What changes when we start relying on AI regularly?

Now, what changes now. Not in some distant science-fiction future. In the work we do, the skills we keep, the things we stop practicing, and how well we can tell when AI gets something wrong.

If AI does something for you, do you slowly lose the ability to do it yourself?

A new paper called Adaptive Complementarity in Human-AI Systems looks at something we rarely consider when evaluating AI.

Usually, we ask: Did it help me do this task better today?

The researchers argue that we should also ask what the arrangement does to us over time. If AI repeatedly takes over part of a task, we may stop building the experience, habits, or relationships that made us capable of doing it ourselves.

The research does not show that AI is making people less intelligent. It is a theoretical model.

But the question is useful anyway.

Imagine you use AI to write every difficult email. At first, that's simply convenient. Six months later, are you still as comfortable handling a delicate conversation without it?

The question isn't whether to use AI.

It's: What do I still want to be good at?

Having a human “check the AI” may not mean very much

You hear this advice constantly: Keep a human in the loop.

That sounds reassuring. But what if the human isn't actually making a meaningful decision? A revised paper called Cognitive Amplification vs Cognitive Delegation explores this through computer simulations.

In the model, keeping human ability intact wasn't enough by itself. The human contribution had to actually affect the final result. That distinction feels surprisingly relevant to everyday AI use.

There is a big difference between:

AI gives you information and you make the decision.

And:

AI produces a polished answer and you click Approve.

In both cases, technically, a human was involved.

But only one required much judgment.

So the more useful question may be: Does my review actually change anything? If not, we may be calling something human oversight when it is really just human confirmation.

AI can finish the job before you know whether the job was done well

Researchers behind LLMVul examined more than 21,000 pieces of C/C++ computer code associated with AI-assisted development. Their analysis identified vulnerabilities in 1,540 of them.

That doesn't prove AI-written code is generally worse than human-written code. But it highlights a problem that applies to far more than programming. AI can produce something that looks finished.

A report. A spreadsheet. A legal explanation. A piece of code. A travel plan. A medical summary. The finished-looking answer can arrive long before we have developed the expertise needed to know whether it is actually good.

That creates an odd new situation:

Producing an answer is becoming easier. Knowing whether the answer deserves to be trusted may become more important.

It might also be worth asking: What is this making easier? What am I still learning? Where does my judgment matter? And what do I want to remain capable of doing myself?

AI can give us more time and less friction.

What we do with that is still ours to decide.

Want a starting point? Take the attention quiz.

Stay curious, A. Karacay