Is AI Making Us Dumber?

I have conducted roughly one hundred and fifty practical interviews over the past four years. Fifty for data specialist roles. A hundred for advertising and performance marketing specialists.

Almost every one of them involved sitting down with a candidate over a practical task — something close to a real problem we actually need to solve at the company. Not theory. Not trivia. Applied problem-solving.

Over time, I started noticing a pattern.

It was not sudden. There was no single moment where the quality of candidates collapsed. But if you do this long enough — if you keep the same tasks and the same standards — you begin to see the curve.

And the curve is going down.

THE BEST RESULT I EVER SAW — WITHOUT AI

The strongest solution I have ever received came from a woman who had no AI.

No ChatGPT. No Copilot. No language model running in the background during her hour of preparation. She had the data, the task description, and her own mind.

She is a mathematics teacher. She works in education, promoting innovative Montessori-inspired approaches to teaching — the kind of person who spends her days figuring out how to make complex things understandable.

It showed.

I cannot share what she did specifically — the interviews are still ongoing and the tasks are still in use. But on a scale of one to ten, she scored a nine.

There were several moments during her presentation where I thought: this is a different league. Not one moment. Several.

She did not just answer the question. She noticed things in the data that were not part of the question. Things I had not explicitly asked about. Things that mattered.

That is the difference between AI-assisted output and genuine human problem-solving. AI answers what you ask. She answered what mattered.

To this day, no candidate with AI access has come close.

THE DECLINE OF INDEPENDENT THINKING

Three years ago, roughly one in five candidates scored above six out of ten. Their solution was solid enough that the interview became genuinely interesting — a conversation, not an evaluation.

One in twenty scored eight or higher. That meant they solved the task independently, mostly correctly, with a level of critical thinking that made it clear they understood what they were doing.

Today — for data specialist roles — I would struggle to find anyone above five.

I am not talking about a subtle shift. I am talking about a measurable decline in independent problem-solving that has reshaped what I expect when I walk into an interview room.

For advertising specialists, the picture is slightly better. But the trend is the same.

THE AI PRODUCTIVITY PARADOX — MORE TOOLS, WORSE RESULTS

Data specialist candidates receive the dataset and the task one hour before the interview. They have time to prepare. They have access to any tools they choose — including AI models that are orders of magnitude more capable than anything that existed when I started doing these interviews.

More time. Better tools. Worse results.

In any other context, we would call this a crisis.

I am not the only one seeing this pattern. A randomised controlled trial by Grace Liu and colleagues (1,222 participants across multiple experiments) found that people who used AI assistance performed significantly worse once the tool was removed — even after only ten minutes of exposure. The researchers suggest that AI conditions people to expect immediate answers, denying them the experience of working through challenges on their own. The mechanism has a name: cognitive offloading.

An MIT Media Lab study of 54 participants using EEG measurements found that brain connectivity systematically decreased with the level of AI assistance — with the ChatGPT group showing the weakest neural coupling in areas associated with executive function, semantic processing, and attention regulation. The study is a preprint and has not yet been peer-reviewed — but the direction is consistent with what I observe in my interviews.

These are different domains — mathematics, essay writing. My domain is analytical problem-solving in data and marketing. But the underlying pattern is difficult to dismiss. Better tools. Less effort. Weaker independent performance.

Here is what it looks like in practice. The candidate takes the data, feeds it into a model, and brings the output to the interview. Sometimes the output looks clean. Sometimes even impressive — well-structured, visually polished.

Then I ask the first follow-up question.

"Why did you approach it this way?"

And the room goes quiet.

The answer was never theirs.

WHEN AI REPLACES THINKING — THE BLACKBOX GENERATION

The most striking interviews I have experienced recently involve young candidates who are skilled programmers. They know how to fine-tune models, adjust algorithms, work with complex tooling.

They present their solution confidently. The output looks structured, sometimes even impressive. But after a few follow-up questions, the picture becomes clear. They fed the data into a model, treated it as a blackbox, and accepted whatever came out. They do not say it that bluntly — but that is what happened.