Are We Outsourcing Our Thinking to AI?

The promise of artificial intelligence is that it can make us more productive, knowledgeable and capable. But a growing body of research raises an uncomfortable question: Could the very technology designed to make us smarter gradually weaken our ability to think independently?

AI can solve problems, write essays, explain complicated subjects, summarize research and produce answers in seconds. It can also spare us the frustration of working through difficult questions.

That convenience is enormously useful. But there is a potential downside.

Learning isn't simply about obtaining the right answer. It involves the mental work required to reach that answer: remembering, questioning, comparing, making mistakes, correcting them and gradually developing an understanding of the subject.

If AI routinely performs that work for us, we may become better at getting things done while becoming less capable of doing them independently.

This isn't merely a philosophical concern anymore. Researchers are beginning to test what happens when people use AI to perform cognitive tasks and then have to perform those same tasks without assistance.

The findings deserve attention. But they also require some careful interpretation. The evidence does not establish that AI makes people permanently less intelligent. It does suggest that certain patterns of AI use can interfere with learning, persistence and independent performance.

Let's examine what the research actually tells us.

1. The study that brought the question into focus

One of the most relevant investigations is a 2026 study by researchers Grace Liu, Brian Christian, Tsvetomira Dumbalska, Michiel Bakker and Rachit Dubey, titled AI Assistance Reduces Persistence and Hurts Independent Performance.

The study was presented at the 2026 Conference on Language Modeling (COLM) and examined 1,222 participants across three randomized experiments.

Unlike surveys asking people whether they thought AI was making them less capable, this research tested what participants could actually do.

What did the researchers do?

In the first experiment, 354 participants were asked to solve basic fraction problems.

• One group worked without AI.
• The other group could use ChatGPT to help solve the problems, including providing direct answers.
• After 12 problems, the researchers unexpectedly removed access to AI.

The AI-assisted group initially performed better. That wasn't surprising: they had access to an additional source of answers.

But once that assistance disappeared, their performance deteriorated.

In a subsequent experiment involving 667 participants, the researchers attempted to rule out alternative explanations, including differences in access to previously learned solutions. The pattern persisted.

A third experiment examined reading comprehension rather than mathematics. Again, participants who had used AI performed worse when they subsequently had to work independently.

Across the experiments, researchers found two concerning effects:

1. Reduced independent performance. People who had relied on AI performed less successfully when required to complete similar work without it.
2. Reduced persistence. They were more likely to abandon difficult questions rather than continue trying.

The researchers reported that these effects could emerge after approximately ten minutes of AI-assisted work.

Why the persistence finding matters

The second finding may be even more important than the first.

Being unable to solve a problem immediately isn't necessarily a disadvantage. Struggling with a difficult question forces us to explore alternatives, retrieve relevant knowledge and develop new approaches.

That process is part of how we acquire skills.

If we become accustomed to receiving immediate answers, the experience of working through uncertainty can become less appealing.

We may gradually develop a habit of seeking assistance before making a serious attempt ourselves.

The researchers suggest that AI changes our expectations about how much effort a problem should require. Once accustomed to instant answers, ordinary intellectual effort can feel disproportionately frustrating.

The danger isn't simply that AI knows the answer. It's that we may become less willing to find the answer ourselves.

One important qualification: the experiments establish short-term causal effects on the tasks studied. They do not demonstrate that ten minutes of AI use permanently damages someone's ability to think, or that the effects accumulate over years. That remains an important question for future research.

2. Does AI interfere with learning and memory?

The distinction between completing a task and learning something is fundamental.

Imagine two people studying the same subject.

The first reads a chapter, struggles with unfamiliar concepts, takes notes, tries to explain the material and tests their understanding.

The second asks ChatGPT to summarize the chapter, explain the important points and produce a set of answers.

Both might finish the exercise with an apparently good understanding.

But will they remember the material a month later?

A 2025 randomized controlled trial by André Barcaui, published in Social Sciences & Humanities Open, investigated precisely this issue.

The 45-day retention experiment

The study enrolled 120 undergraduate business students learning about artificial intelligence.

Participants were assigned to one of two groups:

• An AI-assisted group that used ChatGPT as a study aid.
• A traditional learning group that studied without generative AI.

Approximately 45 days later, participants took an unexpected knowledge-retention test.

The results were:

The AI-assisted group remembered significantly less of what they had studied.

This is particularly interesting because the study measured retention weeks after the learning experience, rather than simply asking which group completed the original task more successfully.

There are limitations. It was a relatively small university sample, participants were studying a particular subject, and almost 30% of the original group did not complete the delayed assessment.

