What is your programming stuff that SO can help with that generative AI can't?

I think you are conflating the concepts. Getting help from SO does not always mean asking a new question on SO. In fact in most cases it doesn't, because after so many years of development SO has built up a fairly comprehensive database of knowledge that most simple questions are already answered. I would say most of the time I need help with programming I found SO is the site that brings me the final answer, usually without me asking a new question. Let's say SO has helped me with 1000+ concrete programming questions without me asking a question in the end, you are not going to see this 1000+ data anywhere¹ from the raw statistics. You're asking for personal experience, and this is my personal experience.

How much research effort is expected of Stack Overflow users? A lot. For me, using a generative AI is just part of my research. I typically take generative AI as another form of glorified search engine, with the caveats kept in mind.

• Results from a traditional search engine can land me on the page that contains the necessary context to evaluate the quality of the content, which is missing in the case of genAI. Compare the following scenarios:
  1. Scenario 1: A traditional search engine lands me on a blog. Then I can go and check another blog of the same blogger. If I find that the other blog shows fundamental misunderstanding (for example, claiming that an array is a const pointer in C), I will put less trust in the blog and maybe move on to the next search result.
  2. Scenario 2: A traditional search engine lands me on an SO answer. The vote system can give me a score that gives me a general idea whether it is trustworthy or not. If I discover something I find ridiculous in the answer, but turns out the answer is from someone with a golden tag badge, I am inclined to believe that it is me that has the fundamental misunderstanding and double check and learn. In fact this is exactly how I got rid of the "An array is a const pointer in C" misunderstanding that I acquired from elsewhere.
  3. Scenario 3: Traditional search engines give me nothing that immediately answers my question and I decide to give genAIs a shot. Now I have absolutely no external measures to evaluate the quality of the response. Examining other responses by the same genAI tells me nothing about the quality of this response. Nor are there anything like "golden tag badges" to ensure the quality of the response is up to par. I need to vet the response carefully myself.
• GenAIs are very reluctant to disobey or even just disagree with the user. Even though they try hard to appear as unbiased as possible, in most cases it just ends up with distorted responses. This is one thing I particularly dislike about genAIs. Even if you ask about something that is generally (not yet univocally) considered bad practice, genAIs will list 3 points of pros and 3 points of cons and end the response with the signature phrase "Ultimately, ... depends on...". This kind of response is not entirely garbage and can actually help if you have enough critically thinking, but it is kind of misleading. On SO you'll see people point out various defects of your code in comments. Admittedly sometimes the comments are blunt or even borderline rude, but such feedback with strong disagreement is necessary if you are honestly seeking improvement. GenAIs spoil every asker in the same way that impedes them from learning from mistakes they unconsciously make. I digress a bit here, but our sister site, Code Review, can never be replaced by GenAIs in the slightest way. Don't even get me started on XY problems. GenAIs can never generate responses that say "Why do you do X? It seems strange to me, what is the Y you're seeking? There may be better approaches." This is exactly due to the lack of the ability of genAIs to disagree with the asker.

In a sense, the attitude taken by genAIs is preposterous. Typically, a genAI assumes the user is almost always correct, and at the same time it assumes it knows everything, evidenced by that it chooses to make up plausible responses instead of just admitting "I don't know". However, "the user is always right" and "I know the answer to everything" is already a logical contradiction and when the contradiction manifests itself in practice you see all kinds of dramas in genAI responses.

With all these said, it doesn't mean genAIs are completely useless and many times genAIs have brought me attention to the blindspots in my initial evaluation of the problem. I didn't have the slightest idea that a keyword is what I should look for, or that a convenient library function exists and after asking a genAI I do and can investigate deeper. The take-away message is that using a genAI is just another part of my research before I would actually ask a question on SO. GenAIs offer extra tooling, but in my mind they are just another form of glorified search engines with novel caveats that traditional search engines don't have.

^{1_{Not really, the number of my saves and upvotes can actually kind of tell that.}}

My suggestion is that Stack Overflow should deviate from that aspect of what generative AIs can do and maybe SO should focus more on new directions and features that can greatly help many users.

Stack Overflow doesn't control what users ask here: we can only answer what people feel like bringing up here if they can do so in a way that is constructive and useful.

Unfortunately... this trend of just dumping coding questions into GPT and running with the result is going to make it far less likely that new questions asked here will be constructive and useful. It trains users to ask questions in a way that gets results from an LLM, but doesn't provide what a real developer would need to reach useful outcomes without wasting valuable time.

Asking a group of users on the internet who don't have access to your codebase how to solve a given problem is an acquired skill: it requires many of the same skills that you'd need to solve the problem on your own.