When AI Gets Smarter – and We Unlearn How to Think

Some headlines are so stark that they are hard to ignore.

“We are leading ourselves straight into a dictatorship of idiots.”

That is how the Austrian daily newspaper Heute summarises a warning by French neuroscientist Michel Desmurget.

His thesis: educational performance is declining, people are spending ever more time in front of screens, and at the same time artificial intelligence systems are becoming more powerful. If we delegate more and more mental work to machines, we may eventually no longer be able to adequately assess whether what an AI presents to us is correct, false, manipulative or simply fabricated [1].

It sounds spectacular.

Perhaps even alarmist.

And that is precisely why I wanted to know:

How much science is actually behind this warning?

Is this merely another culturally pessimistic story claiming that every new generation is supposedly becoming “stupider”?

Or are we in fact witnessing a development in which an increasingly powerful technology meets people whose fundamental abilities to read, understand, verify and think independently are simultaneously weakening?

After my research, the answer is:

Desmurget’s stark prediction of the future has not been scientifically proven. But the problem he draws attention to is very real.

And some of the latest data are indeed remarkable.

Who is Michel Desmurget?

Desmurget is not a social-media commentator warning about digitalisation based on a gut feeling.

He is a neuroscientist and research director at the French Institut des Sciences Cognitives Marc Jeannerod, a joint research institution of France’s CNRS and Université Claude Bernard Lyon 1.

There, he leads the research team “Neural and Cognitive Control of Action”. Among other things, it examines how voluntary actions arise, how sensory and motor functions are organised, and how brain functions reorganise after injuries [2].

That matters.

But precision is worthwhile here as well.

Desmurget’s institutional research does not focus primarily on screen time, education and artificial intelligence.

His publicly known statements on these issues are based to a considerable extent on his review and interpretation of a broader scientific literature.

That does not make his arguments worthless.

It merely means:

Even the statements of a renowned neuroscientist must be checked against the underlying data.

“Dictatorship of idiots” is not a scientific diagnosis

The Heute article draws on an interview published in the French newspaper Le Figaro on 12 September 2026.

In it, Desmurget essentially warns that the collapse of school performance, combined with the rise of artificial intelligence, could lead to a “dictatorship of idiots” [3].

The stark phrase does indeed come from this interview context.

But one thing must be clear:

“Dictatorship of idiots” is not a scientific term.

There is no recognised model from which one could calculate:

declining PISA performance + artificial intelligence = dictatorship.

Desmurget is using a polemical warning metaphor.

The interesting question, then, is not whether such a “dictatorship” can be scientifically predicted.

It is:

Are the mechanisms he warns about actually plausible?

And this is where things become considerably more interesting.

Because the new PISA data give cause for concern

Just a few days before the interview, the OECD published its PISA 2025 results.

Around 760,000 15-year-olds from 91 countries and economies took part.

And the results are striking.

Across the OECD on average, mathematics and reading reached their lowest measured levels to date.

Between 2015 and 2025, average mathematics performance fell by 22 PISA points.

In reading, the decline was as much as 28 points [4].

The drop in reading is especially relevant.

Because in a modern information society, reading does not simply mean being able to decipher words.

It means:

Understanding information.

Following arguments.

Comparing sources.

Recognising connections.

Distinguishing facts from opinion.

And noticing contradictions.

The OECD also observes that the share of so-called “hasty readers” – students who work through texts very quickly and give incorrect answers – nearly doubled between 2018 and 2025 [5].

And this is exactly where education meets artificial intelligence.

An AI can provide a perfect answer – without you having learned anything

This may be the decisive distinction in the entire debate.

An AI can solve a mathematical problem.

Summarise a scientific text.

Create a presentation.

Write an essay.

Code.

Translate.

Argue.

Research.

And yet, at first, the quality of the result says surprisingly little about what the person themselves can do afterwards.

Research is now examining precisely this distinction between performance and actual learning.

One of the most interesting studies on the subject appeared in the Proceedings of the National Academy of Sciences in 2025.

Nearly 1,000 students were given GPT-4

Hamsa Bastani and colleagues studied nearly 1,000 students in mathematics classes.

