Digital Dementia: When AI Removes the Practice Behind Judgement

I have built AI into the way I work.

It helps me research, organise ideas, test arguments and get moving when procrastination begins to sound suspiciously like careful preparation. I build my own AI skills. I am writing a book about how humans can thrive with this technology. I am an enthusiastic user.

AI is also the first technology I have used that can remove almost the entire visible process of thinking. Give it a vague question and it can supply the structure, choose an angle, write the argument, add examples and deliver a polished answer before you have decided what you believe.

That possibility keeps bothering me.

Ruben Hassid recently made a sharp economic observation: AI has made the first serious attempt close to free. A memo, market scan, sales email, slide outline or set of campaign ideas can appear in seconds. When the cost of an attempt collapses, the number of attempts rises. Average output loses value. Judgement becomes more valuable. His piece on cheap intelligence is right about all three.

My concern begins one step earlier.

Judgement grows through bad first attempts, corrections, embarrassing mistakes, difficult conversations and repeated exposure to consequences. What happens when the next generation can skip the work from which judgement is built?

The label overstates the evidence

Manfred Spitzer popularised the phrase “digital dementia” in 2012. The phrase does too much. Researchers use it as a hypothesis about technology and cognition. Clinical diagnostic systems use established dementia categories. The World Health Organization describes dementia as a syndrome caused by diseases that damage the brain and progressively impair memory, thinking and daily functioning.

The broad claim that ordinary technology use causes dementia is unsupported. A 2025 meta-analysis in Nature Human Behaviour examined 136 papers involving adults over 50. The 57 studies suitable for meta-analysis covered 411,430 people. Digital technology use was associated with lower odds of cognitive impairment and lower rates of cognitive decline over time. The authors were careful: these observational findings do not establish which way causation runs.

That evidence matters. Phones and computers can stimulate us, connect us, teach us and compensate for limits in memory. Panic makes for a good headline and bad science.

A narrower concern survives the evidence. Digital tools change which mental tasks we perform ourselves.

Researchers call this cognitive offloading. Evan Risko and Sam Gilbert define it as using an external action or tool to reduce the mental demands of a task. A calendar reminder, contact list, notebook or calculator all qualify. Cognitive offloading is often sensible. Our brains have never stored everything. We have always spread memory across people, paper and tools.

But offloading changes what we practise.

In four experiments published in Science in 2011, Betsy Sparrow, Jenny Liu and Daniel Wegner found that people who expected information to remain available later remembered less of the information itself and more about where to find it. The researchers called the internet a form of external or “transactive” memory. The study found a specific change in what people encoded. General intelligence was outside its scope.

Navigation offers another clue. A 2020 study of 50 regular drivers found that greater lifetime GPS use was associated with worse spatial memory during navigation without GPS. Only 13 participants returned for a three-year follow-up, so grand conclusions would be silly. That small group still showed an association between heavier GPS use and steeper decline in spatial-memory measures.

The useful question is therefore specific: which mental task does the tool perform, and do we still need to learn that task ourselves?

AI can offload the path from question to answer

Google can spare us from remembering a fact. GPS can spare us from building a route. Generative AI can take over a much longer chain: framing the problem, selecting assumptions, finding information, producing options, writing the conclusion and wrapping it in fluent language.

The output may be useful. The learning may be close to zero.

That gap matters because polished work can hide missing competence. You can submit a good-looking analysis without understanding the model behind it. You can publish a confident argument you could not defend in conversation. You can approve a source that does not exist because the summary sounded right and you were already late for lunch.

The research on generative AI and cognition is still young. It gives us signals rather than a final verdict.

In a study presented at CHI 2025, researchers from Microsoft and Carnegie Mellon surveyed 319 knowledge workers about 936 real examples of AI-assisted work. Higher confidence in generative AI was associated with less self-reported critical-thinking effort. Higher confidence in one’s own ability was associated with more. AI also shifted critical thinking towards verification, integration and responsibility for the task. The study relied on self-reports, so it cannot prove that AI weakened anyone’s brain. It does show where the work moves when AI enters the process.

A large field experiment in high-school mathematics gives us firmer evidence about learning. Nearly 1,000 students received access either to a standard GPT-4 interface, a version designed to behave more like a tutor, or no AI. The standard interface improved performance during practice by 48 per cent. When the AI was removed, those students scored 17 per cent lower than the control group. The guarded tutor produced larger gains during practice and largely avoided the later loss. The paper was published in PNAS.

The two AI systems used the same underlying technology. The design of the interaction determined whether students practised the skill or obtained the answer.

Older learning research helps explain this. In a well-known 2011 experiment, students who practised retrieving and reconstructing knowledge learned more than students who used repeated study or concept mapping. The advantage also appeared on questions that required comprehension and inference. Retrieval practice works because the act of bringing knowledge back is part of the learning.

AI removes friction. We need to distinguish wasted friction from the effort that trains memory, reasoning and judgement.

We are creating an apprenticeship problem

Much of today’s professional judgement was built while producing work was expensive. People wrote the weak memo, made the poor forecast, watched the client lose interest, tried again and slowly learned what good looked like.

Young professionals now enter a workplace where the first pass can arrive in seconds. We tell them that their future value lies in judgement, then automate many of the repetitions that produce it.

That is a poor training system.

Imagine a junior analyst generating 50 market scenarios before lunch. The volume looks impressive. Without a mental model of the market, the analyst cannot see that 38 scenarios share the same weak assumption. The senior can spot it, for now. Ten years later, who becomes the senior?

AI can improve apprenticeship when we make it ask questions, reveal assumptions, offer counterexamples and explain errors. The guarded maths tutor points in that direction. Designs that hand over answers can create dependency. Designs that demand participation can accelerate learning.

The workflow strongly influences which one we get.

Five guardrails for keeping your mind in the work

I think we need practical rules, at school and at work. Banning AI will fail. Unrestricted use rewards the output while ignoring the capacity of the person producing it.

1. Write before you prompt. For an important task, spend five minutes stating the problem, your tentative answer and what you are uncertain about. Give the model something to challenge.

2. Ask for resistance. Request the weakest assumption, missing evidence, strongest counterargument and conditions under which your view would fail. Agreement feels pleasant. Critique teaches more.

3. Use tutor mode when you are learning. Ask for one question or hint at a time. Make yourself complete the next step. A finished answer is useful when you need an answer. It is a poor substitute when you need the skill.

4. Close the chat and reconstruct. Explain the result from memory. Write the final reasoning in your own words. Open the original studies, spreadsheet or source material and check the claims. Fluency is an invitation to verify, never proof.

5. Keep unaided repetitions. If a skill matters to your work, practise it without AI often enough to know whether it is still yours. Write the first page. Build the basic model. Navigate the familiar route. Teach the idea to another person without reopening the chat.

Managers need guardrails too. Measuring only volume rewards cognitive dependency. Review the reasoning behind the work. Ask what the person rejected, which evidence they checked and what decision they would defend if the AI were wrong.

Keep the useful friction

I will keep using AI. Formatting, transcription, routine searches, boilerplate and administrative work deserve to become easier. AI creates momentum and opens paths I would struggle to see alone. It can give us more time for curiosity, relationships, judgement and better questions.

Some friction is where learning happens. The first few minutes of a difficult problem force you to retrieve what you know, expose what you do not know and form a view that can be tested. Removing that moment saves time today and may reduce your ability tomorrow.

Before your next important prompt, write your own answer in three sentences. Then ask AI to find what you missed.