I use AI tools every day. I also don't trust them. Both are true at the same time, and that's the only sane way to use them right now.
The problem isn't that AI gives wrong answers. It's that it gives wrong answers in the same confident tone as right ones — no hesitation, no "I'm not sure about this part." A wrong answer and a right answer look identical until you test them. So I stopped judging answers by how they sound, and started running every important one through three questions first.
They're not complicated. If the test takes longer than the answer saved you, you've lost the plot.
1. Can I check this against something I already know is true?
This is the first question because it's the fastest. If the AI tells me something, I ask myself: is there a fact I already trust that this has to agree with?
A real example: I asked an AI to explain why a Python snippet printed what it printed. It gave a confident explanation about loop variable scoping. Sounded fine — except I already knew one solid fact: loop variables leak out of the loop, they don't get a fresh scope per iteration. The explanation contradicted it. I caught it in ten seconds, not because I knew the right answer, but because I knew one true thing it had to fit with.
You don't need to know everything. You need one anchor: a fact from docs you've actually read, a behaviour you've seen with your own eyes. If I have no anchor at all — the topic is completely new — I don't trust the answer yet. I go and find an anchor first, from a source that isn't the AI.
If you can't check an answer against anything you know, you haven't learned anything from it. You've just rented a sentence.
2. Does it fail the same way twice?
Ask the same question twice and you can get two different answers, sometimes opposite ones. Most people treat that as a bug. I treat it as a test.
When something matters — code I'm about to ship, a concept I'm about to teach — I ask twice, in slightly different words. If both answers land in the same place with the same reasoning, my confidence goes up. Not to certain, just up. If they disagree, I know I'm looking at a guess wearing a confident face, and I go check it properly.
A real example: I asked an AI to write a small function for parsing dates in different formats. First attempt, one approach. Second attempt, a different library and a different approach entirely, no mention of the first. Both looked plausible. But the fact that it couldn't settle told me the problem was harder than it looked, and I ended up reading the library docs myself — where I found an edge case both versions had missed.
Two different answers doesn't always mean both are wrong — sometimes the question has more than one good answer. But an answer that can't survive being asked twice is not one I'd build on.
3. Would I bet my own code review on it?
This is the one that changes my behaviour, because it makes the stakes personal.
Imagine you're in a code review. A senior developer points at a line and asks, "why did you do it this way?" If your honest answer is "the AI told me to," you don't understand it well enough to ship it.
A real example: an AI suggested a clever one-liner to replace a loop I'd written. It worked, the tests passed, it looked smart. But when I imagined explaining it in a review — why this approach, what happens with an empty input, what breaks if the data changes shape — I couldn't. The one-liner did three things at once and I understood maybe one of them. I kept my boring loop. Longer and dumber, and I could defend every line of it.
This isn't about being anti-AI. The clever answer is often genuinely better. But "often right" isn't the standard for code you ship. The standard is: can you explain it, defend it, and fix it at 2am when it breaks? If not, the AI didn't save you time. It just moved the work to later, when it'll cost more.
The short version
Check it against something true. Ask it twice. Imagine defending it in front of someone who knows more than you.
None of this requires distrusting AI. It just stops you from outsourcing the part that was always yours: knowing whether something is right. The AI can write the answer. It can't take responsibility for it.
Top comments (0)