Does Using AI Make Us Think Less?

The question we can no longer ignore

A few years ago, writing a report meant researching a subject, organizing information, building an argument, and finding the right words.

Today, an AI assistant can generate a polished first draft in seconds.

The same transformation is happening across education, programming, financial analysis, research, and everyday decision-making.

We are becoming more capable of producing answers. But are we becoming equally capable of understanding them?

Researchers in cognitive science, psychology, education, and human-computer interaction are beginning to investigate this question.

Their findings are more complicated than the headlines suggest.

The short answer: AI can reduce the mental effort required to complete certain tasks. Under some conditions, that can weaken independent learning and critical engagement. Under others, AI can support understanding and improve learning outcomes.

The distinction between completing a task and developing the ability to complete it is central to understanding the impact of AI on human cognition.

1. The science behind cognitive offloading

To understand what AI might change about human thinking, we first need to understand a concept called cognitive offloading.

In 2016, psychologists Evan F. Risko of the University of Waterloo and Sam J. Gilbert of University College London published a review in Trends in Cognitive Sciences examining how humans use external resources to reduce mental demands (Risko & Gilbert, 2016).

Their examples included reminders, physical actions, and other strategies that change how much information the brain must process or remember.

The principle is simple: rather than performing every mental operation internally, we shift some of that work into our environment.

This is not a new phenomenon.

In 2011, Betsy Sparrow, Jenny Liu, and Daniel M. Wegner published research in Science examining how access to online information influences memory.

Across four experiments, they found that when people expected information to remain accessible, they were less likely to remember the information itself and better able to remember where to find it (Sparrow et al., 2011).

This became widely associated with the idea of the "Google effect."

But remembering where to find information instead of memorizing every detail does not automatically mean that our intelligence is declining.

It means the way we distribute cognitive work can change.

Generative AI introduces a new dimension: we can now delegate not only information retrieval, but also parts of reasoning, drafting, problem-solving, and decision support.

That makes the question more significant.

What happens when the external system does not simply store information, but generates the answer itself?

2. What Microsoft Research discovered about critical thinking

In April 2025, researchers from Microsoft and Carnegie Mellon University published a study at the CHI Conference on Human Factors in Computing Systems.

The research team, led by Hao-Ping (Hank) Lee and including Advait Sarkar, Lev Tankelevitch, and colleagues, surveyed 319 knowledge workers and collected 936 examples of how they used generative AI during professional tasks (Lee et al., 2025).

The researchers investigated when users engaged in critical thinking and how AI influenced the mental effort they reported.

One finding was particularly revealing.

Participants who expressed greater confidence in AI tended to report less critical thinking during AI-assisted tasks.

Conversely, participants who expressed greater confidence in their own abilities tended to report more critical thinking.

But the researchers also discovered something important: AI did not simply eliminate thinking.

It changed where people directed their mental effort.

Instead of producing information from scratch, users increasingly focused on checking information, integrating AI-generated responses, and supervising the overall task.

The context mattered.

For routine, low-stakes activities performed under time pressure, participants generally reported expending less critical-thinking effort with AI.

For high-stakes activities requiring accuracy, some reported expending more effort checking AI-assisted work than they would have without it.

What this actually proves

The study identifies relationships between how people perceive AI, how much they trust it, and how they report exercising critical thinking.

It does not establish that AI causes permanent cognitive decline.

It also relies on participants' descriptions and self-assessments rather than directly measuring changes in long-term intellectual ability.

Nevertheless, the research highlights a meaningful concern:

When users become highly confident in an AI system, they may become less inclined to question its outputs.

That is especially relevant when AI is integrated into professional decision-making.

3. What happens inside the brain when we write with AI?

In June 2025, researchers associated with the MIT Media Lab released a study titled Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task.

Led by Nataliya Kosmyna, the team examined differences in cognitive engagement during essay writing (Kosmyna et al., 2025).

Participants were divided into three conditions:

• Writing without external tools.
• Writing with access to a search engine.
• Writing with assistance from ChatGPT.

The researchers used electroencephalography (EEG) to examine patterns of brain activity during the writing tasks.

A total of 54 participants completed the first three sessions, while 18 participated in a fourth session that changed their access to AI assistance.

The findings

The researchers reported differences in neural connectivity across the three conditions.

Participants writing without external assistance displayed the strongest and most widely distributed connectivity patterns in the measured networks.

