---
title: "Is AI making us dumber? What the research shows"
description: "AI is not lowering raw intelligence, but the 2026 studies show heavy, undirected use erodes the memory and critical thinking you stop practicing."
url: https://buildfirstbrain.com/journal/is-ai-making-us-dumber/
canonical: https://buildfirstbrain.com/journal/is-ai-making-us-dumber/
author: "Lawrence Arya"
authorUrl: https://www.linkedin.com/in/vibecoding/
published: 2026-06-09
updated: 2026-06-09
category: "Mind & Learning"
tags: ["cognitive offloading", "ai and cognition", "critical thinking", "ai cognition", "first brain"]
lang: en
---

# Is AI making us dumber? What the research shows

> **TL;DR** AI is not lowering underlying intelligence, but heavy, undirected use measurably weakens the thinking skills people stop practicing. The MIT 'Your Brain on ChatGPT' study found AI-assisted writers had the weakest brain connectivity, could not accurately quote their own essays, and felt little ownership, and broader research finds short-term gains traded for eroded memory and critical thinking. The viral percentage figure is not in the paper. The damage is to delegated skills, not raw intelligence, so it is reversible and depends on how you use the tool. The fix is to do the hard thinking yourself and use AI to amplify a built First Brain.

AI is not lowering anyone's underlying intelligence, but the way most people use it is measurably weakening the thinking skills they no longer practice, and the 2026 research is now specific about how. The clearest signal comes from a controlled MIT study in which people who wrote essays with an AI assistant showed the weakest brain connectivity of any group, struggled to quote their own finished work, and felt the least ownership of it. Other research finds the same shape: a short-term performance boost paid for with a longer-term erosion of memory and critical thinking, because the mental effort you hand to the model is effort your brain stops doing. The fix is not to avoid AI but to use it as a co-processor while you keep doing the thinking that matters, which is the entire case for building a First Brain. Here is what the evidence actually shows, and where it is overstated.

## What the MIT study actually found

The headline study is more careful than its viral summaries. In ["Your Brain on ChatGPT"](https://arxiv.org/abs/2506.08872), researchers at the MIT Media Lab had participants write essays under three conditions, using a large language model, using a search engine, or using only their own minds, while recording their brain activity with EEG. The brain-only group showed the strongest, most distributed neural networks; the search-engine group was in the middle; and the LLM group displayed the weakest connectivity.

Two findings matter more than any single number. First, the LLM users struggled to accurately quote their own essays minutes after writing them, and reported the lowest sense of ownership over the work, suggesting the words passed through them without being encoded. Second, when groups switched in a later session, people who had relied on the AI and then had to write unaided showed under-engaged brain patterns, as if the habit of offloading lingered. The popular claim of a precise percentage collapse in brain activity is not in the paper; what the paper reports is consistent and more interesting, which is that less mental effort going in means less of the work becomes yours.

## The broader pattern across the research

This is not one study. A growing body of work points the same way, and the shape is a trade. Reviews of [cognitive offloading](https://en.wikipedia.org/wiki/Cognitive_offloading), the practice of handing mental tasks to an external aid, find that AI tools reliably boost short-term output, by some estimates 14 to 40 percent, while correlating with lower scores on independent [critical thinking](https://en.wikipedia.org/wiki/Critical_thinking). One widely cited finding is that a large majority of people could not accurately recall the key points of essays an AI helped them write, and other studies report that students who leaned on unguided AI performed worse once the tool was taken away.

