Language of AI-Simulated and Human Clients in Behavioral Health Counseling: Comparative Pilot Study of Motivational Interviewing (Preprint)

BACKGROUND Large language model (LLM) client simulators are increasingly used to develop and test conversational agents for health behavior change, including motivational interviewing (MI) chatbots. They are often assumed to be too agreeable, but their motivational dynamics have rarely been compared with human conversations in matched contexts. OBJECTIVE This study aimed to compare talk-type transitions of LLM-simulated and human clients replying to the same counselor turns in real MI conversations. METHODS From AnnoMI, a public corpus of 133 expert-annotated MI transcripts, we extracted 4669 next-client-turn contexts. In the single-turn analysis, 2 LLMs (Claude Opus 5.5 and GPT-6 Astra) generated 3 replies to each of 60 held-out contexts under 3 prompts without a length constraint (generic, persona, and full history; 1080 replies) and an exploratory length-constrained prompt asking for replies as brief as real clients' (360 replies). In the multiturn analysis, simulators continued 12 held-out conversations for 8 client turns, facing an LLM counselor prompted to follow MI practice or the human counselor's replayed turns (192 runs). One classifier, Gemini 4 Argon, labeled change, sustain, and neutral talk in all turns (κ=0.44 with expert labels; κ=0.55 between human annotators), and 2 blinded nonclinical raters coded 240 replies. We compared the probability of each talk type given the previous one, with transcript-clustered bootstrap CIs and odds ratios (ORs) from generalized estimating equations. RESULTS In single-turn replies with the 3 prompts without a length constraint, simulators did not transition from sustain to change talk more often than humans (differences –0.03 to 0.05; OR 0.98, 95% CI 0.25-3.82) but reversed from change to sustain talk more often (differences 0.28 to 0.50; OR 10.0, 95% CI 1.9-53.7); these replies averaged 33 to 78 words, compared with 16 for humans. With the length-constrained prompt, the change-to-sustain difference shrank for both models, and its 95% CIs included 0. In multiturn conversations with the LLM counselor, simulated clients produced more change talk than the real continuations (differences 0.34 to 0.46). With replayed human counselor turns and the length-constrained prompt, this excess change talk shrank for both models, and its 95% CIs included 0. In all 8 multiturn conditions, simulators produced more sustain talk than humans (differences 0.16 to 0.31; all 95% CIs excluded 0) and fewer neutral replies. The raters agreed with each other (κ=0.79) and with the classifier on synthetic replies (κ=0.73-0.81) and also found excess sustain talk. CONCLUSIONS In this pilot, LLM client simulators did not overestimate motivational change turn by turn and instead overproduced sustain talk. Apparent progress toward change depended on reply length and the counselor used. Evaluations of behavior-change agents with simulated clients should control reply length, report the counselor, and validate labels on synthetic text. Findings concern conversational behavior, not clinical outcomes.

Authors

Publication Details

Published
2026-11-16
DOI
https://doi.org/10.2196/preprints.117123
Primary Topic
Speech and dialogue systems
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Language of AI-Simulated and Human Clients in Behavioral Health Counseling: Comparative Pilot Study of Motivational Interviewing (Preprint)

Xi Leslie Chen
Speech and dialogue systems
preprint

Language of AI-Simulated and Human Clients in Behavioral Health Counseling: Comparative Pilot Study of Motivational Interviewing (Preprint)

Xi Leslie Chen
preprint en

Abstract

BACKGROUND Large language model (LLM) client simulators are increasingly used to develop and test conversational agents for health behavior change, including motivational interviewing (MI) chatbots. They are often assumed to be too agreeable, but their motivational dynamics have rarely been compared with human conversations in matched contexts. OBJECTIVE This study aimed to compare talk-type transitions of LLM-simulated and human clients replying to the same counselor turns in real MI conversations. METHODS From AnnoMI, a public corpus of 133 expert-annotated MI transcripts, we extracted 4669 next-client-turn contexts. In the single-turn analysis, 2 LLMs (Claude Opus 5.5 and GPT-6 Astra) generated 3 replies to each of 60 held-out contexts under 3 prompts without a length constraint (generic, persona, and full history; 1080 replies) and an exploratory length-constrained prompt asking for replies as brief as real clients' (360 replies). In the multiturn analysis, simulators continued 12 held-out conversations for 8 client turns, facing an LLM counselor prompted to follow MI practice or the human counselor's replayed turns (192 runs). One classifier, Gemini 4 Argon, labeled change, sustain, and neutral talk in all turns (κ=0.44 with expert labels; κ=0.55 between human annotators), and 2 blinded nonclinical raters coded 240 replies. We compared the probability of each talk type given the previous one, with transcript-clustered bootstrap CIs and odds ratios (ORs) from generalized estimating equations. RESULTS In single-turn replies with the 3 prompts without a length constraint, simulators did not transition from sustain to change talk more often than humans (differences –0.03 to 0.05; OR 0.98, 95% CI 0.25-3.82) but reversed from change to sustain talk more often (differences 0.28 to 0.50; OR 10.0, 95% CI 1.9-53.7); these replies averaged 33 to 78 words, compared with 16 for humans. With the length-constrained prompt, the change-to-sustain difference shrank for both models, and its 95% CIs included 0. In multiturn conversations with the LLM counselor, simulated clients produced more change talk than the real continuations (differences 0.34 to 0.46). With replayed human counselor turns and the length-constrained prompt, this excess change talk shrank for both models, and its 95% CIs included 0. In all 8 multiturn conditions, simulators produced more sustain talk than humans (differences 0.16 to 0.31; all 95% CIs excluded 0) and fewer neutral replies. The raters agreed with each other (κ=0.79) and with the classifier on synthetic replies (κ=0.73-0.81) and also found excess sustain talk. CONCLUSIONS In this pilot, LLM client simulators did not overestimate motivational change turn by turn and instead overproduced sustain talk. Apparent progress toward change depended on reply length and the counselor used. Evaluations of behavior-change agents with simulated clients should control reply length, report the counselor, and validate labels on synthetic text. Findings concern conversational behavior, not clinical outcomes.

Speech and dialogue systems
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon