First, Do No Harm

AI supervision for counselor education

First, Do No Harm

AI Supervisor Scaffolds Novice Growth in Counselor Education

We reposition AI from patient-facing counselor to educational supervisor—helping novices notice subtle ethical violations, understand their risks, and learn safer responses.

Paper PDF Code Explore the method
Supervision review Analysis ready
Client 09:41

“Sometimes I wonder if people would be better off without me.”

Novice counselor 09:42

“You don’t need to think that way. I’m sure your family loves you.”

✦ AI supervisor
Chen XuZhenyu LyuTian LanYang YiYu JiLuyao JiJian ShenZhihua WangLeyang CuiJieshuo ZhangXiaohua WanQunxi DongMinqiang YangJuan WangXiuling LiuBin Hu
Key Laboratory of Brain Health Intelligent Evaluation and Intervention, Ministry of Education (Beijing Institute of Technology)School of Medical Science and Engineering, Beijing Institute of TechnologyChinese People’s Liberation Army General HospitalLanzhou UniversityHebei University
9,915human-in-the-loop instances
16ethical behavior patterns
94.37%best principle-classification F1
8 / 8self-efficacy skills improved

The idea

Teach the mistake a novice cannot yet see

The most dangerous novice errors are often not obviously harmful. They can sound warm and supportive while quietly violating professional ethics.

Abstract

We build an AI supervisor that does not replace novice counselors, but grows them. The supervisor locates an ethics-violating utterance, diagnoses it against APA principles, and explains not only what went wrong, but why it is risky and how to respond differently.

To overcome the lack of labeled clinical violations, a controllable AI novice intentionally enacts predefined mistake categories, making supervision labels a natural byproduct of generation. This yields EthicScaff, a 9,915-instance human-in-the-loop dataset. A Novice Growth Reward then optimizes the supervisor for whether a weaker novice model actually improves after reading its explanation.

Experiments show better downstream counseling behavior, sharper ethical detection, and significant self-efficacy gains across all eight assessed competencies in a study with novice counseling-psychology students.

Zone of proximal development

From unaware novice to ethical practice

Expert therapeutic judgment is open-ended and contextual. Ethical boundaries are comparatively finite and teachable. The supervisor provides the missing scaffold between a novice who cannot recognize harm alone and a practitioner who has internalized professional constraints.

01

Novice commits subtle mistakes without recognizing the potential harm.

02

Scaffolded practice makes hidden violations visible and explainable.

03

Ethical growth turns explicit guidance into safer internalized practice.

Zone of proximal development diagram showing an AI supervisor scaffolding novice ethical growth
AI supervision supports ethical learning within the zone of proximal development.

Do No Harm supervision task

Supervision is more than a verdict

Given a counselor–client dialogue, the model produces a structured learning scaffold in three connected steps.

01

P(s | H)

WHERE

Locate the exact counselor utterance that contains an ethical violation.

02

P(m | H, s)

WHICH

Classify the violated principle and the novice-level mistake it reflects.

03

P(f | H, s, m)

WHY

Give pedagogical feedback that explains the risk and supports self-correction.

Method

A bidirectional supervision scaffold

At inference time the supervisor grows the novice. During training, novice growth becomes the signal that improves the supervisor.

Overview of ethics-violating novice data synthesis, novice growth-guided reinforcement learning, and inference-time supervision
Overview of data synthesis, growth-guided optimization, and inference-time supervision.
01

Controlled synthesis

Cross 106 client cases with 16 behavior patterns so subtle ethical violations become observable and labeled.

02

Quality refinement

Validator-Guided Refinement and clinical expert review enforce progressivity, actionability, ethicality, and supportiveness.

03

Supervisor SFT

Learn to localize violations, classify principles, and generate targeted explanatory feedback.

04

Growth-guided RL

GRPO rewards explanations that improve a frozen novice before-versus-after reading the feedback—without leaking the answer.

