Ultimately this is where ChatGPT and similar projects seem most likely to hit a wall. Text generators are statistical machines, but the relationship between statistics and truth is merely probabilistic and not identical. There is no way to train ChatGPT to say things that are true, because the number of true and false statements in the world are far too numerous and too specific to ever reliably delineate. At best, we will be stuck fact-checking ChatGPT using enormously complex knowledge bases and asking it to try again until something doesn't raise a red flag, at which point it may still not be true.
For most business purposes this is unacceptable. We will have machines that can generate untruths in a way that contravene common sense, so long as those untruths don't obviously contradict whatever local database is being consulted for grounding. And the mechanisms required to re-ground the GPT output even to that unhappy state will be so complex in their own right that businesses might as well just implement a similarly complex system to handle the issue directly, without calling out to a complex, licensed GPT. Not sure whether there is a way around this.