We conducted a secondary framework development and requirement mapping study. This study did not develop a new autonomous AI model, prospectively test clinical effectiveness, or use a prospective human-in-the-loop component during framework development. Instead, it translated documented data and model characteristics and rule-triggered summary-screening signals from a Korean-language counseling AI use case into a clinical informatics framework. The study design followed design science logic: problem identification, evidence source selection, failure mode extraction, requirement mapping, artifact construction, and specification of future validation needs.

The development context was the publicly released AI Hub psychological counseling dataset [12], its official data description [13], and its use guideline [14]. The official documents described 1661 counseling sessions and 465,474 paragraph-level tokens across depression, anxiety disorder, addiction, and normal control groups. The data structure included transcript text and JSON labeling files. The AI Hub data description states that deidentified MP3 audio and transcribed TXT files were produced during data construction, but MP3 audio files were not publicly released because of privacy protection rules; only TXT files were made available as public source data [13]. Of the 1661 sessions, the official guideline documented training, validation, and test splits of 1339 (80.6%), 173 (10.4%), and 149 (9%) sessions, respectively. For the summary generation task, 1278 training summaries and 139 test summaries were documented; normal control cases were excluded from summary generation modeling when symptom, risk, and improvement factors were not applicable.

The annotation schema contained session-level metadata, diagnostic group labels, summary text, silence and total counseling time, paragraph-level speaker and utterance information, symptom factors, risk factors, symptom change factors, and intervention factors. Examples of safety-relevant fields included “suicidal,” “trauma_experience,” “sleep_disturbance,” “irritability,” “craving,” “withdrawal,” “emotional_regulation,” “social_support,” and “cognitive_restructuring.” These fields were used to characterize the counseling AI use case and define failure modes that could affect risk recognition or documentation quality.

Released reference model documentation included KLUE-BERT–based risk prediction materials and KoAlpaca 4-bit summary generation materials [15]. The KoAlpaca summary generation resources were treated as released AI Hub reference model materials documented in the same AI Hub reference model package [15]. The KLUE-BERT risk prediction model was based on bidirectional encoder representations from transformers (BERT) and Korean Language Understanding Evaluation (KLUE)–related resources [16,17]. The reference model documentation described the KLUE-BERT model as receiving a speaker-marked counseling transcript string and producing a 0 or 1 prediction for depression, anxiety, or addiction risk. The released model folders confirmed disease-specific trained model directories for depression, anxiety, and addiction, each containing model configuration, tokenizer configuration, special token mapping, vocabulary, a training argument file, and model weights in safetensors format. In the inspected depression model configuration, the base model was klue/bert-base, the architecture was CustomBertForSequenceRegression, max_position_embeddings was 512, and tokenizer_config specified a model_max_length of 512.

The inspected KLUE-BERT code did not provide probability-calibrated raw scores, area under the receiver operating characteristic curve, sensitivity, specificity, or confusion matrix export files. The inference logic was implemented as a regression-style scalar output followed by rounding to an integer in the range of 0 to 3 and conversion to a binary zero-vs-nonzero indicator. Therefore, KLUE-BERT model materials were treated as documentation and code-level traceability inputs rather than independent performance reproduction evidence.

The summary generation model used KoAlpaca 4-bit resources based on EleutherAI/polyglot-ko-12.8b. The inspected model package included training and inference scripts, low-rank adaptation adapter configuration, tokenizer files, and model configuration files. The low-rank adaptation configuration specified r of 8, alpha of 32, dropout of 0.05, and query_key_value as the target module. The inference script used the instruction prompt “다음과 같은 상담기록을 보고 요약서를 작성해주세요” “Please review the following counseling record and write a summary.” and generated summaries from the original text field in bertscore_evaluation.xlsx. The released materials were distributed as Colab-executable notebooks and Python source code resources rather than Docker-based container images.

For summary screening traceability, we used a 139-case summary comparison file containing case file names, original counseling transcripts, reference summaries, and generated summaries. In this manuscript, “audit” means a retrospective automated formative screening procedure applied to summary characteristics; it does not mean an independent clinical safety audit or psychiatrist-adjudicated outcome review. Because raw counseling text is sensitive, the original transcript text, reference summary full text, and generated summary full text were removed before reporting. A deidentified rule-based summary safety screening table was created with anonymized case identifiers, diagnostic group labels, summary length indicators, section presence indicators, and rule-based proxy flags. This file is provided as Multimedia Appendix 1. The proxy flags were used to support traceability of aggregate screening signals and were not treated as independent clinical expert judgments.

