I use AI almost every day now.

I use it for coding, reviewing pull requests, writing articles at work, making decisions, and sparring with ideas when I’m not sure whether my thinking is strong enough yet.

Most of the time, it helps.

It helps me move faster. It gives me alternative angles. It catches things I may have missed. It gives me a starting point when the blank page feels too blank. When used well, it can feel like having a patient collaborator who is always available.

But I’ve also noticed something else.

Sometimes, I’m not using AI to think better. I’m using it to avoid the harder first step of thinking.

That uncomfortable realisation led me back to an older piece of research on media multitasking and cognitive control.

TL;DR
AI is not making us think less by default. But it does make it easier to skip the messy early parts of thinking: forming our own view, sitting with uncertainty, and building judgment. Used well, AI can sharpen our ideas. Used passively, it can make polished output feel like understanding. The real skill is not just prompting better, but knowing when to think first, when to use AI, and how to critically evaluate what comes back.

In 2009, researchers Eyal Ophir, Clifford Nass and Anthony Wagner published a Stanford study on heavy media multitaskers. The study found that people who frequently juggle multiple media streams were not necessarily better at multitasking. In fact, they were often worse at filtering irrelevant information, managing working memory, and switching between tasks efficiently (Ophir, Nass & Wagner, 2009).

That finding feels even more relevant now. But the AI era changes the problem.

The issue is no longer just that we are switching between too many tabs, notifications and apps. The issue is that we now have tools that can write, summarise, code, analyse, decide and suggest on our behalf.

AI does not make us think less by default.

But it does make it much easier to skip the parts of thinking that build judgment.

The old multitasking problem was attention

The 2009 Stanford study introduced a Media Multitasking Index to compare heavy and light media multitaskers. Heavy media multitaskers were more affected by external distractions, more susceptible to interference from irrelevant memories, and less efficient at task switching (Ophir, Nass & Wagner, 2009).

The paradox was interesting: people who multitasked more were not necessarily better at multitasking.

A simple interpretation is that heavy multitaskers may become more open to everything around them. They scan broadly. They notice more. But the cost is that they may become less effective at filtering what matters.

That distinction is important.

Good thinking is not just about taking in more information. It is also about deciding what to ignore.

Later research has made the picture more nuanced. A ten-year review of the original “Cognitive Control in Media Multitaskers” paper found that the relationship between media multitasking and cognitive control is real but mixed. The results depend on how media multitasking is measured, which cognitive functions are tested, and the context in which people are using media (Parry & le Roux, 2021).

So I don’t think the right conclusion is “multitasking destroys your brain.”

A more careful conclusion is this:

When we constantly split attention across many streams, we may become more reactive to incoming stimuli and less practised at sustained, goal-directed thinking.

That matters because AI is not just another stream of information. It is a stream of possible answers.

The AI-era problem is not just distraction

Before generative AI, my multitasking looked something like this: Slack, email, calendar, Confluence, Jira, browser tabs, Figma, BitBucket, maybe a meeting running in the background.

That was already fragmented.

Now AI adds another layer.

When I use AI for coding, I’m not just writing code. I’m reviewing suggestions, checking assumptions, testing whether the generated code fits the architecture, and deciding whether I trust it.

When I use AI to review a pull request, I’m not just reading the diff. I’m checking whether the AI has found something meaningful or whether it is confidently pointing at something irrelevant.

When I use AI to write, I’m not just drafting. I’m deciding whether the output sounds like me, whether the claims are defensible, whether the tone is appropriate, and whether the structure is actually useful.

When I use AI to make decisions or validate ideas, I’m not outsourcing the decision entirely. At least, I shouldn’t be. I’m using it to challenge my framing, surface blind spots and pressure-test assumptions.

This is where the cognitive load shifts.

AI can reduce the effort of producing an answer, but it can increase the responsibility of evaluating one.

That is a different kind of work.

