83% Couldn't Quote Their Own Essay

Over the weekend I was at dinner with a friend's family and the conversation landed somewhere uncomfortable. Someone compared AI adoption to pharma companies releasing drugs and running the long-term studies afterward, except this time the trial includes pretty much everyone with an internet connection. We are all in the experiment, and we are only starting to understand what we traded for the convenience.

A study by MIT tracked 54 people writing essays under different conditions. One group used ChatGPT. Another used Google search. The third relied only on their own thinking, no external tools. The researchers did not just compare the essays. They measured what was happening inside the participants' heads while they wrote.

What they found was a clear hierarchy of neural engagement, and I did not love where I would have landed on it.

What your brain looks like on AI

People writing without external tools showed the strongest, most widespread neural connectivity. Their brains were firing across regions tied to memory formation, creative thinking and critical analysis. The search engine group showed moderate activity, still doing real cognitive work, scanning results, evaluating sources, synthesising.

The LLM group's connectivity was significantly weaker across the board. Worth saying clearly: this was 54 participants from a specific geographic region, so we need larger and more diverse studies to confirm the pattern. But the early signal is hard to ignore once you have seen it. The brain is doing less, the way a muscle you stopped using is technically still there and quietly forgetting its job.

The memory problem

After writing, participants were asked to quote from what they had just written, literally minutes before. The brain-only group reached nearly perfect recall by session two. The search engine group did well with minor hiccups. The LLM group: 83% could not quote their own work in the first session. Even after multiple practice sessions, many still struggled with this basic task.

They had just finished writing essays and could not remember what they wrote, because they never really processed it. They had become sophisticated copy-paste machines. It is the difference between someone who learned to drive in a manual and can handle any vehicle versus someone who only ever drove an automatic and stalls the first time they touch a stick.

The ownership crisis

When researchers asked how much participants "owned" their essays, the brain-only group nearly unanimously said it was completely their work. The search engine group claimed 70 to 90% ownership, mostly mine with some external input. The LLM group fragmented, ranging from full denial to partial credit. Some explicitly said the essay was not really theirs. Others claimed partial ownership but seemed unsure. This is not an intellectual honesty problem. It is a disconnect between effort and outcome that we are only starting to notice.

When the training wheels come off

The most revealing part came in a fourth session, where they flipped the script. People who had been using LLMs were asked to write without them. People who had been writing independently got access to LLMs.

The former LLM users without AI: 78% could not quote anything from their new essays, only 11% produced correct quotes, neural connectivity dropped dramatically, and they appeared to have lost fundamental writing and memory skills. The former brain-only users with AI showed increased brain activity across all frequency bands, used the AI strategically and thoughtfully, kept strong memory and ownership, and used more sophisticated prompts, suggesting better AI literacy.

The sequence mattered. Build the capability first, then add the tool, and the tool makes you sharper. Start with the tool and the capability never forms.

This lines up with something I have been doing for months, a 70/30 approach, letting AI handle 70% of execution while I stay actively engaged in the other 30%. When I was learning to deploy production-level web apps, I would map what I knew versus what I had no clue about, then use AI for explanations and analogies rather than just copying output. But this research makes me think I should flip the ratio when learning is the goal. Maybe 60/40 or even 50/50 for tasks where I am genuinely trying to develop a new capability. The real value of AI as a tutor is that it adapts to you and gives immediate feedback. Most people sprint through the learning part to reach the output and miss exactly that.

Again, limited sample, needs validation with larger and more diverse groups. But the pattern points at something real about the order in which you learn.

The homogenization effect

The essays written with LLMs were remarkably similar within each topic. Not just in structure, but in specific ideas, arguments and language. Human teachers, blind to which group wrote what, could consistently pick out the LLM-assisted work. They called it "conventional" and "homogeneous". It hit all the right points and lost the individual perspective that makes human thinking worth reading. The brain-only group showed "strong variability", each person bringing their own angle.

What this means for working and learning

For learning, start with unassisted effort to build the foundation, then bring AI in for enhancement. Treat it as a sophisticated tutor that adapts to you, while you still do the actual learning. For professional work, use AI to augment your expertise rather than replace your thinking, keep ownership of your ideas and reasoning, and remember that everyone has the same tools so your own thinking is the differentiator. For long-term development, regularly work without AI to keep mentally fit. The researchers flagged this matters most for junior professionals, who might miss building crucial mental models if they learn in an AI-heavy environment from the start.

The part still being figured out

Every time we let an LLM handle something we could do ourselves, we make a trade. Sometimes it is worth it. We should at least be conscious of what we are trading. The participants who kept their brains fully engaged did not just write better essays. They remembered them, owned them, could build on them, and their neural networks grew more complex. They became more capable, not less.

The convenience is seductive, and it should be, because these tools are remarkable collaborators when used well. The researchers called for longitudinal studies with larger and more diverse populations before any definitive conclusions, which is the honest caveat and not a reason to wait.

You do not have to wait for the longitudinal study. This week, pick one task you currently hand entirely to AI and do the first draft yourself before the AI touches it. The differentiator in a world where everyone has the same tools is the one thing that cannot be automated, and it only forms through the unglamorous work of actually thinking.