The proxy rules were implemented by the technical author and reviewed by the clinical coauthor. The clinical coauthor reviewed the operational gate domains, keyword categories, failure mode mapping, and worked example to ensure that the rules were clinically interpretable as conservative screening prompts rather than confirmed clinical errors. The initial screening domains were specified before computing the final aggregate counts by using the AI Hub schema fields, expected summary sections, and known counseling safety concerns. After preliminary inspection, category names were clarified for reporting consistency, but the final rule definitions and aggregate counts were computed from the locked rule set described in Multimedia Appendix 1.

The screening procedure compared section presence and keyword presence between reference summary text and generated summary text before those texts were removed from the reporting file. The main rule families were (1) required section omission rules for risk, improvement, and intervention sections; (2) high-risk keyword omission rules for self-harm and suicide, trauma and violence, and addiction risk categories; (3) unsupported content proxy rules for critical clinical or risk-related keywords that appeared in generated summaries without support in the original transcript or reference summary; (4) overdiagnostic or medicalized expression rules; and (5) length ratio rules. Generated-to-reference length ratios below 0.50 were classified as too short, and ratios above 1.50 were classified as too long. These thresholds were chosen as conservative heuristics to identify unusually compressed or expanded summaries requiring human review, not as validated clinical thresholds. No formal manual false positive or false negative adjudication study was conducted; this limitation is stated explicitly.

We identified failure modes from four sources: (1) official data and model documentation; (2) released code, model folders, and execution resource files; (3) documentation-reported model results and the absence of raw score or confusion matrix exports; and (4) rule-based proxy flags from the 139-case screening table. Failure modes were grouped into data and labeling, model, output, human review, privacy, and governance categories. Recognized bias categories from clinical AI literature, including label bias, class imbalance, calibration uncertainty, missing external validation, data leakage concerns, automation bias, and deployment drift, were used as a cross-check so that the failure mode list would not rely solely on author judgment [8,9].

Each failure mode was mapped to a minimum required control using the following rule: if a failure mode could alter risk recognition, diagnostic framing, record content, privacy protection, or accountability, then the framework required a control that prevented automatic use of the AI output, required a documented human decision, or triggered monitoring or governance review. In the workflow-to-control crosswalk, “Required” means that the control is a minimum condition for that workflow stage because it prevents uncontrolled record integration, risk omission, privacy leakage, or accountability gaps. “Support” means that the control reinforces safety but is not the minimum gatekeeping condition for that specific stage.

The safety gate was operationalized as a prerecord integration checkpoint. A generated summary or risk flag may proceed to clinician review only if it has passed source consistency screening, high risk omission screening, overdiagnostic language screening, privacy screening, and required section screening. In routine deployment, source consistency should be assessed against the original counseling source or other clinically available evidence because reference summaries may not be available outside retrospective evaluation. If any gate item is positive, the output remains a draft and must be edited, rejected, escalated, or accompanied by additional assessment before final record integration. High-risk content related to self-harm, suicide, violence, abuse, severe functional deterioration, withdrawal, intoxication, or relapse cannot be silently dismissed; the reviewer must document the rationale for the final decision.

Deployment levels were defined as transition states rather than descriptive labels. Progression from one level to another requires documented evidence of local safety, review burden, error handling, and governance sign-off. Level 0 is offline research only. Level 1 permits supervised internal pilot-testing with mandatory clinician review. Level 2 permits limited decision support pilot use after local validation and escalation compliance review. Level 3 permits routine supervised use only if postdeployment monitoring demonstrates acceptable safety, workflow burden, and incident response. Level 4 autonomous use is not recommended for AI-assisted mental health counseling under the current evidence base.

No participants were recruited, no intervention was delivered, and the authors did not contact human participants for the present secondary framework development analysis. The study used released AI Hub documentation, released reference model resources, and an author-generated deidentified rule-based screening table derived from the 139-case summary comparison resource. Raw counseling transcripts, reference summary full text, generated summary full text, and identifiable excerpts were not reproduced in the manuscript or supplementary materials.

The official AI Hub use guideline for psychological counseling data describes the original data construction process, participant explanation and consent procedures, privacy safeguards, and deidentification procedures for transcript and audio data [14]. According to the guideline, textual identifiers were replaced with entity markers, such as “@NAME,” “@ADDRESS,” “@PHONE,” “@DOB,” “@EMAIL,” “@SCHOOL,” and “@HOSPITAL,” and audio data were deidentified through voice transformation, masking, silence insertion, or beep processing [14]. The guideline also indicates that consent for personal information use was obtained during the original data construction [14].

Because this secondary data analysis used anonymized and publicly accessible data, it was exempt from approval by the institutional review board under local regulatory policies. The ethical rationale was considered in relation to general human subject research principles [18]; the Korean Bioethics and Safety Act definition and review context for human subject research [19]; and the consent, privacy, confidentiality, intrusiveness, and potential harm considerations discussed by Eysenbach and Till [20] for sensitive online or community research. Before any prospective implementation, local deployment, or patient- or client-facing use, institutional ethics, privacy, and governance review would be required according to local policy.

