Is AI making us dumber? The “55% connectivity drop” says yes

Open your recent chats with AI. Now, try to quote a single sentence you “wrote” ten minutes ago. If you can’t, congratulation, you’re part of the 83%.

Research shows that using ChatGPT doesn’t just assist your brain, it creates a “cognitive debt” that slashes neural connectivity by 55% and flattens your unique perspective into a statistical average.

To be clear, EEG connectivity is obviously not a direct measure of intelligence or reasoning ability. What it captures is engagement across cognitive networks during task execution, and the magnitude of the drop is what makes the finding difficult to ignore.

Here is the mathematical proof that large-scale outsourcing of cognition produces convergence, and convergence is the opposite of thinking.

We were told that AI augments human capability. That it’s a bicycle for the mind. That more access to powerful language models means more ideas, more creativity, more cognitive horsepower for everyone.

But…is it? What if the technology we’re using to think is quietly, systematically flattening how we think, and we don’t notice?

I spent three entire weeks digging through the emerging research, and what I found is unsettling. Not because any single study is ultimately damning, but because the evidence travels from so many different directions (mathematics, neuroscience, linguistics, cross-cultural psychology) all pointing toward the same structural conclusion.

This isn’t a suspicion anymore. It’s an emergent property of how these systems are built.

The Mathematical Certainty of Collapse

Let’s start with the hardest evidence: proof.

In 2024, researchers published a paper in Nature demonstrating something that should concern anyone paying attention. Shumailov et al. showed that when generative AI models are trained recursively on their own outputs (even mixed with real data) they inevitably lose information about the true underlying distribution. The technical term is “model collapse,” but what it means in practice is this:

The weird, unusual, minority perspectives disappear first. Then the variance narrows. Eventually, you’re left with what mathematicians call a “delta function”, a single point with zero variance.

The tail disappears before the center. The outliers go first.

“Once absorbed into a delta-function state, the information loss is permanent,” the researchers noted. “No amount of temperature tuning or sampling strategies can recover lost distributional tails.”

Humans, of course, are not language models. But when humans repeatedly accept suggestions generated by systems that are mathematically guaranteed to converge, the human is no longer the generator: they become the fine-tuning loop.

This isn’t even an isolated speculation. It’s a proof that applies to any generative model trained on its own outputs.

And here’s the uncomfortable reality: synthetic data is already everywhere. It’s scraped from the web where LLM-generated text has become ubiquitous. It’s contaminating training corpora right now.

The RLHF Problem Nobody Talks About

So why do some AI tools flatten diversity more than others?

The most rigorous answer comes from a controlled experiment by Padmakumar & He, published at ICLR 2024. They recruited 38 professional writers from Upwork and randomly assigned them to three conditions: writing solo, writing with GPT-3 (base model), or writing with InstructGPT (the RLHF-tuned model that powers ChatGPT).

Each participant wrote 100 argumentative essays while keystroke-logging software tracked every character’s authorship.

The finding that stopped me cold: InstructGPT reduced lexical diversity. GPT-3 did not.

Despite users incorporating similar amounts of model-generated text (about 32-35%), the instruction-tuned model produced measurably more homogenized outputs. Five-gram repetition increased. Pairwise similarity between different writers’ essays increased. The statistical significance held across multiple metrics.

Why? Reinforcement Learning from Human Feedback—the technique that makes ChatGPT “helpful”—trains models to maximize a reward function learned from human preferences. And human annotators tend to prefer clear, concise, grammatically correct, on-topic text over colorful, unconventional, or stylistically distinctive writing.

“The model learns to avoid risky, unconventional, or stylistically distinctive outputs in favor of safe, high-reward text,” the researchers explained.

This is the diversity–quality trade-off nobody in AI marketing wants to discuss: the system is not optimizing for thought, but for approval.

The same technique that makes these tools feel useful is the one that makes them cognitive homogenizers.

Your Brain When Chatting With AI

At this point, you might be thinking: “Sure, the text is more similar. But that’s just words on a page. It doesn’t change how people actually think.”

The neuroscience disagrees.

A controlled study tracked brain activity across three groups writing essays: no assistance, search engine assistance, and ChatGPT assistance. The EEG findings were stark:

- Up to 55% reduced neural connectivity in the ChatGPT group compared to unassisted writers

- Decreased frontal theta activity, indicating reduced working memory engagement

- Weakest overall neural coupling across memory and decision-making networks

But the most troubling finding was this: 83% of LLM users were unable to quote from essays they had just written.

