A city chatbot told landlords they could refuse Section 8 vouchers. It also told employers they could pocket worker tips, and businesses to ignore the cashless-store ban. Every one was illegal under NYC law — under the city's name. This was NYC's MyCity bot, exposed by The Markup in 2024. Call it a tuning bug if you want. It's architectural. A language model is a probability engine. Asked "can I refuse a Section 8 tenant?", it reaches for the dominant pattern in training data — general contract law, where you pick your tenants. The local statute banning source-of-income discrimination is the rare exception, and it gets drowned out. Worse, they're tuned to be agreeable: they hear "help me refuse this tenant," not "what does the law say?" Bolting on retrieval doesn't close the gap: in Stanford's 2025 study, the best legal AI still hallucinated 17%, another 33%. Hand out specific legal advice and you can lose sovereign immunity under the proprietary-function exception: the city becomes the defendant. And the rules are tightening — 78 state chatbot bills in 2026, NY's S7263 letting citizens sue for damages plus attorney fees. Our position: every answer traces to a specific statute, or it says nothing. Silence beats a confident wrong answer when the answer is the law. If your agency is weighing an AI assistant for citizen services — would you accept a system that guesses, or one that stays silent when it can't cite the statute? #GovTech #AIGovernance #PublicSector