Cognitive offloading is useful, until it replaces understanding

Cognitive offloading is not new. We have always used external tools to reduce mental effort. Notes, calendars, calculators, checklists and search engines all help us think by moving some burden outside our heads.

Risko and Gilbert define cognitive offloading as using physical action or external tools to reduce the information-processing demands of a task (Risko & Gilbert, 2016). In that sense, AI is part of a long history of humans extending cognition through tools.

The problem is not offloading itself.

The problem is offloading too early, too often, or too passively.

There is a difference between asking AI to help after I have formed my own view, and asking AI before I have done the work of understanding the problem.

The first can sharpen my thinking.

The second can quietly replace it.

For example, if I ask AI to critique an article after I have written a rough argument, it can help me improve the piece. But if I ask AI to create the argument before I know what I believe, I may end up editing someone else’s reasoning rather than developing my own.

That distinction feels small, but it matters.

The “struggle phase” is not wasted time. It is where understanding forms.

AI can make us feel productive while reducing critical effort

A 2025 Microsoft Research study surveyed 319 knowledge workers and collected 936 examples of generative AI use at work. The researchers found that confidence plays an important role: when people had higher confidence in AI, they tended to apply less critical thinking effort; when they had higher confidence in themselves, they tended to apply more critical thinking effort (Lee et al., 2025).

That finding resonated with me.

The danger is not that AI gives bad answers all the time. The danger is that AI often gives fluent, plausible answers quickly.

Fluent answers can feel like progress.

But fluency is not the same as truth. Speed is not the same as understanding. A polished draft is not the same as a considered point of view.

This is especially risky in knowledge work because the output can look finished before the thinking is finished.

But if the human has not actively questioned the output, the work may only be superficially strong.

“Cognitive debt” is a useful warning, but not settled science

There was research exploring whether heavy reliance on AI tools may reduce cognitive engagement.

A 2025 MIT Media Lab preprint studied participants writing essays using ChatGPT, search engines or no tools, while measuring EEG activity. The study reported that participants in the ChatGPT group showed lower brain connectivity, lower ownership of their essays, and weaker recall of their own writing compared with participants who wrote without tools or used search (Kosmyna et al., 2025).

The phrase “cognitive debt” has been used to describe this possible pattern: AI may reduce mental effort in the short term, but if overused, it may leave people less practised at independent thinking.

That is a powerful idea.

But it should be treated carefully.

The MIT study was a preprint, not settled scientific consensus. A later comment on the paper raised methodological concerns, including sample size, reproducibility, EEG analysis and interpretation of the results (Stankovic et al., 2026).

So I would not claim that AI is “damaging the brain” or “weakening the prefrontal cortex.” That is too strong.

A more responsible claim is:

Early evidence suggests that passive or over-reliant AI use may reduce cognitive engagement in some contexts, especially when people use AI to produce work before they have formed their own understanding.

That is enough to take seriously without turning it into moral panic.

The real question: did AI help me think, or did it think for me?

This is the question I’m trying to ask more often.

When I use AI now, I try to notice which mode I’m in.

Sometimes I’m using AI as a collaborator:

• challenge this assumption

• show me what I’m missing

• critique this argument

• give me another way to frame this

• help me test whether this decision is sound

Other times, I’m using AI as a shortcut:

• write this for me

• decide this for me

• make this sound smart

• give me the answer so I don’t have to sit with the ambiguity

The first mode can build thinking.

The second mode can weaken it if it becomes the default.

AI is incredibly useful for expanding possibilities. But judgment still needs to belong to the human.

That means the most important skill is not prompting.

It is knowing when to prompt, when not to prompt, and how to evaluate what comes back.

How I’m trying to use AI without outsourcing my thinking

I don’t want to stop using AI. That would be unrealistic and, honestly, unhelpful.

Instead, I’m trying to use it with more intention.

1. Think first, prompt second

Before asking AI for an answer, I try to spend a few minutes forming my own view.

This gives me something to compare the AI response against. Without that, it is too easy to let the AI set the frame.

