Safety is an engineering problem: you write the test, run it, watch it pass. But no test ever starts with 'am I swimming in whale poop?' How we test Una, and why real children are the only benchmark that matters.

Our AI guide Una passed every child safety test in English and missed nearly one in six in French. Here is why AI safety does not transfer across languages, and why we now test in every language.

We made our children's AI guide safer, and she lost her spark. The week we broke Una taught us that safety alone does not make an AI a child should talk to.