Search-engine users showed intermediate patterns.

Participants using ChatGPT showed the weakest connectivity patterns under the study's measurements.

The AI-assisted group also reported a lower sense of ownership over their essays and had greater difficulty accurately quoting text they had recently submitted.

In the fourth session, participants who switched from ChatGPT-assisted writing to writing without assistance showed signs of reduced engagement in some measured neural networks.

These results raise questions about how AI-supported writing influences cognitive engagement, memory, and the development of writing skills.

What the study does not prove

This distinction is critical.

The research was released as a preprint and, according to the MIT project's published information, had not completed peer review when the initial paper was released.

It involved a relatively small sample, a specific essay-writing task, and only 18 participants in the final session.

EEG connectivity patterns are also not a direct measure of intelligence.

Less connectivity during a task cannot, by itself, establish that the brain has been damaged or that a person's intellectual capacity has declined.

The researchers themselves explicitly warned against presenting their findings as evidence that ChatGPT makes people less intelligent or causes brain damage (MIT Media Lab, n.d.).

The scientifically defensible conclusion is narrower: different forms of writing assistance were associated with different patterns of cognitive engagement and recall under the experimental conditions.

Whether these differences persist, generalize to other activities, or influence long-term cognitive development requires further research.

4. The most important distinction: completing work versus learning

Perhaps the clearest experimental evidence of this distinction comes from education.

In 2025, Hamsa Bastani, Osbert Bastani, Alp Sungu, and their collaborators published a study in the Proceedings of the National Academy of Sciences (PNAS).

The researchers conducted a field experiment involving nearly 1,000 high-school mathematics students (Bastani et al., 2025).

They compared three approaches.

One group completed mathematics exercises without access to a generative AI tutor.

A second group used GPT Base, a tool resembling a conventional AI chatbot.

A third group used GPT Tutor, a modified AI system designed to encourage learning by providing hints rather than simply revealing solutions.

The findings showed a striking difference between immediate performance and independent learning.

Note: These percentages are relative differences in measured grades compared with the control group. They are not percentage-point changes or measurements of intelligence.

The researchers also examined how students interacted with the tools.

GPT Base users frequently asked directly for answers and copied solutions.

GPT Tutor users engaged in more substantive interactions, including requesting help and attempting problems independently.

Why this matters

The students using unrestricted AI could perform better when the technology was available, yet some performed worse when they had to work independently.

The improved results during practice did not necessarily translate into stronger retained skills.

The GPT Tutor results suggest that the design of the assistance matters. Providing guidance without immediately supplying the entire answer largely mitigated the negative effect observed with unrestricted assistance.

This was a study of high-school mathematics, not a demonstration that the same outcome occurs in every profession or cognitive activity.

Still, it illustrates an essential principle:

AI can improve the work we produce without necessarily improving the abilities we develop.

5. But research also shows that AI can improve learning

So far, several studies suggest that excessive reliance on AI may reduce certain forms of cognitive engagement.

But treating these findings as proof that AI is universally harmful would ignore a substantial body of research pointing in the other direction.

In May 2026, Changxin Fan, Lele Ke, Zexiong Chen, and Pin Lv published a meta-analysis in Frontiers in Psychology.

The researchers synthesized 36 studies involving 7,229 participants and 132 effect sizes in higher education (Fan et al., 2026).

They reported an overall positive effect of generative AI on learning outcomes, with a standardized effect size of Hedges' g = 0.499.

The review also identified positive findings for higher-order learning outcomes, including understanding and cognitive or creative activities.

Instructional method emerged as an important factor, with particularly favorable results in collaborative and blended-learning settings.

However, another 2026 review urges greater caution

In August 2026, Chiara Boolzen and colleagues published a systematic review and meta-analysis in Artificial Intelligence Review focusing on cognitive learning outcomes in STEM education.

The researchers identified 85 eligible studies, including 49 studies suitable for quantitative meta-analysis (Boolzen et al., 2026).

Their conventional statistical model initially indicated a substantial positive overall effect.

But the results varied dramatically across studies.

The researchers reported very high statistical heterogeneity and evidence of possible publication bias.

After applying a robust Bayesian approach to account for publication bias, the analysis no longer supported a consistent positive or negative overall effect.

These reviews should not be treated as a direct contradiction in a controlled experiment: they included different bodies of evidence, used different statistical methods, and focused on partly different learning outcomes.