One group worked without AI.

A second was given access to a GPT-4-based assistant.

A third also used GPT-4, but with educational safeguards. This version was designed, for example, to provide hints rather than simply outputting complete solutions.

While the teenagers were allowed to work with AI, their performance improved substantially.

At first, that sounds like a clear success.

Then the AI was removed.

And suddenly, a very different picture emerged.

The students who had previously used the largely unrestricted GPT assistant scored 17 percent worse in the subsequent test than the group that had not used AI at all [6].

With the pedagogically limited AI, this negative effect largely disappeared.

That is a remarkable finding.

AI can improve your immediate performance while impairing your actual learning.

If it takes over the very mental work through which learning arises in the first place.

Cognitive offloading – when we outsource thinking

The OECD addresses precisely this issue in its Digital Education Outlook 2026.

Generative AI could significantly support learning.

But only if its use is designed accordingly.

General-purpose AI systems, by contrast, could lead students to complete tasks successfully without developing the underlying capabilities.

The OECD speaks of cognitive offloading:

outsourcing mental work.

And it warns of possible “metacognitive laziness” – a mental passivity that can emerge when machines constantly take over essential steps in thinking [7].

This is not a warning against artificial intelligence as such.

It is a warning against a particular way of using it.

Because there is also research showing the exact opposite

Anyone who now concludes:

“AI makes people stupid”

is making the same mistake in the other direction.

A randomised study at Harvard University showed in 2025, for example, that a specially developed AI tutor could produce substantial learning benefits compared with conventional active instruction.

194 students took part in the experiment.

The crucial difference:

The AI was deliberately aligned with findings from learning psychology and pedagogy.

It guided students through the learning process.

It did not simply provide answers [8].

This is precisely why the simple question:

“Is AI good or bad for education?”

is actually the wrong one.

The better question is:

Which cognitive work does the AI take over – and which must the human still do themselves?

AI as scaffolding or as a crutch?

Learning is demanding.

That is not a design flaw in our brains.

The effort is part of it.

You may need to read a difficult passage several times.

Try out a mathematical problem.

Make mistakes.

Remember.

Connect concepts.

Develop hypotheses.

That is exactly how knowledge changes.

If an AI permanently removes this uncomfortable part, a paradoxical outcome can arise:

You work faster.

Your text looks better.

Your solution is correct.

But your own knowledge hardly grows.

The OECD refers to this necessary mental engagement as productive cognitive struggle [7].

And suddenly, part of Desmurget’s warning no longer sounds so exaggerated.

The better AI gets, the more invisible the problem becomes

A poor AI is relatively easy to monitor.

If a chatbot produces obvious nonsense, you may immediately notice:

Something is wrong here.

But modern systems are becoming increasingly persuasive.

Their language is improving.

Their arguments are becoming more coherent.

Errors are becoming less frequent.

And this is precisely what increases the temptation to simply adopt their results.

This creates an interesting paradox:

The more capable the machine becomes, the less you may notice how dependent on it you have become.

As long as it works.

It becomes problematic when it is wrong.

When information is missing.

When data have been manipulated.

Or when someone is deliberately trying to influence you.

How do you monitor an AI if you lack the knowledge yourself?

Imagine an AI gives you a perfectly plausible medical explanation.

Technical terms.

Statistics.

Studies.

A logical argument.

How do you know whether it is correct?

You can check sources.

But to do so, you need to understand which sources are relevant.

You can examine the argument.

But for that, you need foundational knowledge.

You can spot contradictions.

But only if you know what might be contradictory in the first place.

Critical thinking does not work in a vacuum.

It needs knowledge.

Language.

Concepts.

Points of comparison.

This is why the idea that factual knowledge becomes unimportant in the age of Google and AI is problematic.

If machines provide ever more information, people may not need less knowledge.

They need more judgement.

There have long been warning signs when it comes to this judgement

The Heute article refers to research by Stanford University.

This research does indeed exist.

In 2016, the Stanford History Education Group tested 7,804 school and university students of different ages.

The results were sobering.

Many participants struggled to distinguish advertising from editorial content.

They did not reliably recognise political interests behind information.