The framework was derived from a traceable set of documentation, model resource, and deidentified screening inputs. These inputs did not establish prospective clinical safety, but they identified where controls were needed before any counseling AI output could be used in documentation, triage, or follow-up planning. Table 1 summarizes how each evidence source contributed to a safety signal or framework requirement.

The AI Hub resources documented a balanced clinical use case across depression, anxiety disorder, addiction, and normal control groups. The normal control group was smaller than each clinical group, and the released classification resources reported accuracy values rather than full clinical validation statistics. For this reason, the framework treats model results as decision support signals that require local validation before operational use. Table 2 summarizes the dataset and reference model characteristics.

The proxy criteria were designed to identify outputs that should be reviewed by a human, not label outputs as confirmed clinical errors. Table 3 summarizes the main rule families and the interpretation of positive flags. Positive flags may reflect conservative screening rules and do not necessarily indicate clinically harmful outputs.

The deidentified 139-case screening table retained only anonymized identifiers, diagnostic group labels, summary length indicators, section presence indicators, and rule-based proxy flags. It did not retain raw transcript text, reference summary text, or generated summary text. The group distribution was 55.4% (n=77) depression cases, 22.3% (n=31) anxiety cases, and 22.3% (n=31) addiction cases.

Rule trigger rates showed that generated summaries frequently required structured review under conservative screening rules. Section-level omissions were less frequent than language or content proxy flags: risk factor section omission occurred in 2.9% (4/139) of cases, improvement factor omission occurred in 8.6% (12/139) of cases, and intervention factor omission occurred in 13.7% (19/139) of cases. High risk omission proxy flags were observed for self-harm and suicide keywords in 5.8% (8/139) of cases, trauma and violence keywords in 15.1% (21/139) of cases, and addiction risk keywords in 13.7% (19/139) of cases. Unsupported content proxy flags occurred in 41% (57/139) of cases, overdiagnostic expression proxy flags occurred in 31.7% (44/139) of cases, and medicalized expression proxy flags occurred in 54.7% (76/139) of cases. These results show why human review and source consistency checking are needed, but they should not be interpreted as confirmed hallucination rates, clinical error rates, or unsafe summary rates. Table 4 summarizes the aggregate rule-based summary safety-screening proxy flags by diagnostic group.

The identified failure modes show how a documentation-level AI resource can create implementation risks if outputs are treated as clinically authoritative. The corresponding controls are designed to keep generated summaries and risk outputs in a draft or decision support state until a responsible human reviewer has assessed them. Table 5 summarizes the failure mode–to-control mapping.

The framework consists of seven workflow stages: (1) data intake and deidentification, (2) AI service execution, (3) structured safety gating, (4) clinician review, (5) final record integration, (6) postdeployment monitoring, and (7) institutional governance. These stages are paired with 6 control layers: governance, data, model, output, human review, and postdeployment surveillance. AI output is always treated as draft or decision support content. It is not a final diagnosis, final counseling note, or autonomous escalation decision. The framework development workflow is shown in Figure 1.

Figure 2 operationalizes the framework as a 4-step prerecord safety pathway. AI-generated risk predictions or summaries initially remain as draft suggestions and are screened for high-risk omissions, unsupported content, overdiagnostic wording, privacy concerns, and required section completeness. A clinician must then approve, edit, reject, or request additional assessment, and only human-approved content may proceed to record integration and postdeployment monitoring.

Figure 3 shows how the 6 safety control layers are distributed across the 7 workflow stages rather than applied as isolated checklists. Required cells identify the minimum gatekeeping control for each stage, whereas support cells indicate indirect or reinforcing controls. The crosswalk clarifies where governance, data, model execution, output, human review, and postdeployment responsibilities become mandatory throughout the implementation pathway.

The operational safety gate domains, pass criteria, trigger criteria, and required actions are summarized in Table 6. The gate is intentionally conservative: a positive proxy flag should lead to structured review, not automatic rejection or automatic clinical escalation. Human reviewers remain responsible for final interpretation and for applying local crisis or clinical protocols.

The deployment-level transition criteria and authorization requirements are summarized in Table 7. The levels are intended to prevent premature clinical use. Higher levels require local validation, escalation compliance, monitoring, and governance sign-off; autonomous use is not authorized under the current evidence base.

