I once spent 8 to 9 hours a day talking with AI while planning my career. I kept reading GPT-6 Pro’s long answers, but by the time the conversation ended, I felt drained and the conclusion was still unclear. At some point, I found myself thinking in the direction explained by the AI rather than the direction I had originally considered. I want to share this experience so that others can avoid some of the same trial and error I went through.
When You Don’t Even Know What You Don’t Know
At first, I didn’t even know what I didn’t know. So I threw entire problems with many variables, such as career planning and starting a business, at the AI. I didn’t adequately explain the life I wanted, the risks I could tolerate, or the resources I currently had. The AI would then give me plausible and wide-ranging answers. But I had no basis for judging whether those answers had actually identified the core of my problem.
The longer the AI’s answer was, the more it felt as though it had thought more deeply. Instead of evaluating its suggestions, I chose the next question from within them and read another long answer. The conversation continued, but my thinking actually became narrower. Some suggestions concerned problems someone had already solved, while others were not problems I wanted to solve in the first place.
This was also where I felt the difference when asking an expert. The difference was not simply the amount of knowledge, but which conditions someone who had actually faced the problem would check first. AI can also help reveal what I don’t know, but I had to recognize that gap and keep digging into it myself.
What I Can Entrust to AI and What I Have to Do Myself
AI is useful for organizing questions, suggesting overlooked perspectives, and pointing out gaps in my thinking. However, it is difficult for AI to make “the decision that is good for me” on my behalf. It cannot fully possess my experiences, priorities, and responsibilities. If I write down my hypothesis first, before reading the answer, I am less likely to be blindly pulled along by what the AI says.
My questions also needed constraints. Rather than asking, “What career should I choose?” it was more useful for me to ask, “I can spend 20 hours a week over the next six months, have almost no capital, and want to build something myself. Given these conditions, what experiments could I use to compare the two directions?” Constraints help AI provide better answers while also making me check what I consider important.
That is why I came to feel that domain knowledge remains important in the AI era. The more you know, the more narrowly you can frame your questions and the more easily you can spot flaws in the answers. Contrary to the expectation that better AI means we need to learn less, using AI properly required me to know what to ask it to do and what to verify.
Why Does It Seem Exceptionally Strong at Coding?
When I think about why coding agents are so powerful, programming has many explicit rules and rapid feedback loops. You can run the code, test it, and fix errors. Of course, real software development also becomes complex when requirements, users, and operating environments are intertwined. Even so, the fact that the results can be checked relatively clearly in many tasks makes it different from making decisions about one’s life.
I forgot this difference and expected to receive a clear right answer to career problems, just as I could with code. But the more I handed a complex world over to AI as-is, the more likely its answers were to drift toward average directions. Conversely, it became helpful when I set the goals and constraints myself and reframed them as questions that could be tested through small experiments.
Before Reading a Long Answer
These days, before asking AI, I try to check three things first: What do I currently believe? What do I still not know? What action will I take after getting this answer? I use the AI’s suggestions as material for testing these thoughts. I make the final decision myself, then think again based on the results.
I still believe that AI can be a tool for drawing out tacit knowledge. But I have to be the one who asks the tool questions, doubts what it says, and verifies it through action. More important than a long answer is whether my own thinking remains after the conversation ends.