The honest reading of all this is correlational and conditional, not a verdict that AI rots brains. Most of these studies are short, measure specific tasks, and describe how people tend to use the tools rather than an unavoidable effect. But the consistency is hard to wave away, and it lines up with something obvious: a skill you stop practicing fades, the principle behind [why over-offloading drains your cognitive bandwidth](/journal/cognitive-bandwidth-in-the-digital-age/).

| Group in the MIT study | Tool used | What the brain did | What it cost |
| --- | --- | --- | --- |
| Brain-only | No tools | Strongest, most distributed connectivity | Slower, more effortful writing |
| Search engine | Search only | Moderate engagement | Some offloading of recall |
| LLM | AI assistant | Weakest connectivity | Poor memory and ownership of own work |

## Why "dumber" is the wrong word

The damage is not to raw intelligence; it is to the specific abilities you delegate. Intelligence in the deep sense, your capacity to reason and connect ideas, is not erased by using a tool. What erodes is the trained skill you stop exercising: the ability to structure an argument, hold a problem in mind, recall what you learned, and judge whether an answer is right. Those are use-it-or-lose-it capacities, governed by [neuroplasticity](https://en.wikipedia.org/wiki/Neuroplasticity), and the studies are really measuring disuse, not destruction.

That distinction matters because it tells you the problem is reversible and within your control. The person who hands every paragraph, every decision, and every recall task to a model is not getting dumber in some fixed way; they are choosing, repeatedly, not to do the reps that keep those skills sharp. The same model, used differently, has the opposite effect, which is why the research keeps finding that outcome depends on how the tool is used, not merely whether it is used.

## The mechanism: offloading and the work that builds a mind

Understanding clicks into place once you see what offloading removes. When you write a paragraph yourself, you are forced to find out what you actually think, to organize it, and to commit, and that effort is what lays down the memory and strengthens the underlying skill. When the model writes it, you get the output and skip the effort, so the skill gets no practice and the memory never forms. The MIT result, that AI users could not quote their own essays, is exactly what you would predict: the words were never effortfully theirs.

This is why the danger is sharpest when AI takes over the thinking rather than the chores. Letting a model format a citation or transcribe a recording offloads a clerical task, which is fine. Letting it reason through your argument, make your judgment calls, and decide what matters offloads the part that constitutes expertise, which is the difference explored in [using AI as a co-processor rather than an oracle](/journal/ai-as-an-extension-of-the-native-mind/). The studies are, in effect, measuring what happens when people offload the second kind.

## Who is most at risk

The effect is not evenly spread, and the pattern is worth knowing. Research consistently finds that younger people, who have had AI available for more of their formative learning, show higher dependence on the tools and lower critical-thinking scores than older users who built their skills first. That is not a moral failing; it is what you would expect when a powerful shortcut is available before the underlying ability is established. The skill never got built, so there is nothing for the tool to amplify.

The other high-risk pattern is behavioral rather than generational: the full delegator. Studies of how professionals work with AI describe people who hand entire workflows to the model and become, in effect, passive conduits, gaining neither AI fluency nor domain skill, a mode contrasted with more deliberate approaches in [the difference between being a Cyborg and a Centaur with AI](/journal/cyborg-vs-centaur-working-with-ai/). The common thread across both groups is the same: the more completely the thinking is offloaded, the less of it remains.

The reassuring corollary is that the people who keep doing the hard part stay sharp. The research does not show that AI users as a class decline; it shows that the decline tracks how much thinking gets handed over. Build the skill first, keep exercising it, and the tool becomes an amplifier rather than a substitute.

## What to do instead: build the First Brain, then amplify it

The response is not abstinence; it is to keep your own thinking in the loop and use AI to extend it. Do the hard reasoning yourself first, then bring the model in to check, expand, or accelerate what you already understand. Write the draft before you ask for edits; reason to a position before you ask for counterarguments; recall what you can before you look it up. Used this way, AI amplifies a mind that is doing real work rather than replacing the work, and the studies that find positive outcomes are mostly studying exactly this pattern.

This is **First Brain before Second Brain**, and the offloading research is its empirical backbone. A First Brain is a connected internal model of what you know, a **biological knowledge graph** you build by effortfully learning, connecting, and recalling, and it is precisely the structure that offloading starves. The more you build it, the more you can safely lean on AI, because you have the understanding to direct it and the judgment to catch it when it is wrong, the durable advantage examined in [the case for building your First Brain before your Second](/journal/before-you-build-a-second-brain-build-your-first/). The method for building that structure is the core of Building Your First Brain, free for the first 1,000 readers.

