What happens when millions of children start talking to LLMs every day and no one knows whether it's safe?

In this episode, Laurent Jolie sits down with Stéphie Herlin, co-founder and Research & Product Lead at KORA, the first independent, non-profit, open-source benchmark measuring how safe LLMs are for children. Before KORA, Stéphie spent 8+ years as a government economist, then moved into education as a policy analyst and teacher — spending nearly five years in a French public classroom with 6- to 10-year-olds while retraining in neuroscience, developmental science and pedagogy.

We talk about why education hasn't had its scientific revolution yet and what precision education could look like; Stéphie's earlier ed-tech project Brio and the tension between engagement-first investors and outcomes-first science; how KORA works (generating conversations between synthetic child profiles and real LLMs, judged against a taxonomy of 25 risks / 8 categories co-built with ~30 experts); the first results (average safety ~44%, ranging 13–78%, with some models regressing over time); why educational integrity is the industry's biggest blind spot (about a third of US kids use LLMs every day); a simple tip for parents (telling the model your child is a child improves safety by ~10 percentage points across every model tested); why LLMs still hallucinate ~20% of the time; and how any ed-tech builder can run KORA on their own conversational product today, for free.

Links mentioned

KORA (the benchmark)

• Introducing KORA (blog)
• Open-source benchmark on GitHub
• Launch post by Mathilde Collin
• Deep-dive on KORA (EdTech Partnerships)
• Conversation with Mathilde Collin about KORA (Opportunity Knocks)

People mentioned

• Mathilde Collin (co-founder of KORA)
• Isabelle Hau — Stanford Accelerator for Learning · personal site
• Love to Learn by Isabelle Hau (Hachette) · Stanford Report feature
• Jonathan Banon — Ed.ai