[yonatron]

Yeah. These "lies" are just artifacts of the way that LLMs work. They're meant to predict likely text given a prompt. And they do. If tasked with "write some marketing or a buying guide for product X", they will simulate likely marketing blurbs, nothing yo do with truth, that's not their wheelhouse. Predictive is a very different function, algorithm and problem-set than something like "accurately summarize existing reviews". This is a feature, not a bug. If you use something off label, you'll get off label results. MSFT should know better.