In movies, when detectives, police officers, or lawyers do extraordinary things, I am in awe. But when a programmer “hacks” a bank by typing some HTML, I can’t help thinking - are you kidding me, aisa thode hi na hota hai!

It makes me wonder - are the detective scenes equally wrong? I can’t tell.

I feel the same way about AI now. When I vibe-code in a language or domain I know nothing about, the output looks brilliant. When I review work in an area I know well, I keep finding flaws.

This is the Dunning-Kruger effect with a new twist. When we know little about a topic, we can’t see how much we don’t know. AI makes this gap harder to see, because it fills it with confident, well-formatted answers.

And then length starts to feel like depth. You read 15 pages of documentation to realize that only a couple of pages are actually useful - but if you were to just glance through it, you wouldn’t know which two.

I see it in RCAs too. Many of them read well and say very little. And I don’t think the problem is AI. It’s how we use it.

If you don’t give AI enough context, it makes “probabilistic” guesses. A few guesses are right, a few sound right, and a few are plainly wrong. If you don’t have the context yourself, you can’t tell which is which.

This is true for a product spec, a marketing plan, a contract summary, or a business strategy. Not only for code.

None of this is a reason to stop using AI. I worked on chatbots in 2015, and I didn’t think we would get here this soon. Today, millions of people, including me, use it every day.

But one simple test helps. Ask AI about something you know very well, and count the mistakes you catch. Then remember that the same mistakes exist in topics where you can’t catch them.

There is still a lot of context in people’s heads that no model has.