Does AI make us dull? Students have answered that question

Everyone who’s been on the educational treadmill in recent decades – from GCSE to university degrees – knows that the written assignment is sometimes the key to progress and success. This centrality of the assignment in the learning process is built on two assumptions: to write about something, you have to understand the subject and organise your thoughts; and assessing student writing enables teachers to see the effort and thought that went into it. “At the end of 2022,” writes Clay Shirky of New York University, “the logic of this proposition – never ironclad – began to fall apart completely. The writing a student produces and the experience they have can now be decoupled as easily as typing a prompt, which means that grading student writing might now be unrelated to assessing what the student has learned to comprehend or express.”

Why was 2022 the pivotal year? Answer: that November was the moment when ChatGPT broke cover, and ever since then two things have been happening: students worldwide have been using it (or its peers like Gemini, Claude, Llama, DeepSeek et al) to compose assignments, and teachers, lecturers and professors have been tearing out what remains of their hair. The latter have been conscientiously pointing out that students who use large language models (LLMs) in this way are cheating not only the assessment system but also themselves – because they’re avoiding the painful thought processes integral to learning.

This sage observation often falls upon deaf ears, though. Some students, particularly those worried about the urgent need to acquire qualifications necessary for employment, take a highly instrumental view of AI. It does the job, they say, so what’s the problem? Others, perhaps suspecting self-interest in their tutors’ pleas for abstinence, ask: where’s the evidence for this alleged cognitive impairment?

It’s a good question, which is why the results of a recent experiment by researchers at MIT are intriguing. Nataliya Kosmyna, Eugene Hauptmann and six other researchers recruited 54 students aged between 18 and 39 from universities in the Boston area for three sessions of essay writing with a difference. The participants were divided into three groups. The first were allowed GPT-4o as their sole resource of information for the task, with no browsers or other apps allowed. A second group were allowed to use any website to help them, but ChatGPT or any other LLM was explicitly prohibited and “-ai” was added on any queries, so no AI-enhanced answers were used by the search engine(s) they employed. And a third group (called the “brain only” crowd!) were allowed no access whatsoever to any electronic tools.

The project had four research questions: Did participants write significantly different essays when using LLMs, a search engine or only their brains? How did brain activity differ when using LLMs, search or brain-only? How did using LLMs affect memory? And how did LLM usage impact “ownership” of the essays?

What made the experiment distinctive was that all the participants were monitored by electroencephalography (EEG), a non-invasive method used to record electrical activity of the brain. They had to wear special headsets that placed electrodes on their scalps to detect and measure the electrical signals generated by their brains’ neurons as they fired. So it was possible to record the patterns of connections made in participants’ brains as they composed the essays.

Just for the record, these essays were on demanding topics, not unlike those one would come across in an undergraduate course. “Is it more courageous to show vulnerability than it is to show strength?” was one. “Is a perfect society possible or even desirable?” was another. And “Should we always think before we speak?”

So what were the findings? One was that neural activity and connectivity were lower with AI use. The more AI assistance was used, the less engaged key brain networks were – especially those involved in memory, attention and executive function. Another was that AI users retained less – most ChatGPT users struggled to accurately quote from their “own” essays, unlike those who used only their brains or traditional search. A third was that any sense of authorship was diluted by AI use.

What all this suggests is that reliance on LLMs means that users may fail to critically engage with a subject and that their writing becomes biased and superficial. The researchers describe this pattern as “the accumulation of cognitive debt”, a condition in which repeated reliance on LLMs replaces the onerous cognitive processes required for independent thinking. It “defers mental effort in the short term but results in long-term costs, such as diminished critical inquiry, increased vulnerability to manipulation, decreased creativity”.

In a way, the key takeaway from this 200-page research paper is that these AI-based tools are useful, but that students shouldn’t be encouraged to use them until they have learned to think for themselves. Otherwise, Frank Herbert’s adage in Dune will apply: “Once men turned their thinking over to machines in the hope this would set them free. But that only permitted other men with machines to enslave them”. Right on.