And they had trouble assessing the credibility of digital sources [9].

A later large-scale study focused specifically on high-school students.

3,446 teenagers – 96 percent missed a crucial connection

Between 2018 and 2019, Stanford researchers studied 3,446 students from 16 school districts in 14 US states.

In one task, for example, they were asked to assess the credibility of a website dealing with climate research.

A brief external search would have shown that the organisation behind it had ties to the fossil-fuel industry.

96 percent of the teenagers did not discover this connection.

In another task, around two-thirds could not reliably distinguish editorial content from advertising on a website [10].

That is socially relevant.

But the reason is more complicated than “young people know nothing”

The Heute article strongly links such findings to a lack of general knowledge and linguistic deficits.

That certainly plays a role.

But the Stanford work shows something else as well.

Many students tried to assess a website’s credibility from within the website itself.

Professional fact-checkers did something entirely different.

They left the website.

Opened additional tabs.

Looked for information about the operator, funding and reputation.

This method is called lateral reading [12].

The problem, therefore, is not only:

“People know too little.”

It is also:

Many people have never learned how to verify digital information professionally.

And that can be taught.

Media literacy can be learned

Stanford researchers later examined whether such skills can be taught deliberately.

The answer is:

Yes.

Even relatively short interventions on so-called Civic Online Reasoning improved the ability to verify digital information [13].

Other experiments also show that media-literacy interventions can help people distinguish reliable from unreliable news more effectively [14].

That matters.

Because it means:

We are not at the mercy of inevitable collective dumbing-down.

Education can make a difference.

But permanent mistrust is not enough either

An interesting study in Nature Human Behaviour reveals another side of the problem.

Several experiments tested interventions against disinformation.

They did indeed reduce belief in false information.

But in some cases, an unwanted side effect emerged at the same time:

People also became more sceptical of true information [15].

That is remarkable.

Because democracy does not merely need sceptical people.

It needs people who can distinguish:

When scepticism is warranted – and when trust is justified.

Constant mistrust is not media literacy.

The spectacular China comparison needs a footnote

Among other things, Desmurget compares top mathematics performance in France with China.

The Heute article essentially says:

55 percent of students in China achieved outstanding mathematics results.

In France, it was five percent.

The order of magnitude is almost correct.

But the comparison requires an important qualification.

PISA 2025 does indeed identify five percent of 15-year-olds in France as top performers in mathematics [16].

For B-S-J-Z (China), the figure is 54 percent [17].

But B-S-J-Z stands for:

Beijing.

Shanghai.

Jiangsu.

Zhejiang.

This is not a representative sample for the whole People’s Republic of China.

The performance gap remains impressive.

But particularly in an article about information literacy, impressive figures should also be properly contextualised.

And Austria?

In PISA 2025, Austria remains above the OECD average in mathematics, reading and science.

That is the good news.

The less good news:

Performance in mathematics and reading has fallen again compared with 2022.

Both areas are now below all Austrian PISA results from the period before 2022 [18].

In mathematics, 72 percent of Austrian teenagers reach at least proficiency Level 2.

Eight percent are top performers.

In reading, 72 percent also reach at least Level 2.

Only five percent belong to the highest-performing group.

Austria is therefore by no means in an educational collapse.

But the long-term trend deserves attention.

It becomes particularly interesting when it comes to digital use

According to PISA, Austrian students spend an average of around 1.5 hours per school day using digital devices for learning purposes.

This is joined by an average of around 0.9 hours of digital leisure use during school time.

Twenty-eight percent report that classmates are distracted by digital devices in most or all science lessons [18].

This group achieves lower performance on average.

That does not prove a direct cause.

For example, teenagers with weaker performance may also be distracted more often.

But the relationship is relevant enough to take seriously.

And AI has long since arrived in Austrian classrooms

Perhaps one of the most interesting figures in the new PISA data:

50 percent of Austrian 15-year-olds use AI chatbots for learning at least once a week.

Thirty-seven percent use AI regularly for initial research.

Thirty-four percent have texts summarised that they were actually supposed to read themselves.

Thirty-one percent use AI when drafting written assignments [18].

So we are not talking about a technology that may one day arrive in the education system.

