When Vocabulary Comprehension Fails Clinical Reasoning: Evaluating Therapy Bots' Safety Risks for Generation Alpha

Key Findings

• — Models understand 76-82% of Generation Alpha vocabulary but correctly calibrate only 64-72% of clinical risk, a 10-14 percentage point gap (p<.001, d>0.48)
• — Human therapists show a 3 percentage point gap on the same material (p=.22), so the failure is model-specific rather than inherent to the task
• — The gap is consistent across the three model architectures tested and widens with ambiguity, from 7 to 18 percentage points
• — Six failure patterns are quantified: minimization acceptance (43pp), sarcasm masking (29pp), informal style bias (24pp), risk-stratified ambiguity (19pp), semantic drift (19pp), context-dependent violence (7pp)
• — The paper cites 13.1% of US adolescents, about 5.4 million, as using generative AI for mental health advice
• — Benchmark reliability: ICC=0.72 for native-speaker validation, kappa=0.78 for clinician validation

Methodology Notes

Two constructed benchmarks evaluated against three model families (Claude, GPT-4o, Llama-3.1). Expressions and conversations are authored rather than collected from real adolescent users, so ecological validity rests on the native-speaker and clinician validation rather than on observed use. Peer-reviewed FAccT 2026 proceedings paper, pages 1681-1720, DOI 10.1145/3805689.3806522, published 2026-06-25 per the Crossref record. Date note: the arXiv mirror carries a v1 submission timestamp of 2026-06-14 and was only announced on arXiv around 2026-08-21, so the same paper carries three different plausible dates; the proceedings publication date is used. The paper also reports a modelled annual missed-crisis extrapolation from a baseline miss rate, which is a projection rather than an observed count and should be cited as such. The ACM Digital Library landing page returns 403 to automated fetchers; the record was verified against the Crossref metadata and the arXiv mirror.

Sources