring real professional help. The data suggests AI therapists may reinforce a kind of "emotional consumerism": instant gratification of being heard, while avoiding the discomfort of being challenged, educated, and asked to change—elements that are necessary conditions for change, not good UX.

The limits of technical optimism

Three caveats:

1. Behavioral alignment ≠ therapeutic outcome. The paper measures "behaving more like human therapists," not client improvement. 2. Human therapist behavior is not necessarily optimal—the baseline itself carries biases and cultural limits. 3. Psychotherapy's core is the relationship, not the moves. The therapeutic alliance is the strongest predictor of outcomes; whether AI can build genuine relationships remains open.

Mirror or window?

A good therapist is a window—through it you see patterns you never noticed. Most current AI therapists are a mirror—faithfully reflecting what you already know. The data suggests AI mostly does mirror work (inquiry, reflection) and rarely window work (interpretation, psychoeducation, confrontation).

Chapter 5: From Paper to Practice

For developers: 1. Don't optimize only for user satisfaction—a client who merely "feels heard" may be delayed from real professional intervention. 2. Actively introduce "uncomfortable" elements: prompt for interpretation, psychoeducation, gentle confrontation at appropriate moments. 3. Monitor inquiry rate—an AI asking questions in 80% of turns is a red flag. 4. Consider hybrid mode: AI handles early support and information gathering; critical moments (suicide risk assessment, deep intervention) must transition seamlessly to human professionals.

For users: 1. Distinguish venting from therapy—persistent distress warrants licensed professionals. 2. Watch for the questioning trap; ask the AI for feedback or analysis when it only asks questions. 3. Protect your privacy—conversations may be used for training.

Conclusion

The paper's real value is that it asks the right questions: What is the nature of AI-provided support? How does it differ from human professionals? What do those differences mean for users? How can we design and deploy these systems responsibly? Knowing the *names* of therapeutic moves is not the same as truly understanding them—but we are no longer groping in the dark.

Reference

Baldo, A., Pitorro, H., Vassilopoulos, A., Areias, A. C., D'Eon, M., Costa, F., Rei, R., & Guerreiro, N. M. (2026). Move by Move: Measuring and Steering How LLMs Conduct Psychotherapy. arXiv:2608.21325v1.