Together, they demonstrate why broad statements about AI making people smarter or less intelligent are scientifically premature.

The evidence increasingly suggests that the effects of AI depend on how the technology is used, what users are trying to learn, and how outcomes are measured.

6. Can AI help us think better instead?

The most interesting question may not be whether people should use AI, but how AI systems should interact with human reasoning.

In another paper presented at CHI 2025, Leon Reicherts, Zelun Tony Zhang, and colleagues investigated two approaches to AI-assisted decision-making (Reicherts et al., 2025).

Their research examined how different forms of AI support affected participants completing an investment decision task.

One system, called RecommendAI, offered recommendations.

The other, called ExtendAI, encouraged users to articulate their reasoning and then provided feedback that built on their own thinking.

The researchers found that these approaches influenced decision-making differently.

RecommendAI could introduce new options and perspectives while requiring less cognitive effort.

ExtendAI integrated more naturally into participants' existing reasoning processes and was associated with slightly better outcomes in the small comparative study.

The researchers also identified trade-offs.

Recommendations sometimes felt disconnected from users' own reasoning, while feedback based on existing arguments could reinforce initial assumptions.

The study's limited sample does not establish that one approach is universally superior.

But it points toward an important design question:

Should AI systems primarily provide answers, or should they also help users develop the reasoning needed to evaluate those answers?

7. The real risk may be invisible dependence

Taken together, the evidence suggests three potential risks worth monitoring.

1. Mistaking performance for understanding

A person may produce excellent work with AI assistance without being able to reproduce the underlying reasoning independently.

This distinction was demonstrated in the mathematics experiment by Bastani and colleagues (Bastani et al., 2025).

2. Reducing critical verification

When users place considerable trust in AI-generated responses, they may devote less effort to questioning those responses.

The Microsoft Research survey found an association consistent with this concern, although it did not establish a causal relationship (Lee et al., 2025).

3. Delegating learning-relevant mental work

When AI provides complete solutions, people may miss opportunities to practice analysis, problem-solving, and independent reasoning.

This is a plausible mechanism supported by some educational experiments, but its long-term implications across the general population remain uncertain.

These risks should not be confused with proof of declining human intelligence.

They are questions about how reliance on external systems may change what we practice, what we remember, and what we can accomplish without assistance.

8. How to use AI without abandoning independent thinking

The available evidence does not justify avoiding AI altogether.

It suggests that we should pay closer attention to how assistance is designed and used.

Based on the research discussed above, several practices are worth considering.

Think before prompting

When the objective is learning or developing expertise, attempt to formulate your own explanation or initial solution before asking AI for help.

Ask for guidance, not just answers

Request hints, alternative approaches, and feedback rather than immediately obtaining a completed solution.

The GPT Tutor experiment provides direct evidence that this distinction can matter in an educational setting (Bastani et al., 2025).

Verify important outputs independently

For consequential information, examine primary sources, identify assumptions, and check whether the reasoning actually supports the conclusion.

Ask AI to challenge your reasoning

Instead of seeking confirmation, use the model to identify missing evidence, counterarguments, and potential weaknesses in your analysis.

Test what you can do without assistance

When learning a new skill, occasionally complete comparable tasks without AI. The goal is to distinguish improved performance with assistance from improved independent capability.

These are practical recommendations informed by the research, not a clinically validated protocol guaranteed to prevent cognitive decline.

The objective is not to maximize mental effort in every situation.

It is to preserve the forms of thinking that matter for understanding, learning, and responsible decision-making.

9. Mindzy Perspective: AI should expand human capability, not replace human judgment

At Mindzy, we believe the most useful way to evaluate an AI system is not simply by measuring how quickly it produces an answer.

We should also ask what role it leaves for human understanding, supervision, and decision-making.

In a business environment, automating repetitive work may create significant value.

But a system that produces recommendations, analyses, or decisions without effective oversight can also introduce new risks.

The challenge is to determine which activities should be delegated, which require verification, and which should remain under meaningful human control.

AI Employees, enterprise agents, and automated workflows make this distinction increasingly important.

The objective should not be to remove human thinking from every process. It should be to use AI where it creates value while preserving human accountability and the capacity to question its outputs.

The future of AI should not be measured only by how much work machines can perform for us.

It should also be judged by how effectively humans can understand, direct, and evaluate the systems they rely on.