But that's not how modern chatbots are made!

There's no program with these words written into it.

Instead, they take a huge "corpus", some massive collection of human-generated text from all sorts of sources, millions of pages, and do statistical analysis on it to build a statistical model that spits out text based on the patterns from the text in the corpus.

This is called "training" the model, and it takes a long time, and at the end you get a black box with a name like a "neural network" that spits out convincing answers, but no explanation as to why those answers came out.

So there is no way to really "fix it the next day". There aren't any dials to just tweak these models with. The neural network (or other system) is just billions of numbers, no one of which has any real meaning at all on its own.

It isn't even a "programmer error" in any reasonable way! It's an artefact of the training method and the corpus - an "emergent property". That doesn't make it any better - it makes it worse. An error, you could fix.

The best you could do is to have a handwritten blacklist of embarrassing mistakes and filter those out, but you can see why that could be a Whack-A-Mole problem.