It is already there.

That is why a simple AI ban would be the wrong answer

Austria is currently pursuing two strategies at once, interestingly enough.

Since May 2025, students up to and including Year 8 have generally not been permitted to use mobile phones and comparable devices during school operations – unless they are expressly allowed for school purposes [19].

At the same time, Austrian schools continue to be systematically equipped with tablets and notebooks [20].

And since 2026, AI literacy is to be integrated more deliberately into the education system [21].

That is not a contradiction.

At least not if we distinguish between two things:

Using technology.

and

Delegating mental work entirely to technology.

Does screen time actually make children less intelligent?

Desmurget is known for very clear statements about screen use.

But the scientific evidence is more nuanced precisely in this area.

A major systematic review in JAMA Pediatrics analysed 58 studies involving almost 480,000 children and adolescents in total.

The surprising finding:

Overall screen time was not fundamentally associated with poorer academic performance.

Certain forms of use, however, showed negative associations – for example, heavy television consumption or certain forms of video gaming [22].

Research on young children’s language development also shows:

High-quality educational content can certainly have positive effects [23].

The term “screen time” is therefore often far too broad in scientific terms.

One hour of screen time is not the same as another

One hour of TikTok.

One hour of mathematics class.

One hour of video calling with grandparents.

One hour of research.

One hour of programming.

One hour of computer gaming.

Technically, all of it is screen time.

But cognitively, something entirely different is happening in each case.

This is why the blanket statement:

“Screens make children stupid”

does not do justice to the research.

A more precise statement is much better supported:

Digital distraction, excessive leisure use and replacing one’s own mental work with automated processes can impair learning.

That is less spectacular.

But much more precise.

Behind many platforms lies an attention economy

This brings us to another spectacular figure from the Heute article.

It gives the impression that France could lose around three percent of its gross domestic product because of declining educational performance.

The underlying source says something else.

In 2025, France’s Directorate General of the Treasury published an analysis of the so-called attention economy.

This refers to digital business models that monetise human attention.

The analysis attempts to quantify various social follow-on costs:

Productivity losses.

Health effects.

Loss of time.

And possible long-term effects of impaired cognitive abilities.

Over the long term, the French finance ministry arrives at an estimate of two to three percentage points of GDP.

However, the authority explicitly points out that this estimate rests on numerous assumptions and must be interpreted with caution [24].

That does not mean: “Poorer education costs France three percent of GDP”

This distinction matters.

The French model calculation attempts to estimate the long-term effects of an entire digital attention economy.

Education and possible changes in cognitive abilities play a role in it.

But turning that into a precise forecast that:

“France will lose three percent of GDP because of poorer schools”

would not be justified.

At this point, the Heute article condenses a complex economic model calculation more strongly than the primary source allows.

Educational quality is nevertheless economically relevant

The fundamental relationship between education and long-term economic development has been extensively studied.

Eric Hanushek and Ludger Woessmann, for example, modelled the economic effects of possible improvements in European PISA performance.

Over very long periods, such models produce substantial macroeconomic effects [25].

Here too, the following applies:

These are model calculations.

Not laws of nature.

But the idea that educational quality has nothing to do with future productivity and prosperity would be just as wrong.

And now for the big question: Is our democracy at risk?

Desmurget’s “dictatorship of idiots” is obviously a polemical phrase.

Yet it contains a serious question:

What happens to a democracy when fewer and fewer people can understand and assess complex information?

Democracy does not need a society of scientists.

But it does need citizens who can, at least in principle:

Contextualise information,

Compare claims,

Understand arguments,

Assess sources,

Recognise contradictions

and weigh political alternatives against one another.

A systematic review published in 2026 analysed 359 empirical studies on the relationship between education and democratic competencies.

Overall, it found a predominantly positive relationship between certain forms of education and democratic knowledge, political skills, attitudes and participation [26].

Education is therefore genuinely relevant to democracy.

But poor PISA scores do not create a dictatorship

This must also be said clearly.

Democracies do not automatically collapse when reading literacy declines.

Political institutions.

The rule of law.

Media pluralism.

Economic conditions.

Social inequality.

Corruption.

