We are running a massive, uncontrolled experiment on our children. AI is now in the hands of millions of kids and teens, and we still do not really know whether it helps them thrive or puts them at risk. KORA starts from a simple idea: you cannot improve what you do not measure. We built the first independent, open-source benchmark to evaluate how safe frontier AI systems are for children, working with more than 30 child-safety experts to define 25 critical risks and test 35 models on hundreds of thousands of realistic synthetic conversations.
This talk walks through KORA's taxonomy, scoring method, and main findings: educational integrity as the industry's weakest spot, child-aware prompting as a powerful lever (+25 points across all models), and human redirection as the strongest predictor of overall safety. We also share what KORA does not yet capture, the limits of synthetic evaluation, and where we need pushback, replication, and collaboration to move the field forward.
Mon, 8 Jun 2026 16:00 - 16:45 (UTC+00:00) Dublin, Edinburgh, Lisbon, London
Register now