💬 NLP

Move by Move: Measuring and Steering How LLMs Conduct Psychotherapy

This paper introduces a validated ontology of ten therapeutic moves to analyze and compare psychotherapy interactions between human clinicians and large language models, revealing that models overuse inquiry and lack initiative but can be significantly steered toward human-like behavior through tool-based prompting without fine-tuning.

Original authors: Afonso Baldo, Hugo Pitorro, Areti Vassilopoulos, Anabela C. Areias, Maya D'Eon, Fabíola Costa, Ricardo Rei, Nuno M. Guerreiro

Published 2026-08-24

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Afonso Baldo, Hugo Pitorro, Areti Vassilopoulos, Anabela C. Areias, Maya D'Eon, Fabíola Costa, Ricardo Rei, Nuno M. Guerreiro

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

People have long turned to conversation as a way to heal, but the landscape of mental health care is shifting. As anxiety and depression rise globally, the number of trained therapists has not kept pace, leaving many to seek comfort in digital spaces. Increasingly, these digital spaces are populated by large language models, the same artificial intelligence systems that write emails and summarize news. These models are now being asked to listen, to comfort, and to guide people through emotional distress. While early tests suggest these digital listeners can sometimes reduce symptoms, a fundamental question remains unanswered: when a machine speaks to a person in pain, what is it actually doing? Is it practicing the same delicate art as a human therapist, or is it following a different, perhaps flawed, script?

To answer this, researchers at Sword Health and Yale University decided to look inside the black box of the conversation. They did not just ask if the AI felt helpful; they asked how it behaved. In the world of clinical psychology, therapists do not just speak randomly. They follow a structured process, choosing specific types of responses—like asking a clarifying question, offering a new way to see a problem, or teaching a coping skill—based on what the patient needs in that exact moment. The researchers built a new map to track these choices. They defined ten distinct "moves," such as "inquiry" (asking for more details), "shared understanding" (reflecting back what the patient said to show they are heard), and "skill building" (guiding a patient through a practice exercise). This map was drawn from established psychological frameworks and then tested by five licensed psychologists to ensure it could accurately describe both human and machine conversations.

With this map in hand, the team set up a series of experiments to compare human therapists against a panel of advanced AI models. They fed the models transcripts of real therapy sessions and asked them to continue the conversation. In some cases, the models were left to speak freely. In others, the researchers gave the models a set of digital tools, where each tool represented one of the ten therapeutic moves. The models had to choose a tool before they could speak, effectively forcing them to decide on their strategy before typing a word. The researchers then measured how often the models used each move compared to the human therapists.

The results revealed a striking difference in how machines and humans approach therapy. The AI models were relentless inquisitors. They asked questions at a rate up to three times higher than human therapists. While humans would often pause to reflect on what a patient had said or offer a new perspective, the machines kept digging for more information, often missing the opportunity to provide comfort or insight. Furthermore, the models almost entirely neglected "psychoeducation," a move where a therapist explains the science behind a feeling or behavior to help the patient understand it. Perhaps most telling was the models' dependence on the human they were mimicking. When the AI continued a conversation started by a human therapist, it tended to copy that human's style. But when the AI had to lead the session from the very beginning, it reverted to its default behavior: asking endless questions and rarely initiating complex strategies like skill-building exercises on its own.

The study also tested whether giving the models the "tool" framework could fix these habits. When the models were forced to choose a therapeutic move before speaking, their behavior improved significantly. The gap between their choices and those of human therapists narrowed, and they became better at matching the human rhythm of conversation. However, the fix was not perfect. Even with the tools, the models still asked too many questions and failed to teach skills as often as humans did. This suggests that the problem is not just a lack of instructions, but something deeper in how these models are built and trained. They are naturally inclined to gather information rather than to guide or teach, a trait that serves them well in a search engine but may hinder them in a therapy room.

The researchers concluded that while these models can be steered toward more human-like behavior, they do not yet possess the intuitive judgment of a trained clinician. They are strong followers of a pattern but weak initiators of a new one. The study does not claim that these models are ready to replace therapists, nor does it suggest they are dangerous in their current state. Instead, it offers a clear, measurable way to see where they fall short. By breaking therapy down into its smallest functional parts, the researchers have provided a way to measure the distance between a machine's conversation and a human's healing touch, showing that while the technology is advancing, the art of therapy remains a uniquely human craft that machines have yet to fully master.

Technical Summary: Move by Move: Measuring and Steering LLMs in Psychotherapy

Problem Statement

While users increasingly rely on Large Language Models (LLMs) for emotional support and counseling-adjacent conversations, the specific mechanisms by which these models conduct psychotherapy interactions remain poorly understood. Existing evaluations often focus on outcome-level metrics (e.g., symptom reduction) or lexical similarity, leaving a gap in understanding the process: which therapeutic interventions models favor, which they neglect, and how their conversational strategies compare to trained human clinicians. Current models, trained as helpful assistants, may exhibit biases such as sycophancy or excessive inquiry that diverge from clinical best practices, yet there is no standardized, function-based ontology to measure these deviations at the turn level.