Political parties.

Political culture.

International developments.

Many factors determine the stability of democratic systems.

Desmurget’s phrase is therefore not a scientific forecast.

It is a warning.

Yet a lack of information literacy does change democratic societies

When people struggle to assess sources, distinguish advertising from information and understand complex arguments, that changes the information environment of a democracy.

Generative AI can further exacerbate this problem.

Because disinformation once had to be produced.

Texts had to be written.

Images edited.

Videos cut.

Translations created.

Today, such content can be generated automatically, personalised and produced at scale.

The cheaper convincing information becomes, the more valuable the ability to verify it becomes.

We are experiencing something historically new

Never before has the average person had access to so much knowledge.

And it has probably never been so easy to appear as though one had this knowledge oneself.

An AI can produce a persuasive text about quantum physics within seconds.

About medicine.

Law.

Politics.

Economics.

History.

You can copy and publish that text.

And perhaps understand very little about the subject yourself.

That is new.

We are beginning to separate two things from one another:

the quality of a visible result

and

the actual competence of the person presenting that result.

Perhaps this is precisely the strongest core of Desmurget’s warning.

In the past, you often had to be able to do something to appear competent

Not always.

But at least more often.

A good scientific text normally required a certain degree of scientific understanding.

A functioning program required programming knowledge.

A good translation required language skills.

A professional graphic required design expertise.

Generative AI changes this relationship.

People can suddenly produce results whose quality far exceeds their own competence.

That is fantastic.

And problematic at the same time.

Fantastic because capabilities are being democratised.

Problematic because it becomes harder to recognise who actually understands what they publish.

And perhaps one day we will even deceive ourselves

The problem is not only that other people could deceive us.

We can deceive ourselves.

If an AI writes a good text for me, its quality may eventually feel like my own competence.

If it corrects my mistake before I notice it myself, I may learn less from it.

If it provides every argument for me, I practise building one myself less.

If it summarises every difficult text, I practise actually understanding difficult texts less.

Not necessarily.

But the risk exists.

The experiment with mathematics students demonstrates precisely this mechanism under controlled conditions [6].

The problem, therefore, is not artificial intelligence

The research to date does not justify a nostalgic call for:

Computers out of schools.

Nor does it justify:

Ban AI.

High-quality digital offerings can improve learning.

Pedagogically designed AI systems can support learning processes.

Austria is therefore attempting to combine AI literacy with critical thinking, data literacy and reflection [21].

The scientifically more interesting dividing line is not between:

analogue and digital.

Human and machine.

Book and screen.

But between:

active mental work

and

complete mental delegation.

It is precisely because of AI that we need foundational knowledge

Why should a child still learn facts if an AI can look up everything?

Because you can only verify information if you have comparative knowledge.

Why still learn to write if an AI writes texts?

Because language structures thought.

Why learn mathematics if computers can calculate?

Because mathematical understanding is needed to interpret results.

Why read long texts if AI can summarise them?

Because complex arguments cannot be fully replaced by summaries.

The existence of the calculator did not make mathematical understanding obsolete.

GPS did not make spatial orientation meaningless.

And artificial intelligence does not automatically make human thinking dispensable.

Perhaps knowledge is becoming even more important

Information used to be scarce.

Today, it is nearly unlimited.

The problem has shifted.

No longer:

“Where do I find an answer?”

But:

“Which answer is correct?”

And perhaps soon:

“Which parts of this perfectly worded answer are wrong?”

That requires judgement.

And judgement does not arise simply from the instruction:

“Think critically.”

It arises from:

Knowledge.

Experience.

Comparison.

Language.

Logic.

And practice.

What Desmurget gets right, in my view

After reviewing the current research, I consider several basic assumptions of his warning to be well founded.

Educational performance, especially in reading and mathematics, has declined significantly across the OECD on average [4].

Digital distraction in class is widespread and associated with weaker performance [18].

Uncontrolled AI use can improve immediate performance while impairing actual learning [6][7].

Many teenagers have difficulty correctly assessing digital sources and underlying interests [9][10].

AI systems must be monitored and their statements verified [7].

And education is demonstrably linked to democratic competencies [26].

