AI output is only as good as its user

I believe AI output is only as good as the person using it. I didn't come up with that idea. I learned it from my mentor, who has been saying it for as long as I have been listening, and somewhere along the way it stopped being his line and became the way I actually work. The best way I can put it is that AI is like a pill that gives you superpowers, except the superpowers are built on the skills you already have. Bring a lot, and it multiplies a lot. Bring very little, and it multiplies that too, only more fluently.

I should say clearly what I don't mean. I don't mean that every bad answer is the user's fault, or that beginners can't learn anything from these tools. They can, and I am proof of that. Models have real limits, some tasks are beyond them however skilled the person asking, and experts still get bad output. Access matters too, since not everyone has the same tools or information. But in my experience, for most of the everyday work people bring to AI, the biggest difference between a poor result and an excellent one is not the model. It is what the person brought to the conversation.

↳I knew it was wrong because I had built the thing myself. No amount of clever prompting would have caught that for me. I could have asked it to check the actual code, but you only think to do that when you already know enough to doubt the answer.

That last one stays with me because it shows exactly where the value sits. The AI was fluent, specific and completely wrong, and the only thing standing between that answer and a bad decision was that I knew my own work. If I hadn't, I would have believed it. That is the uncomfortable part of this idea: you cannot see that an answer is thin unless you know enough to see it, so the people who most need to question the output are often the least equipped to do it.

My mentor describes this in a way I haven't been able to improve on. The first answer you get, he says, puts you in the middle of the pack, close to everyone else who asked the same thing. You can only move beyond that by going deeper into a field than the machine can, to the edge where its knowledge starts to thin out and you begin to notice its gaps. And the proof that you have really reached that edge has to come from outside the machine, from people who practise the craft, from primary sources, or from your own work. I find that framing both humbling and motivating, because it means the tool is not the ceiling. My own understanding is.

So what does a good user actually bring? When I look at what changed in me, the first thing is that I now understand my problems well enough to know what a right answer would even contain. I have also noticed that whenever my request is vague, the answer is vague too, even with the best models. They may ask a question or two, but they cannot fill in what I haven't thought through, so a vague answer usually means I haven't finished thinking rather than that the tool has failed. I give more context than I used to, because the model can't know my situation unless I tell it. I hold a higher standard, so I notice the sentence that sounds right but says nothing. Above all, I can now judge what comes back, and I have come to believe that recognising quality matters as much as asking for it.

None of this is really new, and it isn't only true of AI. I noticed the same thing long before, watching people talk with mentors and senior guests: the quality of the questions you ask largely decides what you get out of the conversation. A good question can take a session somewhere nobody planned. AI just makes that truth faster, and much harder to ignore.

It has also changed how I set the tools up. My mentor warns against building a mirror instead of a coach, an assistant so eager to please that it only tells you what you want to hear, and I have tried to take that seriously. The instructions I give my own AI tools are less about being helpful and more about holding me to a standard. Read the source yourself instead of relying on someone's summary. Tell me when you disagree. Judge the work against what I am actually trying to achieve, not against the rules I happened to write down last month. When the output is generic, I have learned to look at my own input first, because more often than not, that is where the weakness was.

The same holds well beyond design and software. An accountant who knows when the numbers don't reconcile, a lawyer who spots the clause that will cause trouble later, a teacher who can tell whether a student has really understood, and a writer who can hear when a sentence is hollow each bring something the tool cannot supply on its own. In every case, the person who understands the work deeply can push the tool much further than the person who doesn't, and can tell when it has gone wrong. That gap does not shrink as the tools improve. If anything, it becomes easier to see, because AI can make shallow work look capable to people outside a field right up until it meets someone who knows the craft, or real users, or the job of maintaining it.

So if I had one piece of advice, it would be to keep improving the thing the tool amplifies, which is you. Where you already know the work, start with your own attempt and let AI improve it, instead of the other way round. Where the work is new to you, use AI to learn it, and ask it to teach you rather than only do it for you, which is how I learned to build software in the first place. Do some of the work by hand, study why good work is good, and notice when you are accepting something you can't assess, because that is a signal to slow down and learn rather than to speed up. It irks me a little when people use AI to skip the thinking entirely. I understand why, because it is faster, but the aim should be to become so good at what you do that the machine has more to multiply. The people who keep getting better at their craft are the ones these tools will lift furthest.