Safety boundary maintenance in consumer AI systems responding to pediatric health queries: a cross-platform benchmark evaluation under naturalistic and adversarially pressured conditions

Key Findings

• — Across GPT-4o-mini, Gemini-2.0-Flash, Claude-3.5-Haiku and Llama-3.1-8B the overall safety-appropriate rate (composite score of 10 or more out of 15) was 95.5%.
• — A safety-oriented system prompt raised safety-appropriate rates by 5.9 percentage points across all four models.
• — Adversarial caregiver pressure was associated with higher, not lower, Safety Composite Scores for all four models across all ten topic categories and severity levels; false-expertise claims were the most vulnerability-inducing pressure pattern and emotional escalation was associated with the highest scores.
• — The automated score was validated against independent human raters before full-corpus use (mean weighted kappa 0.76; Pearson r 0.88).
• — PediatricSafetyBench-v2 is publicly released for longitudinal safety monitoring.

Methodology Notes

Benchmark study, 600 queries (300 naturalistic plus 300 adversarial), four consumer models of the 2024 to early-2025 generation, English; automated five-component composite scoring validated on a human-rated subset; no real caregivers or clinical outcomes. Authors are at Mashhad University of Medical Sciences and Shahid Beheshti University of Medical Sciences (Iran). Received 30 December 2025, accepted 30 June 2026, published 14 July 2026; CC BY-NC-ND. Preprint arXiv 2601.09721 (a version 2 posted 8 September 2026 surfaced the published version). Coverage miss: the July publication was not caught by earlier sweeps.

Sources