So the core of the problem is by no means invented.

Where Desmurget goes too far

Several of his public exaggerations, however, go beyond what the research permits.

There is no empirical basis for the prediction that Europe is inevitably heading towards a “dictatorship of idiots”.

Screen time is not a uniform harmful quantity. Content, purpose, age, duration and context all play a significant role [22][23].

AI does not inevitably lead to poorer learning. Good AI tutors can produce the opposite effect [8].

The often-cited 55 percent of Chinese top mathematics performers refers to four particularly high-performing regions, not all of China [17].

And the three percent GDP loss comes from an uncertain long-term model of the digital attention economy, not from a direct forecast of a collapse in French education [24].

Precisely when we call for critical thinking, alarmist claims must be critically examined as well.

And what remains of the Heute article?

The original article is not fundamentally wrong.

But it condenses several complex relationships very strongly.

The Stanford research is abbreviated.

The Chinese PISA data are geographically simplified.

The French three-percent estimate appears more definitive than the primary source presents it.

And a scientifically justifiable concern about education and information literacy becomes, through the headline, almost a political forecast of the future.

That is journalistically understandable.

But that is exactly why such a story needs a second look.

Perhaps the truth is less spectacular – and more serious at the same time

We probably do not need to fear an imminent “dictatorship of idiots”.

But we are indeed in the midst of a social experiment.

For the first time, a generation is growing up with tools that can take over almost every mental task, at least in part.

Researching.

Writing.

Calculating.

Coding.

Translating.

Arguing.

Summarising.

Designing.

And no one can say with confidence today what the long-term effects will be if these abilities are regularly outsourced while they are still developing.

That should not create panic.

But attention.

Austria is already in the middle of this experiment

Fifty percent of Austrian 15-year-olds use AI for learning at least weekly.

Thirty-one percent use it to draft written assignments.

At the same time, performance in reading and mathematics is declining.

That does not prove a causal relationship.

PISA itself points out that the relationships between digital use, AI and performance are complex [18].

But these figures show one thing very clearly:

The question of AI and education does not belong on the political agenda at some future point.

It is already there.

Perhaps the most important question is not: What can AI do?

We are currently fascinated by everything artificial intelligence can do.

That is understandable.

Each new generation of these systems produces more impressive results.

But in the long term, another question may be far more important:

What must we humans continue to be able to do ourselves?

Which skills may we outsource?

Which must we consciously train?

When does AI support our thinking?

When does it replace it?

What knowledge do we need in order to judge machines?

And at what point does convenience become dependence?

There are no final answers to these questions yet.

But research is beginning to make initial boundaries visible.

Conclusion: The intelligent machine is not the real problem

Desmurget’s phrase about a “dictatorship of idiots” is too sensationalist for me to adopt as a scientific diagnosis.

But after reviewing the studies, the current PISA results and research on generative AI, I also think it would be wrong to dismiss his warning simply as anti-technology alarmism.

The problem exists.

Perhaps we are simply framing it incorrectly.

The greatest danger may not be that artificial intelligence eventually becomes too intelligent.

It may lie in confusing its capabilities with our own competence.

In being able to produce perfect answers without truly understanding the questions.

In possessing information without building knowledge.

In producing texts without being able to read complex texts.

In spreading arguments without being able to assess them.

And in eventually having to trust a machine because we lack precisely the skills needed to monitor it.

That would not be a “dictatorship of idiots” in the scientific sense.

It would be something less spectacular – and perhaps precisely for that reason more dangerous:

a society with access to more knowledge than any generation before it, while increasingly struggling to decide for itself which of it it can trust.

The good news is:

This path is not inevitable.

Media literacy can be trained.

Reading literacy can be strengthened.

AI can support learning.

Digital tools can make knowledge more accessible.

And well-developed AI tutors can demonstrably help.

The decisive dividing line, therefore, is not between human and machine.

Not between book and tablet.

Not between analogue and digital.

It runs between:

technology that expands our thinking – and technology to which we surrender our thinking.

The more capable artificial intelligence becomes, the more important may become precisely the ability some believed technology would one day make obsolete:

thinking for ourselves.

RECHERCHE