I think this and OP's take is completely wrong. I would say it's going to be mostly non-useful data that's lost.

People will still write Wikipedia articles, they just won't be read as much, but the data on the site will still be valid.

People will still ask questions on stack overflow but there will be fewer questions and the number of trivial questions will significantly reduce as these are easier for AI to answer. But novel and difficult questions will still need to be asked to stack overflow, because AI isn't capable of answering them. And people will still want to and be interested in answering the more novel questions.

Thus, the overall effect will actually be to significantly improve the data, and engage people better. People wanting to answer questions will enjoy the experience much more with the less easy and obvious questions being removed. And they will be able to stumble across the interesting questions more easily and efficiently.

It should actually end up creating a much richer set of data for models to train on.

Think about it. If all the questions on stack overflow that are asked and answered were only questions that the models couldn't answer, that's like literally the most perfect training data. It filters out the questions it already knows the answer too which is not useful for it to read.

And this effect (I think it needs a name if it doesn't have one? - The Flex Effect if I just came up with it 😂😂) will only adapt over time with increased model output accuracy. As the model updates and its answers get better, so too will the questions, and subsequent answers. They'll get better and more and more difficult, matching the critea for what the new model then still needs to learn and train on.