There’s an old piece of advice that sounds almost too obvious to be useful: if you want better answers, ask better questions. I used to hear that mostly as a communication lesson. Ask someone a vague question and you’ll probably get a vague answer. Ask something specific and you have a better chance of getting something useful back. Fair enough. But the more I work with AI, build websites, research ideas and try to solve problems that don’t have tidy little answers, the more I think the principle goes much further than that. The quality of the question can influence the quality of the research that follows it, the decision you eventually make, the system you build and even the problem you decide is worth solving in the first place. Sometimes the answer isn’t the problem. Sometimes the question got there first.
A Correct Answer Can Still Be Useless
Suppose I ask how to increase traffic to a website. There are plenty of technically correct answers. Publish more content. Improve SEO. Build backlinks. Promote the site. Improve internal linking. Target better keywords. Nothing particularly wrong with any of them. But what if traffic isn’t actually the problem? Maybe the site is attracting people who have no reason to stay. Maybe the pages don’t make the purpose of the site clear. Maybe Google hasn’t indexed much of it yet. Maybe there are plenty of impressions and almost nobody is clicking. Maybe the business doesn’t need more visitors nearly as badly as it needs a clearer offer.
“How do I increase traffic?” can therefore produce an excellent answer while sending me in completely the wrong direction. That’s one of the stranger things about problem-solving. We tend to judge the answer because that’s the part sitting in front of us, while the question that produced it gets a free pass. AI makes that especially worth paying attention to because it can answer a poorly framed question with remarkable efficiency. You can be heading down the wrong trail with beautifully formatted bullet points before the shepherd has even finished sniffing the first telephone pole.
When the Answer Makes Me Question the Question
This happens to me constantly when I’m working with AI. I’ll ask something, get a perfectly reasonable answer and still think, No. That’s not quite it. What matters is what happens next. I’m not trying to keep rewording the question until the AI finally gives me the answer I wanted in the first place. That would just be confirmation bias with better grammar. What usually happens is that the unsatisfying answer makes me reconsider the question itself. Maybe I asked it too broadly. Maybe I left out a piece of context that actually matters. Maybe I made an assumption without noticing it. Maybe I’m asking about the visible problem when the real one is sitting underneath it. So I rephrase the question, ask it from another angle and see what changes.
Sometimes the new question produces a better answer. Sometimes it produces another answer that still doesn’t quite fit, which sends me around again. By the time the conversation settles, I may be asking something substantially different from where I started because the exchange has helped me understand what I was actually trying to solve. That distinction matters. I’m not interrogating AI until it agrees with me. I’m using my reaction to the answer as information. The first question hasn’t necessarily failed; sometimes its job was simply to expose the better question underneath it.
The Conversational Vending Machine Has Its Place
There is also a very ordinary and extremely useful way to use AI. Ask a question. Get an answer. Carry on with your day. How do I paint a door without leaving brush marks? What can I cook with what’s left in the fridge? Can you summarize this document? Why is this spreadsheet formula giving me an error? How long should I roast this? There is nothing unsophisticated about using AI that way. Sometimes you need an answer, not a philosophical journey into the nature of door paint.
Research into actual ChatGPT use shows just how important these practical uses are. OpenAI researchers analysed 1.5 million consumer conversations and found that roughly three-quarters fell into practical guidance, seeking information and writing. About 70% of consumer use was classified as non-work rather than work-related. That matters because AI is often discussed as though everybody is either building autonomous agents or preparing to hand their occupation over to a robot. A great deal of actual use is much more ordinary. People are trying to get things done. I sometimes think of that first level as the conversational vending machine. Put in a question, get something useful out. That description isn’t meant as an insult. Vending machines are quite handy when you’re thirsty. The limitation appears only when we assume that getting an answer is all the machine can do.