It cannot establish that every form of AI-assisted learning produces poorer memory.

Nevertheless, it provides experimental evidence that unrestricted AI assistance can come at a cost to long-term retention.

3. What happens inside the brain when AI does the work?

One of the most widely discussed investigations came from researchers at MIT's Media Lab in 2025.

The study, Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task, examined how people engaged with writing tasks under different conditions.

Participants were divided into three groups:

• LLM group: used a large language model.
• Search group: used a conventional search engine.
• Brain-only group: worked without external tools.

Researchers used electroencephalography (EEG) to measure patterns of brain activity while participants wrote essays.

What did they find?

The brain-only group showed the strongest and most distributed patterns of neural connectivity.

The search-engine group showed intermediate engagement.

The AI group showed the weakest connectivity among the three groups.

The researchers also reported that participants who had relied on AI struggled more to accurately quote their own essays and reported a weaker sense of ownership over their work.

In a later session, some participants who had previously used AI were asked to write without it. Their brain activity showed reduced engagement compared with participants who had initially worked without AI.

How should we interpret this?

This is interesting research, but it has received more attention than its design can fully justify.

There were only 54 participants in the first three sessions, and just 18 completed the fourth session.

EEG connectivity is not a direct measurement of intelligence, understanding or cognitive ability.

Lower measured connectivity during a particular task does not establish that the brain has become damaged or that someone has lost intellectual capacity.

The paper was published as an arXiv preprint, rather than an established peer-reviewed clinical study.

It is best viewed as preliminary evidence that different ways of completing a task involve different levels and patterns of cognitive engagement.

The phrase the researchers use, cognitive debt, is an interesting description of the possibility that immediate convenience creates longer-term costs in learning and independent capability.

But it should not be confused with a medically established condition.

4. The counterargument: AI can also make us better learners

It would be misleading to examine only the negative findings.

There is substantial research showing that generative AI can improve learning outcomes under certain conditions.

A 2025 systematic review and meta-analysis published in Computers & Education examined 69 experimental studies, with 62 included in its quantitative analyses.

The researchers found that ChatGPT generally improved academic performance and affective-motivational outcomes. They also found improvements in higher-order thinking propensities, although AI use was associated with reduced mental effort.

That last finding is particularly interesting.

Reducing mental effort isn't automatically bad.

Consider the difference between:

• Spending an hour trying to understand an unfamiliar scientific concept.
• Having an AI tutor explain it in three different ways, provide examples and answer follow-up questions.

The second approach may reduce unnecessary frustration and make learning more accessible.

But there's a difference between reducing the effort required to understand something and eliminating the effort required to learn it.

A 2026 systematic review and meta-analysis published in Artificial Intelligence Review offers an important qualification.

It examined 85 eligible studies, with 49 included in the quantitative analysis.

The conventional statistical analysis suggested an overall positive effect of generative AI on STEM learning. However, after accounting for possible publication bias, that overall positive effect was no longer convincing.

The researchers also found substantial variation between studies.

Their analysis suggested that AI was more promising when it augmented students' own learning activities rather than substituted for them.

In other words, the research does not support a simple conclusion that AI either improves or damages learning.

The way we use AI appears to matter enormously.

A particularly revealing experiment: AI tutors that give hints instead of answers

A 2025 study published in the Proceedings of the National Academy of Sciences (PNAS) examined nearly 1,000 high school mathematics students.

Researchers compared two AI tutoring approaches:

• GPT Base, which behaved much like a conventional ChatGPT interface.
• GPT Tutor, which was designed to provide structured hints rather than simply give away answers.

Both approaches improved performance while students had access to AI.

The conventional GPT Base group achieved a 48% improvement in practice grades, while the safeguarded GPT Tutor group achieved a 127% improvement.

But when AI assistance was withdrawn, the GPT Base group performed 17% worse than students who had never received AI assistance.

The carefully designed tutor substantially mitigated this negative effect.

This is one of the most important findings in the emerging research.

It suggests that AI can be designed to encourage learning rather than replace it.

The distinction is between an AI that does your homework and an AI that helps you become better at doing your homework.

5. Cognitive offloading: we've been doing it for centuries

The phenomenon behind these concerns has a name: cognitive offloading.

It means using an external tool or resource to reduce the mental effort required to complete a task.

We do it constantly.

• We use calculators instead of doing arithmetic mentally.
• We use GPS instead of remembering directions.
• We use calendars instead of remembering appointments.
• We write shopping lists instead of memorizing everything we need.
• We use search engines instead of remembering every fact we've encountered.

There is nothing inherently wrong with this.

In fact, cognitive offloading can improve performance and free up mental resources for more demanding activities.

