Scale MI with AI: ai motivational interviewing

Fine‑tuned LLMs can reproduce many core Motivational Interviewing behaviors and reach MITI adherence comparable to real MI dialogs for some models, but they lag on complex reflections and reflection: question balance.

Use MI‑LLMs for scalable augmentation, reproducible fidelity checks, and pilot support, not as a replacement for trained counselors.

• In the JMIR study (24 Jun 2026), researchers transformed 2, 040 Chinese counseling dialogs and fine‑tuned three open-source LLMs on 2, 000 dialogs.
• Fine‑tuning improved automatic metrics and MITI adherence: ChatGLM total MI‑adherent ratio 0.78, Baichuan 0.75, Llama‑3 0.72, compared with real MI dialogs at 0.73.
• The primary gap was complex reflections: real dialogs 0.37 vs MI‑LLMs ~0.24–0.31.

This study, published 24 Jun 2026, transformed 2, 040 counseling dialogs into MI-style conversations and fine-tuned three Chinese-capable LLMs.

• Source: Runze Hu et al., JMIR Formative Research, published 24 Jun 2026.
• Dataset: 2, 040 MI-style dialogs created by transforming public Chinese counseling corpora (CPsyCounD + PsyDTCorpus); 2, 000 used for training, 40 for testing.
• Models fine-tuned: Baichuan2-7B-Chat, ChatGLM-4-9B-Chat, Llama-3-8B-Chinese-Chat-v2 using LoRA.
• Evaluation: automatic (BLEU‑4, ROUGE) and manual MITI 4.2.1 coding (global scores, behavior counts, summary ratios).

This section explains the study methods in plain English: the team screened public Chinese counseling corpora, transformed dialogs into MI-style conversations, fine-tuned LLMs, and evaluated outputs automatically and manually.

They screened five public Chinese counseling corpora for quality (comprehensiveness, professionalism, authenticity, safety), chose the top two, and used a structured GPT‑4 prompt to convert 2, 040 dialogs into MI‑style multiturn conversations between Jan 1 and Feb 28, 2025.

Three open-source Chinese-capable LLMs were fine‑tuned using low‑rank adaptation (LoRA) on 2, 000 dialogs (3 epochs, LR 1e‑4) and evaluated automatically (BLEU/ROUGE) and manually using MITI 4.2.1 coding of 60 deidentified transcripts (30 simulated MI‑LLM dialogs vs 30 translated real MI dialogs).

Evaluators were trained graduate students; manual coding included global dimensions (cultivating change talk, softening sustain talk, partnership, empathy) and behavior counts (reflections, questions, affirmations).

This two‑week plan reproduces the study’s validation steps for a compact pilot.

• Day 1–2: Collect and deidentify 200–500 counseling transcripts or chat logs, and map speaker roles and target behaviors.
• Day 3–5: Use Evidano to transcribe/translate (if needed), apply a prompt‑based MI transformation on a sample, and import the transformed dialogs.
• Day 6–8: Fine‑tune an internal model or generate candidate MI replies via an LLM; export model responses back into Evidano as a new corpus.
• Day 9–11: Run automated similarity metrics (BLEU/ROUGE) and MITI‑style behavior counts; visually compare global scores and reflection: question ratios by segment.
• Day 12–14: Blind manual coding on a randomized subset to validate automated counts; prepare a stakeholder brief with co‑occurrence networks and quote anchors.

This JMIR study (24 Jun 2026) demonstrates a practical, reproducible way to build MI training corpora and fine‑tune LLMs that approximate many MI behaviors, while human counselors still lead on deep reflections and relational nuance.

If you are running pilots or need reproducible fidelity checks, Evidano can ingest your transcripts, run MITI‑style automated counts, compare model versus human outputs, and produce shareable visuals for stakeholders.

Ready to validate MI‑LLMs on your corpus? Start a pilot and see immediate gains in coding speed and cross‑segment insight: Try Evidano for free.