Is it true that AI makes people dumber?

Research describes a narrower risk: an accurate AI hint improves the solution now and takes away the practice a person needs to learn to manage on their own. So it is better to keep your own skill in use: your own version first, then the hint.

What is cognitive offloading, in simple terms?

It is when a person hands part of their mental work to a tool: a notepad, a calculator, an AI assistant. For things they have already mastered, it saves time. For things they are still learning, it means lost practice.

If I double-check AI, does that mean I don't trust it?

Double-checking is the very route to calibrated trust - trust based on merit. The model in the arXiv paper links trust in an assistant to how often a person checks its reasoning, and the history of those checks to whether they remember the rule. By checking, a person learns.

How can I tell whether a task is at the edge of my ability?

It is still hard on your own but works with a small hint - that is the zone of proximal development. If the task feels easy, it is time to take on something harder; if you cannot get anywhere without a ready-made answer, ask the assistant for one step instead of the whole solution.

Where in Tallinn can I learn to work with AI in Russian?

At the Intellektor school: the programme 'Applying AI in Organisational Work Processes' is taught in Russian, with sessions in a training room in Tallinn or online, based on participants' own work tasks. Enrolment and questions: office@intellektor.ee.

Articles

Does AI make us dumber? What research says and how adults can learn with an assistant

Published 2 November 2026

The spreadsheet is finished and the email to the client is ready, all before the coffee has gone cold: an AI assistant has handled a task a beginner would have spent an evening on. A week later an identical task arrives, and the person realises they still can't do it on their own. This is what the risk of learning with AI looks like: the assistant gets there faster, and an adult getting to grips with new work hands over the practice along with the task. So when asked whether AI makes us dumber, researchers give a more precise answer: an AI hint helps solve the task now and takes away the practice your own skill needs to grow. At Intellektor, a continuing education institution for adults in Tallinn, AI courses are taught in Russian and built around participants' own work tasks, and the approach that preserves your own skill comes down to three techniques: your own draft first, then the hint; double-checking the assistant's reasoning as part of learning; tasks at the edge of your ability instead of ready-made answers.

Does an AI assistant make us dumber?

The short answer: a lot depends on which of the two does the mental work. People have long been used to handing it over to tools, and there is a name for this - cognitive offloading: the term for passing part of your mental work to an external aid, whether a notepad, a calculator, a satnav or an AI assistant.

When someone can already do a task, offloading only helps: the assistant saves time on things they have already mastered. But for an adult who is just starting out in new work, the practice goes along with the task. This risk is described by the VLT theory in a recent paper on arXiv, an open archive of scientific preprints: an accurate AI hint improves the solution here and now, but deprives the person of the practice they need to learn to manage on their own.

That is why a different question is more useful: which of the two is lifting the weight? If a coach lifts the barbell for someone, the record goes in the book, but their muscles are no stronger. It is the same with an AI assistant: the task is done, while the skill the person took on the new work to develop stays exactly where it was.

Why does checking the assistant's answers help you remember the rule?

Because checking is the very practice that a hint otherwise takes away. The same arXiv paper describes a model that links two things: how much a person trusts the assistant and how often they double-check its reasoning. The model then links the history of these checks to whether the person will remember the rule.

This is where a second term comes in useful - calibrated trust: trust in an assistant based on its merits, where a person relies on it calmly where it is usually reliable and checks more often where it tends to make mistakes.

An assistant can be wrong and still sound polished: the tone is confident, the text is tidy, and yet the rule in it is muddled. Blind trust lets that kind of mistake through. That is why double-checking the reasoning is part of learning. It has nothing to do with distrust: the person works out why the assistant reached its conclusion and picks up the rule along the way.

How hard should a task be for it to teach you something?

Right at the edge of your ability. This is known as the zone of proximal development: tasks a person cannot yet manage alone but can manage with a little help.

A good illustration is the ProACT system, which is also discussed in a paper on arXiv. Firefighters train in virtual reality (VR), and the system adapts the difficulty of each exercise to the trainee's current progress, so that the task is neither too hard nor too easy. The reason is clear: live fire drills are expensive, dangerous and hard to scale.