A 2025 review in Nature Reviews Psychology examined the benefits and potential costs of offloading information from memory.

It found that external aids can improve performance, but there can be a disadvantage when those aids become unavailable unexpectedly. People may perform worse when they have relied on external information rather than maintaining it internally.

The important distinction is what we are outsourcing.

Using a calculator to perform arithmetic doesn't necessarily prevent us from understanding mathematics.

Using GPS doesn't necessarily prevent us from understanding geography.

But if we never learn how to calculate, navigate or reason independently, we become increasingly dependent on the tools.

AI introduces something different from many earlier technologies.

It doesn't just store information or perform a narrow operation.

It can perform activities that have traditionally required substantial human cognitive effort:

• Constructing arguments.
• Evaluating evidence.
• Summarizing complex information.
• Identifying patterns.
• Generating explanations.
• Making recommendations.
• Writing and problem-solving.

It can potentially take over much of the process through which we develop knowledge and judgment.

That makes the consequences of cognitive offloading particularly important.

The danger of confusing access to knowledge with possession of knowledge

Imagine asking an AI to explain the causes of World War II.

You receive a comprehensive explanation in seconds.

You understand it while reading.

But could you explain the causes yourself tomorrow, without looking at the answer?

Could you distinguish the major contributing factors from the minor ones?

Could you identify an inaccurate historical claim?

Those are different abilities.

Having access to an explanation doesn't necessarily mean you have developed the knowledge needed to evaluate it.

And being able to repeat an answer isn't necessarily the same as understanding it.

This distinction becomes especially important when AI produces convincing but incorrect information.

6. Does AI affect critical thinking?

This is probably the most consequential question for everyday users.

Critical thinking involves more than acquiring information.

It requires the ability to:

1. Identify assumptions.
2. Distinguish evidence from opinion.
3. Recognize misleading arguments.
4. Evaluate the credibility of sources.
5. Consider alternative explanations.
6. Recognize uncertainty.
7. Change your mind when the evidence warrants it.

AI can help with every one of these activities.

It can also undermine them.

Consider two approaches to researching a health claim.

Approach A: Delegating judgment

You ask AI whether a supplement prevents heart disease.

It gives you a confident explanation, lists some studies and concludes that the supplement may be beneficial.

You accept the answer.

Approach B: Using AI to challenge your thinking

You ask AI to identify the strongest evidence for the claim, the strongest evidence against it, distinguish human trials from animal studies, identify methodological weaknesses and explain what remains unknown.

You then examine the original research and decide what the evidence actually supports.

The technology is the same.

The intellectual outcome can be very different.

A 2026 study examining AI-assisted and manual reasoning tasks involving logical fallacies found that AI assistance improved immediate accuracy and reduced reported mental effort. However, the researchers also observed different patterns of independent reasoning, verification and metacognitive monitoring between participants. They did not find that AI uniformly improved or diminished critical thinking.

A separate 2026 systematic review of 89 studies examining higher-order cognitive skills found mixed and conditional outcomes. Over-reliance was the most frequently identified cognitive risk, while positive outcomes were also reported. The authors emphasized the importance of how AI is incorporated into learning.

The evidence points toward a distinction between using AI to think more effectively and using it to avoid thinking.

7. What about older adults? Is there a particular risk?

This deserves attention because AI adoption is often discussed in terms of younger people, students and the workplace.

Older adults can benefit enormously from AI.

It can help explain unfamiliar technology, interpret complicated information, make sense of government correspondence, organize personal projects and provide access to knowledge that might otherwise be difficult to obtain.

It can also make learning more accessible.

However, there is a particular issue worth considering: AI can sound authoritative even when it is wrong.

A fluent, well-organized answer can create an impression of expertise that the underlying information doesn't deserve.

A 2026 article in The Gerontologist, titled From Access to Agency: Empowering Older Adults for Critical AI Use, discusses this problem.

The authors argue that older adults need more than access to AI and instruction on how to operate it. They also need the skills to evaluate AI-generated information critically.

They describe a risk they call epistemic vulnerability: people may trust AI outputs because their confident presentation resembles the authoritative sources they have traditionally relied upon.

Importantly, this is a conceptual framework, not evidence that older adults are inherently less capable of evaluating AI.

There is currently insufficient direct experimental evidence to conclude that AI causes greater cognitive decline or loss of independent reasoning in older adults than in younger people.

We should not assume that age itself makes someone more vulnerable.

But regardless of age, anyone who accepts AI-generated information without questioning it risks becoming dependent on a system that can make mistakes.

An opportunity as well as a warning

Older adults bring something valuable to the AI conversation: accumulated experience.