Not essays from last week. Essays they had finished moments before.

The researchers described this as “cognitive debt”, users offload reasoning to the system, resulting in weaker memory consolidation and reduced sense of authorship.

You’re not enhancing your thinking. You’re outsourcing it. And your brain is adapting accordingly.

The Effect That Doesn’t Go Away

But it gets worse.

Moon et al. (2025) documented what they called a “continuing effect.” When ChatGPT became unavailable after five days of use, participants “continued producing homogenized content”. Diversity growth rates remained suppressed—approximately 2× lower than pre-ChatGPT baselines.

This suggests that LLM interaction doesn’t just assist cognition: it trains it. And it trains it toward statistical normality. The homogenization isn’t limited to moments of direct AI use. It persists as a learned behavioral pattern.

You’re not just using a tool. You’re being trained by it.

The WEIRD Alignment Problem

The Stanford-NYU team tested 27 LLMs across 155 topics covering 12 countries. They introduced a metric called “epistemic diversity”, the diversity of real-world claims about a topic in LLM outputs.

What they found: for country-specific topics, LLMs reflected English-language knowledge far more than local-language knowledge, even when queried about non-Western countries. Western knowledge dominated in 5 out of 8 countries tested.

As counter-intuitive as it may look, “larger models produced less diverse outputs”. Scaling improves accuracy on common queries but at the cost of tail knowledge.

A separate cross-cultural experiment by Agarwal, Naaman, & Vashistha tested 118 participants from India and the USA writing with AI suggestions. Indian participants adopted Western writing styles, shorter sentences, less formal constructions, when receiving AI suggestions from the Western-trained model.

“AI suggestions altered not just what is written but also how it is written,” the researchers noted, “reducing the uniqueness of Indian participants’ expression”.

The mechanism isn’t censorship. It’s the mundane operation of statistical optimization toward Western-dominant training data. The model was trained on Western English text. It encodes Western norms as “good writing.” Non-Western users who accept suggestions are nudged toward this norm, gradually, one autocomplete at a time.

What looks like cultural bias is actually something more mechanical: when diversity lives in the tails, optimization erases it first.

The Only Thing That Works

I went looking for solutions. What I found was sobering.

Temperature scaling? Can’t recover lost distributional tails once they’re gone.

Prompt engineering with identity-coded personas? Often produces stereotypes rather than authentic within-group diversity.

Debate-based frameworks? Captures the “socially correct mean” rather than genuine pluralism.

Retrieval-augmented generation (RAG)? Helps, but benefits are uneven, countries with more diverse English-language web sources benefit more, while underrepresented groups see minimal improvement.

The only intervention with demonstrated success: hybrid human-AI systems that preserve human cognitive diversity alongside LLM accuracy.

Kraft et al. (2025, accepted at IJCAI) found that LLM-only crowds exhibited diminishing returns as group size increased due to limited diversity, whereas human crowds showed steady performance gains.

Hybrid crowds outperformed both.

The implication is uncomfortable but important: LLMs don’t enhance collective intelligence. They replace it with statistical convergence. The value isn’t in the AI alone, it’s in the human variance the AI can’t replicate.

One Rule That Preserves Cognition

Never let AI generate the first draft of original thought.

Use AI after you have already externalized your reasoning (notes, outlines, ugly drafts). At that point, the model is operating on your distribution, not replacing it.

In every study where homogenization dominates, the common factor is sequence: model first, human second. Reverse the order, and you preserve variance while still benefiting from compression, recall, and synthesis.

Dongu Said

I want to be careful here. This isn’t an argument against using AI tools. I use them. A lot. You probably use them. They’re genuinely useful for certain tasks.

But the evidence suggests we need to gain awareness of what we’re trading away.

The homogenization isn’t a correctable flaw. It’s structural, emerging from training objectives that optimize for high-probability outputs, reinforced learning that encodes consensus preferences, and recursive feedback loops that amplify convergence across generations.

I’m not suggesting that we should be quitting AI, that wouldn’t be wise either, but we should make a willing effort to be brave enough and keep things difficult.

We have to fight for the 'messy' human parts of our process, the slow and inefficient bits where original ideas actually live.

If we let AI smooth everything over for the sake of speed, we lose the rare insights that make us valuable.

And according to the math, once variance collapses, no amount of prompting can resurrect what was never reinforced.