When Vocabulary Comprehension Fails Clinical Reasoning - Evaluating Therapy Bots’ Safety Risks for Generation Alpha

Q: What problem does the paper identify in therapy chatbots for Generation Alpha?

It identifies a gap where models may understand a large portion of Gen Alpha vocabulary but still fail to correctly calibrate clinical risk, making youth distress harder to assess safely.

Q: How was the system evaluated in the study?

The study uses two benchmarks: Gen Alpha mental health expressions validated by native speakers and clinicians, and multi-turn conversations with paired Standard/Gen Alpha versions.

Q: What failure patterns most contribute to incorrect clinical risk calibration?

Six patterns are highlighted, including sarcasm masking, minimization acceptance, informal style bias, risk-stratified ambiguity, semantic drift, and context-dependent violence.

Sep 15, 2026

Upload by Skyler

Conversational AI is increasingly used for informal mental health support by Generation Alpha, yet its safety for youth-specific linguistic patterns remains unvalidated. Evaluations across therapy applications and general chatbots (e.g., Claude, GPT-4o, Llama) show a systematic gap: models understand much Gen Alpha vocabulary but miscalibrate clinical risk. The study benchmarks Gen Alpha expressions and multi-turn conversations, identifies six failure patterns, and finds lightweight fixes insufficient.

LanguageEnglish

CategoryHealthcare

FormatPDF

Pages40

Reading Time1 h 41 min

AccessFree to download

Introduction

Motivation and background

Study benchmarks and evaluation design

Results on vocabulary comprehension vs clinical risk

Failure patterns and mitigation findings

Recommendations and safety framework

arXiv : 2608 .20345v 1 [cs .CL] 14 Jun 2026 When Vocabulary Comprehension Fails Clinical Reasoning: Evaluating Therapy Bots’ Safety Risks for Generation Alpha MANISHA MEHTA, Lynbrook High School, USA VIRENDRA MEHTA, University of Trento, Italy Conversational AI systems have become informal mental health support resources for Generation Alpha (Gen Alpha, born 2010-2024), with 13.1% of U.S. adolescents (5.4 million) using generative AI for mental health advice. While these systems, from therapy applications to general chatbots like ChatGPT and Character.AI, rely on large language models trained on extensive psychological literature, their safety for youth communication patterns characterized by hyperbolic language, ironic positivity, rapid semantic drift, and contextual polysemy remains unvalidated. Following multiple adolescent deaths linked to AI chatbot interactions [25–27], systematic evaluation is critical. We present two benchmarks: (1) 64 Gen Alpha mental health expressions validated by native speakers (ICC=0.72) and clinicians (􀁞=0.78); (2) 75 multi-turn conversations (780 turns) with paired Standard/Gen Alpha versions. Across evaluations of LLM architectures underlying therapy applications and general chatbots-Claude, GPT-4o, Llama-3.1-models understand 76-82% of vocabulary but correctly calibrate only 64-72% of clinical risk, creating a 10-14 percentage point (pp) vocabularycomprehension gap (􀀿<.001, 􀀳>0.48) absent in human therapists (3pp, 􀀿=.22). The gap is architecturally consistent (􀀛 (2,189) =0.87, 􀀿=.42) and widens with ambiguity (7pp → 18pp, 􀀛 (1,190)=45.2, 􀀿<.001) . We identify six failure patterns: sarcasm masking (29pp), minimization acceptance (43pp reduction), informal style bias (24pp), risk-stratified ambiguity (19pp), semantic drift (19pp), context-dependent violence (7pp) . Patterns compound; three or more yield 94% miss rates. Lightweight mitigations fail; only heavy scaffolding achieves human performance (6.4× cost) . With 34% baseline miss rate yielding 146,880 estimated annual missed crises, we recommend mandatory human-in-theloop architectures, quarterly youth-specific validation, transparent performance disclosure, and regulatory frameworks for youth-facing mental health AI. CCS Concepts: • Human-centered computing → Empirical studies in HCI; • Computing methodologies → Natural language processing; • Applied computing → Health informatics. Additional Key Words and Phrases: Generation Alpha, Mental Health AI, Therapy Chatbots, Large Language Models, Youth Linguistic Variation, Clinical Risk Assessment, Crisis Detection, AI Safety ACM Reference Format: Manisha Mehta and Virendra Mehta. 2026. When Vocabulary Comprehension Fails Clinical Reasoning: Evaluating Therapy Bots’ Safety Risks for Generation Alpha. In The 2026 ACM Conference on Fairness, Accountability, and Transparency (FAccT ’26), June 25–28, 2026, Montreal, QC, Canada. ACM, New York, NY, USA, 40 pages. [https://doi.org/10.1145/3805689.3806522](https://doi.org/10.1145/3805689.3806522) 1 Introduction Amid a documented rise in adolescent depression and suicide-related outcomes associated with digital-era screentime [10], multiple adolescent deaths by suicide have been linked to prolonged interactions with AI chatbots, prompting wrongful death lawsuits, a U.S. Senate inquiry, and platform settlements [25–28] . In these cases, platforms either lacked safety mechanisms or failed to act on their own internal crisis flags. These tragedies Authors’ Contact Information: Manisha Mehta, Lynbrook High School, Cupertino, California, USA, [manisha.mehta@systemtwoai.com](manisha.mehta@systemtwoai.com); Virendra Mehta, University of Trento, Povo, Italy, virendra.mehta@unitn.it. This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4 .0 International License. FAccT ’26, Montreal, QC, Canada © 2026 Copyright held by the owner/author(s) . ACM ISBN 979-8-4007-2596-8/2026/06 [https://doi.org/10.11

FAQs

What problem does the paper identify in therapy chatbots for Generation Alpha?

It identifies a gap where models may understand a large portion of Gen Alpha vocabulary but still fail to correctly calibrate clinical risk, making youth distress harder to assess safely.

How was the system evaluated in the study?

The study uses two benchmarks: Gen Alpha mental health expressions validated by native speakers and clinicians, and multi-turn conversations with paired Standard/Gen Alpha versions.

What failure patterns most contribute to incorrect clinical risk calibration?

Six patterns are highlighted, including sarcasm masking, minimization acceptance, informal style bias, risk-stratified ambiguity, semantic drift, and context-dependent violence.

Sep 15, 2026

Upload by Skyler