Will AI make us (more) stupid?

Spiral Staircase — Photo by Gianni Salinetti

This blog post is about the potential impacts of an unaware usage of AI, especially assistants, on the cognitive behaviors. The idea of writing this article started when I found a paper on Arxiv titled AI Assistance Reduces Persistence and Hurts Independent Performance.

The Paper

The abstract of the paper begins with the following introduction, followed by an open-ended question:

“People often optimize for long-term goals in collaboration: A mentor or companion doesn’t just answer questions, but also scaffolds learning, tracks progress, and prioritizes the other person’s growth over immediate results. In contrast, current AI systems are fundamentally short-sighted collaborators – optimized for providing instant and complete responses, without ever saying no (unless for safety reasons). What are the consequences of this dynamic?”

The paper tries to demonstrate that people who develop a habit of constantly using AI assistants for everyday tasks are prone to perform worse without AI and also prone to give up more easily. To support this thesis, the researchers conducted 3 different experiments:

Experiment 1: AI impairs unassisted performance and persistence. Here a set of 354 volunteer subjects was asked to perform fraction-solving tasks. A subset of participants (N=191) were assigned to the AI condition, which means they had free access to an AI assistant during the test, while the remainder (N=163) was assigned to the control condition. The AI condition group was then asked to solve other fractions with no AI assistant. As a result, when the AI was removed, their performance dropped and they were also less likely to persist with problems.
Experiment 2: Replicating the results and ruling out confounds. The second experiment replicated the first experiment with a larger group and some improvements, like adding a pretest of easier fraction problems to exclude poorly performing subjects to address general skill-level gaps, and by adding a sidebar displaying pretest solutions to the control group to eliminate interface asymmetry between groups. Once again, performance and persistence declines are concentrated among participants who obtained direct solutions from AI.
Experiment 3: Convergent evidence from reading comprehension. Here the researchers tried to verify if the impairment effect of AI goes beyond arithmetic problems by replicating the exercise in the domain of reading comprehension. Participants in the AI condition were presented with a series of 5 reading comprehension problems, with an AI assistant (GPT-5) available in a sidebar. The AI assistant was then removed, and participants were asked to solve 3 additional reading comprehension problems. At the same time the control condition group was presented 8 problems all without AI assistance. Again, the solve rate dropped in the AI condition while the skip rate increased slightly.

The paper is very short, so I suggest you read it thoroughly and evaluate the results by yourself. The conclusion of the paper is that current AI assistants represent a new kind of cognitive scaffold that solves anything instantly without refusing to help (with some guardrail exceptions). Even a 10-15 minutes interaction with AI assistants has an impact on the performances and on persistence.

The greatest loss

The described decline of persistence is the most concerning part in my opinion. Our learning processes and skill development are inherently based on a large component of persistence. If, somehow, this behavior is weakened, we could be under a dangerous evolutionary threat.

According to the researcher, there are two main reasons of the decline of persistence:

the resolution speed of problems from AI assistants, which make human-based parallels too much effortful
the removal, caused by AI, of the productive struggle that leads to the development of knowledge and self-knowledge.

That struggle is what leads us to develop new skills. Again, let me make a musical example: I learned (and still learn) to play guitar during years of repeted links, chords, passages. When I approach a new song I need many hours to internalize its structure, changes, voice leading and transfomr into something that I can “hear” in my head and play. That large amount of time is blessed and nobody can skip it if they want to be serious with music.

Every learning journey takes time and demands persistence and self-compassion to accept the errors and retry.

AI is not the first case of cognitive offloading in the history of humanity. Writing, calculators, computers, they helped us to achieve complex tasks by notating, decomposing or programming a machine. However, the current massive generative AI usage is creating a huge gap from the previous examples.

In the best case of day by day usage, people use GenAI for faster searches and are quite prone to accept the not so uncommon allucinations or underterministic results as a tradeoff that can be mitigated with prompting tecniques (for the most advanced users, though).

But wait, don’t you work with AI too?

Yes, I work for an open source software company, Red Hat that is massively investing and believes in AI platform solutions. I personally work with customers in public and private sectors to design their own Sovereign AI Infrastructure that can allow them to be compliant with the EU AI Act and keep control over users data and intellectual property while leveraging open weight models for transparency and auditability. And those use cases provide a foundation for great use cases, often based on agentic patterns, from healtcare to finance or public services for citizens.

These are corner cases however: the vast majority of users rely on closed models on public services like Claude or ChatGPT that are used to assist the personal work or study, sometimes to replace it, unfortunately. This is also a common concern among teacher for their students using GenAI to do their homeworks without struggling on the resolution of problems or the creative part of writing their essays.

The road less travelled

I’m not trying to provide a solution here, but settle some provocative hints that can be useful to approach the usage of GenAI from a different perspective.

Problem Solving: insted of asking a model to solve our problem or exercise, let’s ask it to generate new or similar ones to practice more and increase the level of challenge as we progress.
Challenging peer: when working in creative or scientific contexts, we can use the model as a skilled peer (by leveraging the proper pronting techniques) whose purpose is to challenge our ideas and push us over our though boundaries. This can be a stimulating practice that is food for mind.
Unvibe Coding: this is how I’d call a variation of the challenging peer on coding. Istead of having agents generating code for us, we have them challeng and evaluate our project to find out all the possible flaws in logic or security. If vibe coding lowers the barrier, unvibe coding raises it to help us become more proficient and overcome our programming weaknessees while the productive struggle stays ours.

So the distinction to me is not wheter to use AI, but when and how. Because I strongly believe that dependencies form when coming to AI before you have personally wrestled with something yourself.

Feel your breathe and use your mind.