The Conversation Itself Has Value
This, to me, is where conversational AI becomes much more interesting. You can stay in the conversation. You can challenge the answer, disagree with it, explain why it doesn’t fit, add a piece of context you forgot the first time, change your mind halfway through or come back to something from earlier and say, “No, that’s actually the part we need to look at again.” You can also do all of that at your own pace, and that creates a very unusual environment for thinking. There is a great deal of public discussion, much of it justified, about the risks of conversational AI. People can become overdependent on it. They can accept confident answers too easily. They can use it as a substitute for thinking instead of something that supports thinking. Those are real concerns, but they don’t describe the only possible relationship with the technology.
Used differently, the same conversation can force you to explain yourself more clearly. It can expose an assumption you hadn’t noticed. It can make you defend a position, reconsider one or admit that you still don’t know enough to decide. I’ve become pretty immersed in this myself. I use conversational AI across a ridiculous range of ordinary life now: writing, research, websites, planning, cooking, shopping, decisions and whatever else happens to be in front of me. A lot of the time I’m not even issuing some carefully constructed command. I’m simply talking about what I’m doing.
If I mention that I’m heading out to run errands and one of those stops connects to something I’ve previously said I need, the useful response isn’t necessarily to wait for me to ask, “What was on my shopping list?” The conversation already has enough context to recognize what matters now and bring it forward. The same principle can apply to something I meant to follow up on, unfinished work, an appointment, a decision I said I needed to make or any number of other loose ends that otherwise depend on me remembering where I left them. None of those examples is particularly earth-shattering by itself. I’m not suggesting civilization has been transformed because I no longer have to remember the mustard. What matters is the continuity.
That continuity can also produce things I wouldn’t necessarily have thought of as AI tasks in the first place. Take something as ordinary as walking the dogs. Over time I can talk about where we’ve gone, the routes we’ve taken and the places that have become part of the routine. Once that information is available, I can turn around and ask AI to pull those pieces together, map the routes and show me the pattern that has accumulated. The individual conversations were just conversations. Put together, they become something else. That’s what I find increasingly interesting about being immersed in conversational AI. I don’t always know in advance what I’m going to do with the information I’m giving it. Sometimes the useful application only becomes obvious later, because enough context has accumulated for a new question to become possible.
The interesting part is the continuity. Sometimes enough context accumulates that a useful next action—or even a better question—becomes obvious later.
The tool stops being somewhere I visit only when I need an isolated answer. The conversation begins connecting information, intentions and decisions across different parts of what I’m doing. I can say what’s happening, and sometimes the next useful action is already sitting in the context. That begins to feel less like consulting a machine and more like working inside an ongoing conversation. Of course AI has flaws. It can be wrong. It can miss context. It can sound remarkably certain while standing on fairly thin ice. Human judgment doesn’t disappear simply because the conversation is useful. But the capability itself is worth appreciating. For most of human history, if you wanted to explore an idea across ten different disciplines at eleven o’clock in the morning, you needed access to a very impressive library and probably several unusually patient people. Now you can begin that conversation from a desk, a phone or wherever you happen to be thinking. Why wouldn’t we learn how to use that well? Any human who has spent enough time around a German Shepherd already understands the value of something willing to remain intensely interested in the same subject long after everyone else in the room has moved on.
The Answer Can Become Material for the Next Question
Once the conversation becomes iterative, the process starts to look less like retrieval and more like this: question → answer → reaction → refinement → better question → better decision. The reaction in the middle is where a lot of the value lives. You challenge an assumption, ask for evidence and point out what doesn’t fit. You introduce information the model didn’t have, or ask it to argue the other side. Sometimes you change the objective entirely and discover that two problems you thought were separate are actually connected. The answer becomes material for the next round of thinking rather than the final product. Eventually you may end up somewhere neither the first question nor the first answer could have taken you.