They have encountered changing technologies, political narratives, advertising claims, financial promises and medical fashions.

That experience can provide useful perspective when evaluating new information.

The challenge is to combine accumulated human judgment with the capabilities of AI, rather than surrendering one to the other.

8. How to use AI without surrendering your intellectual independence

The research suggests that the solution isn't to avoid AI.

It is to change the relationship we have with it.

Here are practical habits that can help preserve independent thinking while taking advantage of AI's capabilities.

1. Think before you ask

Before asking AI a question, spend a few minutes considering what you already know.

Form your own preliminary opinion.

Even if it's incomplete or incorrect, having an initial position gives you something to compare against.

2. Ask for hints before answers

When learning something new, ask AI to guide you rather than immediately solve the problem.

Try:

"Don't give me the answer yet. Help me work it out by asking questions and providing hints."

This encourages active participation.

3. Ask AI to disagree with you

One of the most useful ways to use AI is as an intellectual opponent.

Try:

"Challenge my reasoning. Identify the weakest parts of my argument and give me the strongest opposing case."

This helps expose assumptions and blind spots.

4. Verify important claims

AI can produce inaccurate information with remarkable confidence.

For important decisions, ask for original studies, government documents and authoritative sources.

Read the underlying evidence rather than relying exclusively on AI's interpretation.

5. Test what you have learned

After AI explains something, close the conversation and try to explain the subject in your own words.

If you cannot do that, you may have understood the explanation without retaining the knowledge.

6. Do some things without AI

Continue reading, writing, calculating, remembering and solving problems independently.

Not every task needs to be optimized for speed.

Some mental activities are worth doing precisely because they exercise our abilities.

7. Use AI to expand your thinking, not finish it

Ask for different perspectives, alternative explanations, counterarguments and unanswered questions.

Let AI widen the field of possibilities.

Keep responsibility for deciding what is credible and what you believe.

9. The bigger question: are we trading capability for convenience?

There is a broader issue here that extends beyond education.

Modern technology has progressively reduced the amount of effort required to accomplish everyday tasks.

We no longer need to remember telephone numbers, navigate unfamiliar roads, calculate complicated figures or search through physical encyclopedias.

Most of these changes have brought genuine benefits.

But there is a difference between making a task easier and making ourselves less capable of performing it.

AI takes this development a step further because it can increasingly perform the intellectual work itself.

It can write the argument, find the evidence, formulate the conclusion and explain why the conclusion is correct.

That leaves us with an interesting question.

If we increasingly delegate the activities that develop our judgment, reasoning and understanding, what happens to those abilities over time?

We don't yet have sufficient long-term research to answer that question conclusively.

But we have enough early experimental evidence to take the possibility seriously.

There is also a potential feedback loop:

1. AI makes intellectual tasks easier.
2. We become accustomed to receiving immediate answers.
3. We practice independent problem-solving less frequently.
4. Difficult intellectual tasks become less attractive.
5. We increasingly rely on AI to perform them.

Whether this becomes a lasting pattern of cognitive dependence remains to be established. But the mechanisms are plausible, and the early experimental findings provide reasons for concern.

10. Our verdict: AI is neither the enemy nor the answer

The research doesn't justify the headline that AI makes people less intelligent.

It does provide evidence that certain forms of AI reliance can weaken independent task performance, reduce persistence and interfere with knowledge retention.

Other research shows that AI can improve learning, provide useful explanations and make difficult subjects more accessible.

The difference appears to depend substantially on whether AI supports our own cognitive activity or replaces it.

We should also recognize the limitations of the evidence:

• Many studies involve students rather than representative samples of the general population.
• Some experiments measure performance immediately after AI assistance is withdrawn, not long-term intellectual development.
• The MIT brain-activity study was small and preliminary.
• Long-term effects on general intelligence, memory and critical-thinking ability remain uncertain.
• We have little direct evidence comparing the effects on older and younger adults.

The research is developing rapidly. Stronger, longer-term studies involving more diverse populations are needed.

But one conclusion is already worth taking seriously:

Being able to obtain an intelligent answer is not the same as becoming an intelligent person.

AI can give us access to more knowledge than any previous generation could easily obtain. It can help us explore unfamiliar subjects, challenge assumptions and investigate complicated questions.

But it cannot replace the value of developing our own understanding.

We should use these tools to extend our intellectual capabilities, not gradually surrender them.

Evidence assessment: MODERATE

There are credible randomized experiments demonstrating short-term effects on independent performance, persistence and some measures of learning retention. Evidence for lasting reductions in general intelligence or cognitive decline is currently insufficient.