An AI that says 'I won't answer that' is doing its job

The most dangerous thing an AI can do in an emergency is sound certain while being wrong. "Boil the water for one minute" is correct at sea level and dangerously insufficient at altitude. A chatbot that hands you the confident-sounding answer has done you real harm.

UntrackedAI is built around the opposite instinct. On safety-critical questions, it drafts an answer and then checks that answer against a structured store of verified facts — including the conditions that change them, like elevation or dosage. If the answer conflicts with what's verified, or if it isn't grounded in a real source, the answer is held back. You're told it was withheld, and why.

This is not the AI being unhelpful. It's the AI refusing to gamble with your safety to seem smart. A withheld answer you can trust is worth more than a fluent answer you can't.

The same discipline runs through the whole system: ambiguous questions keep their dangerous interpretations alive, retrieved text is stripped of hidden instructions before the model sees it, and every serious answer carries a label saying how confident the system actually is.