That collaborative pattern also shows up in research. Anthropic’s Economic Index work has separated AI use into automation, where the system largely performs the task, and augmentation, where a person and the AI work through it together. Its analyses have found substantial augmentation use in conversational AI, while API use tends to be more automation-heavy. Those figures shouldn’t be stretched into a universal claim about how everybody uses AI. Different platforms attract different users and different kinds of work. But they do support the broader point: iterative, collaborative use is real. People are not only handing tasks to AI. They are also using it to learn, compare, revise, question and work through problems.
Fast Answers Raise the Value of Good Questions
Before generative AI, weak questions carried a kind of natural penalty. Research took time. Drafting took time. Comparing alternatives took time. If you wanted to investigate an idea properly, there was friction everywhere. AI removes a remarkable amount of that friction, which is mostly a very good thing. It also means we can now investigate a bad premise much faster.
A model doesn’t necessarily know that the question you asked isn’t the question you should be asking. It sees the problem as you presented it and tries to help. If I ask for ten ways to improve a terrible strategy, I may get ten very competent suggestions for improving a terrible strategy. If I ask for evidence supporting an assumption without asking whether the assumption itself holds up, I can subtly steer the entire research process. If I ask AI to choose between Option A and Option B, I may never discover that Option C makes both of them irrelevant. This is where human judgment enters the process in a very practical way. Someone still has to notice when the answer doesn’t make sense in context. Someone has to ask what is missing, whether the problem has been framed correctly and what actually matters.
A 2025 Microsoft Research study involving 319 knowledge workers found something similar. As people used generative AI, some of their critical-thinking effort shifted toward verifying information, integrating AI responses and overseeing the task. In other words, easier production didn’t make judgment disappear; some of the thinking simply moved to a different part of the process. That makes the quality of the question more important.
Better Questions Improve More Than AI Prompts
This is where the subject gets bigger than prompting. A good question can change a research project: What evidence would prove me wrong? It can just as easily change a business decision: What problem are people actually paying us to solve? Ask the same kind of question of a website and you get something like What does someone need to understand before this page asks them to click anything? Ask it of a system and the question might become Why does this problem keep coming back after we fix it? At the strategy level, it could be as basic as What are we optimizing for, and should we be?
None of those questions guarantees a brilliant answer. That isn’t the point. They improve the territory being explored. Sometimes the most valuable question is embarrassingly simple: Why are we doing this? That one has the ability to wreck an afternoon with impressive efficiency.
Asking Better Questions Doesn’t Mean Writing Fancy Prompts
There’s a temptation to turn any discussion about questions and AI into prompt engineering. Add more context. Assign a role. Specify the format. Define the audience. Provide examples. Those techniques can absolutely help. But a beautifully engineered prompt can still be built around the wrong question.
The deeper skill is recognizing what you are actually trying to understand. What do I need to decide? What assumption am I making? What information would change my mind? Why was that answer unsatisfying? Is there another problem sitting underneath this one? You don’t need to interrogate every simple request like that. If I ask how long something goes in the oven, I’m unlikely to convene a strategic review afterward. Some questions are transactions, and they should stay transactions. The point is knowing when the first answer deserves another round.
The Question Behind the Question
Generative AI has made answers extraordinarily cheap: not perfect answers, not necessarily correct ones, and certainly not judgment. But plausible, articulate, useful-looking answers can now be produced almost instantly. That changes where some of the difficult work lives. The scarce part may increasingly be deciding what deserves investigation in the first place, seeing the assumption nobody has challenged, recognizing why a technically correct answer still feels wrong and staying with a problem long enough for the real question to emerge.
Most of us first encounter AI because we want it to answer something, and that makes perfect sense. It is exceptionally good at making answers easier to obtain. But one of its more interesting uses may begin a little later in the conversation. You ask something, it answers, and instead of accepting the response or rejecting it, you work with it. You question it. You question yourself. You change the frame. The conversation continues until the problem becomes clearer than it was when you began.
The most useful part of conversational AI may not be that it always has an answer. It may be that it gives you somewhere to keep asking. And sometimes the real breakthrough comes when you finally realize that the question you started with was never the one you needed answered.