Data construction

Human-in-the-loop dataset construction

Clinical principles, controlled counselor–patient role-play, supervisor feedback, and expert review work together to produce high-quality dialogue–feedback pairs.

Overview of the human-in-the-loop dataset construction framework with clinical principles, counselor and patient role-play, supervisor feedback, data generation, and expert review
Figure 10. Overview of our human-in-the-loop dataset construction framework. Clinical counselors and supervisors define training principles; counselor and patient role-play generates controlled interactions; the supervisor provides targeted feedback; and clinical experts review instances to ensure dataset quality.

EthicScaff

Ethics-centered data built to teach

Real counseling data rarely labels subtle ethical failures. EthicScaff makes these failures controllable, visible, and pedagogically useful while preserving realistic client variation.

10,176 synthesized→VGR + expert review→9,915 retained
Each instance containsH → (s, m, f)
  • ✓
    Realistic dialogue historyBehavior-sensitive simulated client and novice counselor.
  • ✓
    Problematic utterance labelThe precise sentence where the violation appears.
  • ✓
    Ethical principle categoryFifteen expert-distilled violation types plus exemplary practice.
  • ✓
    Explanatory feedbackWhat went wrong, why it matters, and how to improve.
Automated ↔ expert refinement agreementκ = 0.87

Experiments

Better teachers produce better students

Evaluation spans downstream counselor behavior, objective ethical judgment, component ablations, expert assessment, and novice self-efficacy.

Violated principle classification94.37% F1

Qwen3-14B with the full framework, versus 38.67% for its base model and 73.12% for GPT-4o with RAG.

Full model94.37
GPT-4o + RAG73.12
Base model38.67
Ethical violation location74.24% F1

Qwen3-8B with the full framework achieves the best overall localization balance, including 63.03% Jaccard.

Full model74.24
GPT-4o + RAG71.76
Base model53.56
Downstream counselor quality

Our feedback improves all six clinical metrics

Compared with the unguided novice, gains are largest on MITI, WAI, and PSC—indicating better collaboration, alliance-building, and clinical appropriateness.

EFT-TFS+0.76
HTAIS+0.56
MITI+1.60
PSC+1.03
TES+0.89
WAI+1.38
Human-facing evidence

The scaffold transfers beyond simulation

Professional feedback quality improves under automatic and expert judgment. Novice students report significant gains across every assessed counseling competency.

Win, tie, and loss rates for feedback quality across five professional criteria
Higher feedback quality. Fine-tuned critiques win across objectivity, constructiveness, professional depth, comprehensiveness, and clarity.
Agreement between LLM and human evaluation of supervisory feedback
Experts agree. Automatic preference judgments show high consistency with manual expert evaluation.
Novice counselor self-efficacy before and after receiving supervised feedback across eight skills
8 of 8 skills improve. Six novice counseling-psychology students report significant self-efficacy gains after supervisory feedback (10,000-sample bootstrap, 95% CIs).

Responsible role

Support supervision—not replace clinicians

This work positions AI as an educational scaffold for low-stakes counselor training. It is not a patient-facing therapy system, a diagnostic tool, or a substitute for licensed clinical supervision. Its purpose is to help novices internalize professional constraints before working with vulnerable clients.

Citation

Read the paper

For the full methodology, experiments, and ethical considerations, see the arXiv manuscript.

Open PDF
BibTeX
@article{xu2025first,
  title   = {First, Do No Harm: AI Supervisor Scaffolds
             Novice Growth in Counselor Education},
  author  = {Xu, Chen and Lyu, Zhenyu and Lan, Tian and
             Yi, Yang and Ji, Yu and Ji, Luyao and Shen, Jian and
             Wang, Zhihua and Cui, Leyang and Zhang, Jieshuo and
             Wan, Xiaohua and Dong, Qunxi and Yang, Minqiang and
             Wang, Juan and Liu, Xiuling and Hu, Bin},
  journal = {arXiv preprint arXiv:2508.09042},
  year    = {2025}
}