What is the "brain cost" of using ChatGPT?

About the accumulation of cognitive debt when using an LLM.

If you wonder what happens to your brain when you use an LLM to write an essay, check this insightful paper from the MIT Media Lab: “Your brain on ChatGPT” (link in the first comment below). It is potentially the first protocol of its kind, which makes it even more interesting.

When using an LLM for writing a text, did you ever feel a little bit “lazy” or at least not stretching your brain to its full capacity? And therefore, you might have started worrying about the long-term impact of doing so.

Well, this study, which “focuses on finding out the cognitive cost of using an LLM in the educational context of writing an essay” is in essence confirming your fears. No surprise here. Delegating the stretching of your brain to a machine, is like stopping to exercise, stopping to stretch your muscles, at some point they atrophy.

The study created 3 research groups: “LLM group, Search Engine group, and Brain-only group, where each participant used a designated tool (or no tool in the latter) to write an essay”. And while the participants were writing their essay, they were monitored using an EEG helmet.

“We used electroencephalography (EEG) to record participants' brain activity in order to assess their cognitive engagement and cognitive load, and to gain a deeper understanding of neural activations during the essay writing task.”

An EEG offers a cost-efficient way to “look into” people’s brain with no harm. It is a non-invasive technique, which, in contrary to putting people in a functional magnetic resonance imaging (fMRI), enables groups of people to remain in action (instead of laying down in the machine), at the fraction of the cost.

So what were the exact findings of the research? “robust evidence that the groups had significantly different (...) cognitive strategies.” Your brain is activated in different ways depending which support mechanism (or none) you use.

“Brain connectivity systematically scaled down with the amount of external support: the Brain‑only group exhibited the strongest, widest‑ranging networks, while the LLM assistance elicited the weakest overall coupling.”

The least brain connectivity, the least the number of neurons firing together and the least the number of new synapses being created or reinforced. Additionally, “The reported ownership of LLM group's essays in the interviews was low (…)”.

So far this sounds quite intuitive. Though what are the consequences long term, of relying more and more on LLMs for this type of work? The study offers that this might lead to “a possible decrease in learning skills“.

While those findings are preliminary, were produced in a specific context, and need to be further validated, they have to be taken seriously for the health of our brain and the ones of our children’s brain. 

Thanks Markus Voelker for sharing this article!