2. Use AI to challenge, not just generate

The most useful prompts are often not “write this for me.”

They are:

• What are the weaknesses in this argument?

• What would a sceptical reviewer challenge?

• What assumptions am I making?

• What are three alternative interpretations?

• What evidence would make this claim stronger or weaker?

This keeps me in the role of thinker, not just editor.

3. Separate creating from verifying

AI can help create options quickly, but verification needs a different mindset.

When reviewing AI output, I try not to skim. I look for unsupported claims, missing nuance, generic phrasing, false confidence and facts that need checking.

This matters even more when writing publicly. If I publish something, the responsibility is mine, even if AI helped me draft it.

4. Protect deep work from constant AI interruption

It is tempting to keep AI open all day as a permanent thinking companion.

But I’m starting to think AI works better in batches.

If AI is always open, it can become another form of ambient attention capture.

5. Keep some thinking analogue

For complex decisions, I still find paper useful.

A piece of paper or an online whiteboard forces slower thinking. It does not autocomplete my thoughts. It does not offer a polished answer. It makes me sit with the mess a little longer.

That friction is useful.

The human advantage is shifting

If AI can generate, summarise and execute faster than us, then our value is not simply in producing more.

Our value is in judgment.

The old productivity ideal was often about speed, volume and responsiveness.

The new one may be more about discernment, focus and deliberate attention.

In a world of infinite generated content, the scarce skill is not output.

It is taste, judgment and the ability to think clearly before the machine starts completing the sentence.

Final thought

AI does not make us think less by default.

But it does make it easier to confuse output with thinking.

Used well, AI can sharpen our reasoning, expose blind spots and help us move from rough intuition to clearer expression. Used passively, it can help us avoid the very struggle that builds understanding.

But let’s be honest: not every task deserves deep, agonising 'struggle phase' thinking. If you are drowning in corporate boilerplate, repetitive administrative tasks, or syntax debugging, using AI as a pure shortcut isn't a loss of critical thinking, it’s an act of professional survival. The real trick is figuring out where to draw the line between busywork and core craftsmanship.

That is the line I’m trying to pay attention to.

Not “Should I use AI?”

But:

Did AI help me think better?

Or did it help me skip thinking?

I'm realizing that the line between AI as a 'collaborator' and AI as a 'shortcut' moves a little bit every day. For those who use it daily: where did that line land for you on the last work or project you shipped?

---

Disclosure

I used AI to help structure early notes, identify relevant research directions and refine the article. I reviewed, rewrote and fact-checked the piece before publishing. The opinions, framing and final responsibility are mine.

---

References

Ophir, E., Nass, C., & Wagner, A. D. (2009). Cognitive control in media multitaskers. Proceedings of the National Academy of Sciences, 106(37), 15583–15587.

Parry, D. A., & le Roux, D. B. (2021). “Cognitive control in media multitaskers” ten years on: A meta-analysis. Cyberpsychology: Journal of Psychosocial Research on Cyberspace.

Risko, E. F., & Gilbert, S. J. (2016). Cognitive offloading. Trends in Cognitive Sciences, 20(9), 676–688.

Lee, H.-P., Sarkar, A., Tankelevitch, L., Drosos, I., Rintel, S., Banks, R., & Wilson, N. (2025). The Impact of Generative AI on Critical Thinking: Self-Reported Reductions in Cognitive Effort and Confidence Effects From a Survey of Knowledge Workers. Microsoft Research.

Kosmyna, N., Hauptmann, E., Yuan, Y. T., Situ, J., Liao, X.-H., Beresnitzky, A. V., Braunstein, I., & Maes, P. (2025). Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task. arXiv preprint.

Stankovic, M., Hirche, E., Kollatzsch, S., & Doetsch, J. N. (2026). Comment on: Your Brain on ChatGPT: Accumulation of Cognitive Debt When Using an AI Assistant for Essay Writing Tasks. arXiv preprint.

DEV Community. (2024). Guidelines for AI-assisted Articles on DEV.