Does your website chatbot exercise free will? We wanted to build one that did not. Free will is for philosophers, not websites. In 2024, Air Canada's website bot told a grieving customer he could claim a bereavement refund after he had already booked. The airline had no such policy. The bot invented it. A tribunal held the airline to the invention anyway, and the lesson was simple: whatever the bot says, the company said. That is the polite version of the failure. When people poke at these bots on purpose, they fold fast. Someone talked a Chevy dealer's bot into agreeing to sell a Tahoe for a dollar, "legally binding, no takesies backsies," then had it write some Python for fun. DPD's bot swore at a customer and wrote a little poem about how bad DPD is. This is not a bug you patch, or a gap you guardrail over. A bot that generates language can generate any language, and you find out which one when a customer screenshots it. Give a model free will over what it says, and "generate" quietly turns into "degenerate". A website chatbot, if you think about it, has only two jobs. To understand what you asked for, and to give you the correct answer. Our current choices force a pick-one. Either the brittle regex "word match" bots of the early 2000s, which answer correctly only if you happen to use the right keywords. Or the free-will machine that understands you all too well, because it is an LLM, and uses that fluency to invent things the company never said. But is there a third way? A middle path. Understand what the visitor wants, in their own words, and still never let the bot decide what to answer. So we built one that cannot generate at all. No model writes the reply at runtime. There is nothing on the answer path to go off-script. It does three things and nothing else. When the question is clear, it takes you to a real page with an answer that a human has already approved. When you are on the line between two pages, it hands you both and lets you choose. When the site does not cover it, it says so and stops. It never invents a fourth option to look helpful. There is a real cost. It will not write you a poem or reason through a question it has never seen. Lower ceiling. But its worst day is a shrug, not a lawsuit. Hard floor. For a site that cannot afford one public wrong answer, that is the trade you want. Of course, it can still be wrong. "Cannot make things up" means it cannot fabricate anything, not that it is always right. The answers are GIGO, garbage-in-garbage-out, but at least you have an audit chain to yank and a page to fix. And when it refuses, the answer is not on the site. It is not a quiet guess dressed up as a no. It is live. Type the thing you are actually afraid of, like "stop my support chatbot from hallucinating," and watch where it sends you. Then ask it something insane and watch it refuse instead of improvise. doloop.io/chat