Numeracy Illiteracy: AI and the Slow Erosion of Critical Thinking

Artificial Intelligence does not erase our skills, but shifts them. Entrusting complex reasoning to ChatGPT and other LLMs risks triggering a new form of "calcu

Every technological leap in human history has entailed a form of outsourcing: we delegated physical strength to steam engines, long-term memory to writing, and arithmetic calculation to silicon processors. Today, with the pervasive integration of generative artificial intelligence into intellectual workflows, we are witnessing an outsourcing of a profoundly different nature: that of analytical reasoning. The emerging sociological alarm does not concern the pure and simple loss of notions, but the genesis of a new cognitive "calculation illiteracy."

However, a rigorous analysis of the scientific literature requires us to avoid moral panic: artificial intelligence does not automatically "erase" human logical skills, but displaces and transforms them. If used as an intellectual scaffold — to compare divergent hypotheses, verify sources, or stress-test one's own reasoning — AI can support and elevate critical thinking. If, on the contrary, it is adopted as a substitute oracle for laborious calculation, painstaking verification, and ethical decision-making, it risks triggering a dangerous atrophy: it reduces self-efficacy, empties the sense of ownership over one's work, and undermines the individual's ability to solve complex problems autonomously.

Cognitive Offloading: Better Results, Less Competence

The heart of the problem lies in the phenomenon known in psychology as cognitive offloading. A central and decisive study, presented at the CHI 2025 conference by researchers from Microsoft Research and Carnegie Mellon University, investigated the impact of GenAI on the critical thinking of 319 knowledge workers through the analysis of 936 real use cases [2967, 2975]. The data reveal a perverse mechanic: the use of AI drastically reduces the perceived effort required to apply critical thinking, fostering excessive reliance on the machine. The study demonstrates a very clear correlation: greater prior trust in AI is associated with less application of human critical thinking, while greater trust in one's own intellectual abilities translates into a more rigorous analysis of algorithmic output.

This dynamic generates a formative paradox, well illustrated by an experimental study published in Cognitive Processing (2026) in which young adults had to identify logical fallacies in complex texts [2968]. The group supported by ChatGPT obtained a significantly higher number of correct answers, while concurrently reporting lower mental effort. Yet, on deeper analysis, this same group showed alarming deficits in independent reasoning, in the propensity for autonomous verification, and in metacognitive monitoring. An insidious gap is thus consolidated: the company or individual obtains a "better result" in the short term, but the human being surrenders vital shares of "retained competence."

A further investigation published in Societies (2025) on a sample of 666 participants highlighted a negative correlation between frequent use of AI tools and structured critical thinking capacity, mediated precisely by the tendency toward cognitive offloading, while noting however that a robust higher education acts as a protective factor [2976, 2977].

The Illusion of Understanding and the Risk of the Passive Approver

Why do we trust a syntactic calculation so much, and so quickly? The psychological mechanism is linked to the linguistic interface of frontier models. As argued in an essay on arXiv (2025) dedicated to protecting and augmenting human cognition, Large Language Models produce output that is intrinsically coherent, apparently exhaustive, and linguistically complacent [2973]. This formal perfection has the power to bypass our natural skepticism, inhibiting the necessary suspension of judgment and creating in the user a dangerous "illusion of complete understanding." A well-written text formatted with algorithmic confidence proves to be cognitively far more insidious than a text that is blatantly ungrammatical or incorrect.

The operational risk is the transformation of the professional into a mere "passive approver." A Microsoft Research report (The Future of AI in Knowledge Work, 2025) summarizes this dynamic: when the stakes are perceived as very high, professionals apply severe critical thinking to AI output. But in routine tasks, or when operating under strong time pressure, critical sense collapses [2975]. In a healthy ecosystem, the golden rule of the synthetic era should be: delegate procedural execution, never delegate judgment. The center of gravity of intellectual work must inevitably shift from the pure generation of content toward the art of cross-verification, the integration of disparate sources, and the ethical supervision of the task.

Beyond Determinism: The Lesson of the Calculator

The current debate on cognitive decline often risks degenerating into Luddite catastrophism. As emphasized by the IEEE Computer Society (2025), cognitive offloading was not born with ChatGPT. We have already delegated spatial orientation to GPS and arithmetic to the pocket calculator [2990]. Using external tools to reduce low-level cognitive load frees precious resources that the human mind can (and must) reinvest in higher-order tasks. The problem does not lie in the tool itself, but in identifying with precision which human skills we stop exercising daily. Are we delegating the calculation of derivatives, or are we delegating the understanding of the mathematical problem?

In this context, methodological rigor is essential. Researchers at Carnegie Mellon University severely warn that current fears, though founded, are based on preliminary evidence. Observational studies linking heavy AI use to reduced functional brain connectivity during learning or to reduced memorability of information show strong correlations, not definitive causal links [2971, 2972]. It is imperative to maintain caution and not transform an alarming statistical correlation into an ineluctable prophecy about the end of human intellect.

Designing Friction: The Design of Epistemic Agency

If the problem is not the machine but the posture with which we relate to it, the solution must also be sought in software design. Researchers propose a paradigm shift: designing AI systems that aim to augment retained human cognition, instead of brutally replacing it [2974]. An artificial intelligence designed to support critical thinking should not limit itself to providing the perfect answer; it should make explicit its degree of uncertainty, ask the user for the reasons behind a given assumption, show potentially conflicting sources, and leave ample margins, even in the user interface, for exploration and review of reasoning. It must reintroduce a healthy "cognitive friction."

The real question that intellectuals, scientists, and professionals must ask themselves today is not the banal defensive plea "should I stop using AI so as not to atrophy my brain?" The central question is: which specific part of the reasoning do I still claim to do myself?

If the algorithm is employed to calculate probabilistic scenarios, compare immense data sets, and propose structural variants, we are witnessing the greatest extension of human faculties ever conceived. But if we ask that same algorithm to decide autonomously what is true and what is false, which scientific or business hypothesis deserves to be tested, and which moral choice should be made, the human being abdicates their own role. The calculation illiteracy of the 21st century, ultimately, will not consist in the technical inability to write a good prompt to query a language model. It will be the dramatic inability to understand when the model is mathematically right, when it is plausibly hallucinating, and when a decision — by its very nature, ethical or strategic — cannot be delegated to anyone other than ourselves [2967, 2973, 2974].

Bibliographic References and Sources

  • [The Impact of Generative AI on Critical Thinking — CHI 2025; Microsoft Research and Carnegie Mellon University] [2967, 2975]
  • [Cognitive offloading, critical thinking and attitudes towards AI — Cognitive Processing, 2026] [2968]
  • [AI Tools in Society: Impacts on Cognitive Offloading and the Future of Critical Thinking — Societies, 2025; Michael Gerlich] [2976, 2977]
  • [The Future of AI in Knowledge Work — Microsoft Research, 2025] [2975]
  • [Understanding, Protecting, and Augmenting Human Cognition — arXiv, 2025] [2973]
  • [Designing AI Systems that Augment Human Performed vs. Replaced Cognition — arXiv, 2025] [2974]
  • [Cognitive Offloading: How AI is Quietly Eroding Our… — IEEE Computer Society, 2025] [2990]
  • [How AI Impacts Skill Formation — arXiv, 2026] [2971]
  • [AI may blunt our thinking skills — Carnegie Mellon University / New Scientist] [2972]

Article by the Editorial Staff of La Bussola dell'IA