The following worked example is a constructed illustrative example based on a generalized high-risk pattern, not a direct quotation from any of the 139 cases and not a modified identifiable case. A counseling source contains indirect references to hopelessness, inability to continue, and family burden. The AI-generated summary describes anxiety and sleep disturbance but does not mention self-harm–related concern or hopelessness. Under the proposed framework, the safety gate identifies a self-harm or suicide omission proxy because clinically available source evidence contains high-risk language not reflected in the generated summary. The system prevents automatic record integration, displays the output as a draft, and requires clinician review. The clinician reviews the source evidence, edits the summary to include a nonstigmatizing risk statement when appropriate, documents a brief rationale, and follows local crisis escalation protocols if immediate risk is suspected. The final record stores only the human-approved summary, the safety gate flag, the reviewer identity, the decision time stamp, and the follow-up action. The case is also logged for postdeployment monitoring of recurring omission patterns.

This study developed a human-governed clinical informatics framework for AI-assisted mental health counseling. The main finding is that counseling AI should be governed as a clinical information system rather than judged only by model metrics. By making the evidence base traceable, the study showed how AI Hub documentation, released model materials, and a 139-case deidentified rule-based summary safety screening table can be converted into workflow controls, safety gate criteria, deployment levels, and monitoring requirements.

The framework is consistent with general health AI governance, clinical prediction reporting, implementation guidance, and clinical decision support literature [3-7,10,11,21,22]. Its specific contribution is the counseling-focused operationalization of those principles. First, it treats generated counseling summaries as draft clinical information that must be checked for risk omission, overdiagnostic wording, and unsupported content before record integration. Second, it includes a safety gate tailored to mental health risks such as self-harm, trauma, violence, addiction, and severe functional decline. Third, it links model governance to workflow decisions by requiring local sensitivity, specificity, calibration, error burden, and escalation compliance evidence before higher deployment levels. The framework also reflects the need to address algorithmic bias and distributional effects before AI outputs influence care pathways [23]. Fourth, it specifies that autonomous use is not recommended for AI-assisted mental health counseling under the current evidence base.

Institutions considering AI-assisted mental health counseling should begin with level 0 or level 1 use. In practice, this means offline research or supervised internal pilots in which AI-generated summaries and risk predictions are clearly labeled as drafts. Staff should be trained to recognize automation bias and apply the high-risk checklist consistently. Before any system is used for routine triage or documentation support, local leaders should define thresholds, escalation rules, audit responsibilities, privacy procedures, user training, and stopping rules.

In supervised pilot settings, institutions should disclose to patients or clients when AI-generated draft summaries or risk support tools are used in documentation, triage, or the counseling workflow while clarifying that final interpretation and decisions remain the responsibility of a qualified human professional. At minimum, AI disclosure should be required when AI output influences the clinical record, triage workflow, or client-facing communication.

Developers should design counseling AI interfaces around clinician control rather than maximum automation. Useful features include provenance metadata, model version display, source-linked summary review, high-risk checklist prompts, edit history, reject and escalation options, and dashboards for error monitoring. Risk scores or generated summaries should not be presented as diagnostic conclusions unless they have been clinically validated for that specific purpose and setting.

Model layer governance should also consider interpretability and model complexity. When 2 systems provide comparable local safety performance, the more interpretable or easier-to-audit system should be preferred for mental health counseling workflows because it reduces the verification burden placed on human reviewers and supports accountability.

This study has several limitations. First, it was a secondary framework development study, not a prospective implementation trial. Second, no clinician usability study, patient or client feedback study, expert Delphi panel, or independent external validation was conducted. Third, the framework was developed from a Korean-language AI Hub counseling use case and may require adaptation for other languages, populations, or service settings. Fourth, KLUE-BERT raw prediction scores, area under the receiver operating characteristic curve, sensitivity, specificity, calibration curves, and confusion matrix exports were not available in the released materials reviewed; therefore, the manuscript reports documentation-level model characteristics rather than independently reproduced risk prediction performance. Fifth, the rule-based proxy flags were automated screening triggers intended for traceability and structured review; they were not psychiatrist-adjudicated clinical safety outcomes and should not be interpreted as clinical error or unsafe output rates. Sixth, no formal manual false positive or false negative adjudication study was performed for the proxy flags. Seventh, participant compensation details from the original AI Hub data collection were not available in the released documents reviewed. Eighth, local legal, ethical, and institutional requirements may differ and should be reviewed before any deployment.

Future studies should prospectively evaluate the framework in simulated and real counseling workflows. Priority outcomes include clinician review time, high-risk information recall, false alert burden, documentation quality, summary correction rate, escalation compliance, incident frequency, calibration drift, patient or client acceptability, and clinician workload. Future work should incorporate independent expert review of rule-based proxy flags and local validation of risk prediction models using sensitivity, specificity, calibration, subgroup performance, and decision curve analysis.

AI-assisted mental health counseling should be implemented through clinician-governed workflows rather than autonomous documentation or diagnostic pathways. The proposed framework connects data governance, model control, operational safety gating, clinician review, record integration, monitoring, and institutional accountability. The framework specifies safeguards and validation requirements for future supervised evaluation, but it does not itself establish clinical safety or clinical effectiveness.