AI and Cognitive Decline: is AI changing your brain?

AI with ALLIE

The professional’s guide to quick AI bites for your personal life, work life, and beyond.

This newsletter took me weeks to write and rewrite. But I want you all to have these answers. Being a responsible AI user means looking at the downsides too. And hey, weird admission but I love your beautiful minds and hearts. So let’s dive into the good, the bad, and the funky of how AI might be impacting them.

My main focus is AI in business, and I've spent every day for the last decade talking about how founders, executives, business leaders, and individual contributors can use AI to transform and grow their companies and build their careers.

But the world of business doesn't live in a vacuum. AI's impact in business is felt throughout our economy, our environment, our family life, our children's schooling, our grocery stores, our music, our toothpaste, everything.

So it doesn't surprise me that the conversation around health and AI has escalated in the last few months. I pretended I was back in grad school, went deep into the science, and pulled out the what these studies actually show and actions that I think matter.

When students can cheat, they’ll cheat themselves

One of the clearest examples comes from a high-school math experiment in Turkey, published in PNAS in June 2025. Researchers worked with 839 “regular” (non-honors) 9th to 11th graders at one large high school across four 90-minute class sessions, split into three groups: one practiced with only their textbook and notes, one with a basic GPT-4 assistant, and one with a teacher-built AI tutor. Each session ran the same way: the teacher reviewed a topic, students practiced, and then everyone took a closed-book exam in the same class period.

Students given access to a basic GPT-4 assistant (the best model around in fall 2023, though it still got these particular math problems right only about half the time) scored 48% better during practice, relative to the control group. Amazing!

But, on the unaided exam at the end of the same session, when the students were forced to think on their own without GPT-4, their scores were 17% lower than the control group's - about 5.4 percentage points against a control average of 32.1%. The chat logs show why. In the first session, 67% of these students' opening messages simply asked for the answer or pasted in the question (read: mostly "just give me the answer"). And the assistant just handed over the answers. The AI instructions told it to “act” as a tutor, but also to "follow the student's instructions carefully," so when a student asked for the answer, it followed it carefully…and gave the answer!

They were skipping the learning. They were skipping the struggle.

They were just cheating.

But what if we used AI as a tutor rather than a crutch? The third group used the same GPT-4 chatbot with two changes to its instructions. The bot was told to give hints instead of answers, and two math teachers loaded the AI with the correct solutions and a list of common student mistakes, so it could check students' work without making things up. Students with the AI tutor did even better in practice (127% over the control), and their unassisted exam results showed no statistically detectable deficit or advantage relative to the control group. Those two changes removed the measured harm on the students' unaided test. I read that study and immediately appreciated all of my teachers 500x more (shoutout to Mr. Weis for telling me to keep pursuing math!).

The AI groups both walked out feeling good about their performance (and both were wrong). The students with the basic AI assistant didn't think they had learned any less, but their exam scores said otherwise. The AI tutor group believed they had done better on the exam than they actually did. Researchers call this the illusion of competence: when practice feels easy, your brain mistakes that for knowing the material.

Developers show the way you use AI can drastically alter your learning

In a separate study published January 2026, Anthropic researchers ran an experiment with 52 experienced Python developers (most had 7+ years of coding under their belts), recruited and paid through an online platform and working remotely. Everyone got up to 35 minutes to finish two small coding tasks in Trio, a Python library none of them had used before, and they were told to go as fast as possible. Half had a GPT-4o (old model) chat assistant that could write the full, correct code if asked. Half had no AI. Right after, with AI turned off for everyone, they all took the same 14-question quiz on the concepts behind the tasks.

The AI group averaged only 50% on the quiz, versus about 65% for the developers who didn't use AI. That's roughly a 15-point gap, and it's statistically significant. It's also unclear whether AI sped them up: on average the groups finished in about the same time, but about half the AI group retyped the AI's code by hand instead of pasting it. One analysis that set those people aside found the AI users were about 25% faster. But - and there is a BIG but - not all engineers in the AI group used AI in the same way.

Read that again: the way you use AI matters.

Upon review of the AI users' screen recordings, the researchers identified six groups with different AI usage behaviors. The groups are tiny, with two to seven people each, so we have to take this more from an anecdotal angle than a statistics angle, but it passed my sniff test. From best to worst quiz score:

Not all AI reliance is the same! You can "rely" on it as a challenger, a tutor, a background, or a devil's advocate and not see a performance dip. But bad AI reliance does look like it leads to a reduction in mastery. Stick to the green methods above.

If you're managing any junior employees, I want you to walk your team through the takeaways, and make sure they understand whether or not they're cheating themselves out of career growth. A little struggle is a good thing, and a finished task or high-quality output tells you nothing of how they used AI. My advice: require early-in-career professionals to walk you through their full workflow, not just the output.

That one viral MIT study: struggle first, then call in AI

The MIT-led Your Brain on ChatGPT preprint published June 2025 studied 54 people ages 18 to 39, mostly students from five Boston-area schools. They wrote SAT-style essays while wearing EEG caps that tracked how strongly different brain regions were communicating. Each person was placed in one of three groups: ChatGPT only, search engine only (normal 2019 Google-esque with AI answers blocked), or no tools at all. They wrote one 20-minute essay per session for three sessions, stayed in the same group the whole time, and the full study was spread over about four months.

Across those three sessions, the ChatGPT group showed the weakest brain connectivity (up to 55% lower than the no-tools group, with the search group in between) and had much more trouble quoting their own work. After the first session, 83% of the ChatGPT group couldn't quote a single sentence from the essay they had just written, versus 11% in each of the other two groups. Similar to the Anthropic study, most people read the headlines of this study, threw up their hands, and said, "See! I knew it! AI destroys your brain!"