You can stay in the zone of proximal development without a simulator, too. A ready-made answer takes a person out of it: the task is solved and there is no difficulty left. So it is worth rationing hints - asking the assistant for a direction rather than a solution, for just the next step, or for a similar task slightly harder than the last one.

Three techniques: how to learn with an AI assistant and keep your own skill

Together, this adds up to a working routine for any task a person wants to be able to do on their own in time:

Your own draft first, then the hint. Even a rough outline will do: a plan for an email, a first version of a formula, your own version of a reply to a client. The assistant then compares and corrects, and the practice stays with the learner.
Double-checking the reasoning as part of learning. Ask the assistant to explain why it chose that particular solution, and check at least one step: against a document, on a simple example with a known answer, or against common sense. This is how calibrated trust is built and how the rule sticks.
Tasks at the edge of your ability instead of ready-made answers. Ask for a hint covering one step, a leading question, or a task slightly harder than the last one. As soon as the task starts to feel easy, raise the bar.

When this routine falters, and what to do then

The routine is designed for skills a person wants to keep for themselves. In three cases it gets in the way, and each one has a way out:

The deadline is pressing. There is no time for a draft, and a ready-made answer saves the day. What to do: take it, and once the deadline has passed, go through one key step with the assistant and repeat it yourself.
The person does not need the skill. The task is rare, and they have no intention of learning it. What to do: hand it to the assistant without a second thought - here cognitive offloading works in the person's favour and leaves more time for what they are actually learning.
There is nothing to check against. The topic is completely new, and there is no way to verify the assistant's reasoning. What to do: ask it to show its reasoning step by step and check at least one step on an example with a known answer, or show the conclusion to a colleague who knows the subject.

So what should you learn on AI courses for adults?

What the assistant will not do for you: setting the task and signing off on the result. An adult's role next to AI is to brief and to check: they decide what needs doing and with which materials, and check whether the answer is good enough to use in their work. The assistant can be confidently wrong, and the one who will notice is the person who understands the subject themselves.

This is what the Intellektor programme 'Applying AI in Organisational Work Processes' is built on: participants learn to give an AI assistant a task with context, materials and criteria for the result, and to assess the answer against acceptance criteria. At least 70% of contact hours are practice on participants' own work tasks, so the one putting in the work is the learner.

The programme's author and the speaker at the sessions is the school's founder, Andrii Ruban. He has used AI assistants in his day-to-day work since 2024 and is introducing AI tools into the work processes of three organisations; more on the trainers page.

The takeaway: an AI assistant speeds up work and can take away practice, and the 'draft - check - task at the edge' routine gives it back to the person. The next step: on your next work task, sketch your own version first and only then open the assistant. You can follow this path in a group, working through your own tasks, on AI courses for adults in Tallinn - sessions are held in Russian.

Questions and answers

Is it true that AI makes people dumber?: Research describes a narrower risk: an accurate AI hint improves the solution now and takes away the practice a person needs to learn to manage on their own. So it is better to keep your own skill in use: your own version first, then the hint.
What is cognitive offloading, in simple terms?: It is when a person hands part of their mental work to a tool: a notepad, a calculator, an AI assistant. For things they have already mastered, it saves time. For things they are still learning, it means lost practice.
If I double-check AI, does that mean I don't trust it?: Double-checking is the very route to calibrated trust - trust based on merit. The model in the arXiv paper links trust in an assistant to how often a person checks its reasoning, and the history of those checks to whether they remember the rule. By checking, a person learns.
How can I tell whether a task is at the edge of my ability?: It is still hard on your own but works with a small hint - that is the zone of proximal development. If the task feels easy, it is time to take on something harder; if you cannot get anywhere without a ready-made answer, ask the assistant for one step instead of the whole solution.
Where in Tallinn can I learn to work with AI in Russian?: At the Intellektor school: the programme 'Applying AI in Organisational Work Processes' is taught in Russian, with sessions in a training room in Tallinn or online, based on participants' own work tasks. Enrolment and questions: office@intellektor.ee.

← All articles