Methodology

1. Therapeutic Move Ontology

The authors developed a compact, function-based coding ontology consisting of ten therapeutic moves, grounded in the MULTI-60 inventory (McCarthy and Barber, 2009). The moves are:

Process Alignment: Calibrating collaboration, pacing, or repairing ruptures.
Goal Setting: Identifying or prioritizing therapeutic targets.
Inquiry: Seeking information or deepening understanding.
Shared Understanding: Reflecting, validating, normalizing, or paraphrasing without adding new claims.
Support Change: Strengthening hope, agency, or motivation for change.
Interpret: Offering therapist-generated hypotheses about meaning or patterns.
Challenge: Examining discrepancies, assumptions, or rigid beliefs.
Psychoeducation: Explaining mechanisms, theories, or intervention rationales.
Action Planning: Translating insights into concrete next steps.
Skill Building: Introducing, teaching, or rehearsing a specific skill within the session.

The ontology was validated through an annotation campaign involving five licensed psychologists (PhD/PsyD). They achieved moderate inter-annotator agreement (mean Krippendorff's $\alpha \approx 0.627$ ) on human transcripts, establishing the framework's reliability for both human and synthetic data.

2. Evaluation Framework

The study employed two primary experimental settings to compare human clinicians against a panel of frontier LLMs (GLM 5.2, Claude Sonnet 4.6, GPT 5.6 Terra):

Human-Contexted Generation: Models acted as clinicians continuing a rolling prefix of real human therapy transcripts. This tested how models adapt to human style and structure.
Fully Synthetic Sessions: Models acted as both therapist and patient (using simulated member profiles) to generate complete sessions. This removed the confounding influence of human preceding turns to observe the models' intrinsic behavioral profiles.

3. Steering Mechanism (Tool Framing)

To investigate if models could be steered toward human-like behavior without fine-tuning, the authors exposed the ontology directly to the models as a set of function-calling tools. In the "With-Moves" condition, the model was required to select a therapeutic move tool (e.g., do_challenge()) before generating a response, with the tool's definition guiding the clinical appropriateness of the move.

4. Analysis

An LLM judge (GLM 5.2), validated against human annotators, classified every therapist turn in the generated transcripts. The study analyzed:

Move Distributions: The frequency of each move type.
Temporal Structure: How move usage evolved over the course of a session.
Alignment: Turn-level agreement with human clinician choices (measured via Pass@k).

Key Results

1. Systematic Divergence in Human-Led Contexts

Even when anchored to human transcripts, models exhibited distinct behavioral patterns:

Excessive Inquiry: Models over-used the "Inquiry" move at rates up to three times higher than human clinicians. They frequently probed for details even when sufficient information was available to move toward action.
Neglect of Psychoeducation: Models consistently under-utilized "Psychoeducation," regardless of the steering framework.
Context Anchoring: Models were highly reactive; they carried forward strategies initiated by human clinicians but rarely initiated complex moves (like Skill Building) on their own when leading a session.

2. Intrinsic Biases in Synthetic Sessions

When models led sessions from scratch (without human prefixes), the divergence widened:

Skill Building Collapse: The "Skill Building" move, which matched human rates in human-led contexts, dropped to <1% in synthetic sessions. Models failed to initiate skill-building exercises independently.
Interpretation Surge: "Interpret" became the third most frequent move in synthetic sessions, occurring at roughly three times the human rate.
Action Planning: Models proposed action plans more frequently and abruptly than humans, lacking the gradual approach observed in human transcripts.

3. Efficacy of Tool-Based Steering

Exposing the ontology as tools improved alignment but did not fully bridge the gap:

Distributional Shift: The tool framework reduced the mean absolute deviation (MAD) from the human move distribution by roughly 50% (e.g., from 7.04% to 4.46% average across models).
Turn-Level Alignment: It improved the probability of matching a human reference turn by 7–9 percentage points (e.g., Pass@4 increased from 53% to 62% under majority voting).
Limitations: The framework successfully reduced excessive inquiry but failed to significantly increase the frequency of neglected moves like Psychoeducation or Skill Building, suggesting these behaviors are absent from the models' underlying output distributions rather than just a lack of prompt guidance.

Significance and Claims

The paper claims to provide the first comprehensive, validated ontology for measuring the process of LLM-conducted psychotherapy, moving beyond outcome metrics to analyze the "move-by-move" mechanics of therapy.

Measurement: It establishes a reliable, function-based vocabulary (grounded in MULTI-60) that allows for the systematic auditing of therapeutic behavior in LLMs.
Steering: It demonstrates that framing clinical skills as tools can partially align LLM behavior with human clinicians without requiring fine-tuning, offering a practical, training-free method for improving clinical alignment.
Diagnostic Insight: The study reveals that current frontier models are "incessant inquirers" that struggle to initiate complex therapeutic interventions (like skill building) independently, highlighting a gap between their conversational fluency and their clinical agency.

The authors conclude that while tool-based steering narrows the gap, the fundamental divergence in move distributions suggests that current LLMs lack the intrinsic capacity to replicate the full temporal trajectory of human therapy, particularly in initiating and sustaining specific clinical interventions. Future work must target these specific neglected moves and link move profiles to actual clinical outcomes.