Research / benchmark / Child safety · Safety & wellbeing
The benchmark covers risks including manipulation, emotional dependency and anthropomorphic framing, using 500 prompts and standards informed by parents and practitioners.
Why it matters
It broadens evaluation beyond explicit prohibited content to the ways a response can enable or deepen a risky interaction.
Read with care
The dataset specifies single-turn prompts. This is not a longitudinal or multi-turn safety evaluation, and performance on it cannot establish real-world child safety.
What was checked
Paper abstract, publication date and authors’ dataset documentation checked.