In short
- AI does not make you dumber by itself. But people who hand over all the thinking learn less and think less critically. That is what the first studies show.
- The key condition is how you use it. People who ask AI for explanations and counterarguments learn about as well as people who work without it.
- Agree as a team which work people keep doing and checking themselves, and where AI thinks along instead of taking over.
Does AI make you dumber? Not automatically, but it can. When a chatbot does the work, you stop practising that work yourself, and it shows in what people still know and can do afterwards. People who use AI to understand something better are much less affected.
So the difference lies less in the tool than in the agreements about how you use it. For an organisation that is good news, because agreements are something you can make. The Dutch newspaper de Volkskrant covered the topic on 2 November in an episode of its science podcast Ondertussen in de kosmos ("Maken chatbots ons dommer?", in Dutch), with AI editor Laurens Verhagen and psychology and brain editor Kaya Bouma, who also wrote an article about it. You can listen below. In this piece we look at what it means for teams that work with AI every day.
What do the first studies show?
The research is young, the groups are often small and some of it has not yet been peer reviewed. Even so, the early studies point the same way.
Less engaged with your own work. Researchers at the MIT Media Lab asked 54 participants to write essays with ChatGPT, with a search engine or with no tools at all, while measuring brain activity with EEG. The ChatGPT group showed the weakest connectivity, felt the least ownership of their text and struggled to quote from it afterwards (MIT Media Lab). It is a preprint with a small sample, so treat it as a signal rather than proof.
Better while practising, worse without AI. In an experiment with almost a thousand secondary school pupils, students practising maths got either a standard ChatGPT interface or a version that only gave hints. With the standard version their practice grades rose by 48 per cent, but in the exam without AI they scored 17 per cent lower than pupils who never had access. The hint-based version improved practice by 127 per cent and largely removed that drawback (Bastani et al., PNAS).
Less critical when trust is high. Microsoft Research and Carnegie Mellon asked 319 knowledge workers about 936 examples from their own work. The more people trusted AI, the less critical thinking they reported. The more confident they were in their own ability, the more they thought things through. According to the researchers, critical thinking shifts towards checking, combining and steering what AI produces (Microsoft Research). These are participants' own estimates, not measurements.
Does this apply at work too?
Two studies from practice show that this is not just about students.
Anthropic, the company behind Claude, ran a randomised experiment in which 52 developers worked with a programming library they did not know yet. The group with an AI assistant was barely faster, yet scored 50 per cent on the test afterwards against 67 per cent without AI. The biggest gap was in debugging. How people used AI mattered a great deal: those who left everything to it scored below 40 per cent, while those who mainly asked conceptual questions or asked for explanations alongside the code scored 65 per cent or more (Anthropic). This is research by an AI vendor, and those usage patterns rest on small groups.
In healthcare, researchers followed 19 experienced endoscopists in four Polish centres. After AI support for colonoscopy was introduced, their detection rate in procedures without AI fell from 28.4 to 22.4 per cent (The Lancet Gastroenterology & Hepatology). It is an observational study, so other causes cannot be ruled out.
The pattern is consistent: when AI takes over a task, you stop practising it and become worse at it without AI. That matters most for work you are still learning and for work you need to be able to check. A team that can no longer judge AI output can no longer catch its mistakes either. That is why the bottleneck in adopting AI is often people rather than the model.
How do you use AI without getting dumber?
You do not need to use AI less. You do need to decide when it thinks along and when it takes over. A team can adopt these agreements straight away:
- You first, then AI. Write your own outline, hypothesis or draft answer, then ask AI to add to it or challenge it. The thinking stays with you and AI makes it better.
- Ask for the reasoning, not just the answer. "Why does this work?" and "What would go wrong if ...?" teach you more than a ready-made solution. In the Anthropic study, those were the users who scored best.
- Let AI argue with you. Ask for the three weakest points in your plan or the strongest counterargument. It keeps you critical, even when an answer sounds convincing.
- Agree who checks. Critical thinking moves towards reviewing AI output. Make that a fixed step in the process: who approves, what to look for, what to do when in doubt.
- Protect how juniors learn. Decide per role which core skills people keep practising themselves. For learning tasks, set up an assistant that gives hints instead of full solutions, like the hint version in the pupil study. With Claude skills or a custom GPT you record that instruction once.
What the AI Act asks of every organisation in terms of AI knowledge is covered in AI literacy in your organisation. Handling AI output critically is part of that.
When is handing work to AI fine?
Not every skill is worth keeping. A calculator or a navigation app takes over work too, and hardly anyone minds. The same goes for routine work you do not need to learn from: formatting a document, summarising a meeting or producing a first translation that someone reviews.
Be careful with three kinds of work: what you still need to learn, what you need to be able to check, and work where your own judgement is the point, such as advice, strategy or decisions about customers. There it pays to use AI deliberately as a sparring partner.
And keep in mind that the science has only just started. Most studies are small and short. Nobody yet knows what years of AI use will do.
Sources
- de Volkskrant, Maken chatbots ons dommer?, podcast Ondertussen in de kosmos (Dutch), 2 November 2026
- de Volkskrant, Laurens Verhagen and Kaya Bouma, Breinrot door de chatbot? Hoe AI het denkvermogen kan beïnvloeden (Dutch), 29 October 2026
- MIT Media Lab, Kosmyna et al., Your Brain on ChatGPT, preprint, June 2025
- Bastani et al., Generative AI without guardrails can harm learning, PNAS, 2025
- Microsoft Research and Carnegie Mellon, Lee et al., The Impact of Generative AI on Critical Thinking, CHI 2025
- Anthropic, How AI assistance impacts the formation of coding skills, 29 January 2026
- Budzyń et al., Endoscopist deskilling risk after exposure to artificial intelligence in colonoscopy, The Lancet Gastroenterology & Hepatology, 2025
