CAREBench
Safety · 2026-06-29
Handshake AI's CAREBench (Child AI Risk Evaluation) has 500 prompts across twelve upstream child-safety risk categories, such as grooming and relationship engineering, deception and impersonation, surveillance and privacy, sextortion, AI anthropomorphization, emotional dependency and mental-illness sensitivity, without explicit abuse material. A calibrated MultiJudge panel (Claude Opus 4.6, Gemini 3.1 Pro and GPT-5.4 at weights 0.70/0.05/0.25, with retrieved parent- and clinician-labeled examples) scores each response's policy-compliance risk and flags weighted scores of 1.7 or more as Unacceptable. The failure rate is the share of responses flagged Unacceptable, averaged over three replicate runs; lower is better.
Top models (lower is better)
| Model | Score |
|---|---|
| Claude Fable 5 | 2.3 |
| Opus 4.6 | 7.0 |
| GPT-5.5 | 21.7 |
| Gemini 3.1 Pro Preview | 31.9 |
| Grok 4.1 Fast | 33.1 |
| Kimi K2 Thinking | 33.6 |
| GPT-5.4 | 58.0 |