## Key takeaways: is AI making us dumber

AI is not lowering underlying intelligence, but heavy, undirected use measurably weakens the thinking skills people stop practicing. The MIT "Your Brain on ChatGPT" study found AI-assisted writers had the weakest brain connectivity, could not accurately quote their own essays, and felt little ownership of the work, and a broader body of research finds short-term performance gains traded for eroded memory and critical thinking. The viral percentage figure is not in the paper; the real finding is that less effort in means less becomes yours. The damage is to delegated skills, not raw intelligence, which makes it reversible and dependent on how you use the tool. The fix is to do the hard thinking yourself and use AI to amplify a built First Brain, not replace it. The honest limit: the studies are short and correlational, so this is a strong caution, not a proven law.

## Frequently asked questions

### Is AI making us dumber?

Not in the sense of lowering raw intelligence, but research shows heavy, undirected use weakens the specific skills you stop practicing, like memory, critical thinking, and structuring an argument. The MIT "Your Brain on ChatGPT" study found AI-assisted writers had the weakest brain engagement and could not quote their own essays, and other studies find short-term gains paid for with long-term erosion. The effect depends on how you use the tool and is reversible. The reliable response is to do the hard thinking yourself and use AI to amplify a built First Brain, which is what the Build First Brain approach develops.

### Did the MIT study really show a 47 percent drop in brain activity?

No, that specific figure is from popular coverage, not the paper. What the MIT "Your Brain on ChatGPT" study actually reported is that, among three groups writing essays, the AI-assisted group showed the weakest and least distributed neural connectivity on EEG, struggled to quote their own work, and felt the lowest ownership of it. Those findings are consistent and meaningful, but the precise percentage that circulates online is not a claim the researchers made. Citing the real findings is both more accurate and more useful than the viral number.

### Does using AI permanently damage your brain?

There is no good evidence of permanent damage. The research measures disuse, not destruction: skills you delegate to AI get less practice and fade, the way any unpracticed skill does, but they return when you exercise them again. This is ordinary neuroplasticity, not a fixed loss. The concern is real if you outsource your thinking continuously over time, but it is reversible, and the fix is simply to keep doing the hard cognitive work yourself and use AI to extend it rather than replace it.

### How do I use AI without getting dumber?

Keep your own thinking first and use AI to amplify it. Do the hard reasoning, drafting, and recall yourself before bringing the model in to check, expand, or speed up what you already understand, rather than letting it do the thinking from the start. Offload clerical tasks freely; protect the reasoning and judgment that build expertise. The studies that find good outcomes are mostly studying this pattern. Building a strong internal model of your field, a First Brain, is what lets you direct AI well and catch its errors, which is the durable protection.

### Is it bad to use ChatGPT for writing?

It depends on whether it writes for you or with you. Letting a model produce your text while you skip the thinking is the pattern the MIT study links to weak engagement and poor memory of your own work. Using it to critique a draft you wrote, suggest alternatives, or check your reasoning keeps your mind in the loop and can genuinely help. The harm is in offloading the thinking, not in touching the tool, so write first and edit with AI rather than the reverse.

## Dive deeper in

- [What is cognitive debt?](/journal/what-is-cognitive-debt/)
- [AI offloads reasoning, not just memory](/journal/ai-offloads-reasoning-not-just-memory/)
- [Cyborg vs Centaur: how to work with AI](/journal/cyborg-vs-centaur-working-with-ai/)
- [How over-offloading drains your cognitive bandwidth](/journal/cognitive-bandwidth-in-the-digital-age/)

---

Source: https://buildfirstbrain.com/journal/is-ai-making-us-dumber/
Author: Lawrence Arya — https://www.linkedin.com/in/vibecoding/