But they missed the second part of the study. 

The researchers offered an optional fourth session, and 18 of the 54 came back: 9 from the ChatGPT group and 9 from the no-tools group. (The search group didn't get a fourth session.) This time, the groups swapped. The former ChatGPT users wrote with no AI, and the former no-AI writers got to ChatGPT. Everyone picked from prompts they had already written about earlier, so it was a second pass at a familiar topic.

The two groups came out very differently. The people who had struggled on their own first and then got ChatGPT showed higher brain connectivity than the ChatGPT group ever had in its first three sessions, and they remembered their work: 7 of the 9 quoted their essay correctly (weird metric that I wouldn’t care about as much but I digress). The people who had leaned on ChatGPT from the start and then lost it only quoted correctly 1 time in 9. Their brain connectivity landed somewhere in between: stronger than a first-timer's, but short of what the no-AI group reached by its third session.

(Note: annoyingly, the paper mostly compares the struggle-first group to the ChatGPT group's earlier sessions, not to people who stayed tool-free the whole way. So it shows that struggling first beat leaning on AI first but doesn't show that adding AI beat skipping it entirely.)

Said another way, if you're going to use AI, first struggle on your own and then pull AI in for support and as a challenger. If you do it the other way around, you're cheating yourself out of some important work. I'm also 0% surprised that participants who used AI to write their essays couldn't recall the specifics of their essays, but maybe that's a sign we should just…not grade students via essays?

Lower brain connectivity can show up in people with cognitive decline, but it isn't proof of decline on its own. And, summarizing as a non-scientist here, weaker connectivity while writing a 20-minute essay with ChatGPT looks a lot more like less effort on that one task and does not establish brain damage or long-term cognitive decline.

The more you trust AI, the less you check it

One Microsoft Research and Carnegie Mellon study published April 2025 surveyed 319 people who use AI at work at least weekly. Most were under 35yo (72%), and nearly all used ChatGPT (97%). Each walked the researchers through three real tasks in which they had used AI (936 tasks in total) and rated how much they trusted the AI to handle each one on a 1-to-5 scale. They also said whether they actually scrutinized what the AI produced (their own answer, not something researchers tested). They said they did for only 59% of tasks. And for every step up that 1-to-5 trust scale (meaning if I trusted it as a 1 versus a 2), the odds that someone questioned or checked the AI's work fell by about half. The more you trust the AI, the less you verify. The people who kept checking were the ones who felt confident in their own skills on that task. Confidence that they could do the task themselves raised the odds of a real review by about 30%, and confidence that they could spot a bad answer raised it by about 36%. Interestingly, how much someone liked AI in general made no difference. It came down to trust on that specific task.

TLDR: the less you trust the AI on that task, and the more you trust yourself and your ability to spot a bad answer, the more likely you are to inspect the heck out of the output. So the people least likely to check AI's work are the ones who trust the AI and don't feel confident in the subject. If that describes your newer hires, build the review step into how they work.

Workers also said their job was changing shape, with less time researching and drafting and more time checking, editing, and supervising what the AI produced. I’m not wholly convinced that people will be fulfilled by this type of work.

With previous technologies, we see lower odds of cognitive declines with use

A cool April 2025 meta-analysis I found in Nature (Human Behavior) covered 57 studies and 411,430 people age 50+, with an average starting age of 69yo, with longitudinal studies running 1 to 18 years. The authors call this group "digital pioneers," people who first met computers and the internet as adults. General digital technology use was associated with 58% lower odds of cognitive impairment. In the 7 studies that followed people over time and reported it, tech users showed a 26% slower rate of cognitive decline during follow-up (and the finding held up after controlling for education, income, health, and social support). Could still be that people whose minds stay sharp may simply keep using technology longer, and AI could be different than the iPhone or the internet since it actually completes work.

So, Allie, is AI damaging your brain?

After reading these studies cover to cover, my answer is: no, not if you use it well.

But none of the studies here show the opposite either ie that AI was making someone learn more or have higher brain activity than they would have on their own. The best case, for now, was a tie.The kids with the good AI tutor scored exactly like the kids with a paper textbook. The developers who kept asking AI "why" scored about the same as the developers who didn’t touch AI. They just got a lot more done along the way.

The default win with AI is output - output amount, quality, type, and likely speed. If you want the learning part too, you have to set it up that way from the get-go, for yourself, your kids, and your team.

IF you’re an IC: Take the first swing yourself, then bring in AI. When it writes something for you, ask it to explain its choices and quiz you before you move on.

IF you’re a parent: don’t have your kid use AI out the gate. Have them try it on their own first. Turn on ChatGPT's Study Mode or Claude's learning mode so it’s guardrailed and gives hints, not answers. Have your kid explain one problem to you, start to finish, with their laptop closed.

IF you’re a team lead or business lead: Don't just judge the output! Ask your people to walk you through how they made it, their assumptions, the tradeoffs, and understand what they verified and what they trusted, especially for new or more junior hires.

We are going to be learning alongside these systems for a lot longer than the next product launch. I want us thinking about the expertise we're building over the next 30 years, even while we enjoy getting a tedious task done before lunch.

Find joy in the struggle of learning, go to a Lecture on Tap, watch nerdy YouTube videos. That’s your brain getting a well-deserved workout.

Stay curious, stay informed,
Allie

P.S. Share this with a friend, concerned parent, or stressed out business leader.