mischiefmanifold

[mischiefmanifold]

There’s a new (unreviewed draft of a) scientific article out, examining the relationship between Large Language Model (LLM) use and brain functionality, which many reporters are incorrectly claiming shows proof that ChatGPT is damaging people’s brains.

As an educator and writer, I am concerned by the growing popularity of so-called AI writing programs like ChatGPT, Claude, and Google Gemini, which when used injudiciously can take all of the struggle and reward out of writing, and lead to carefully written work becoming undervalued. But as a psychologist and lifelong skeptic, I am forever dismayed by sloppy, sensationalistic reporting on neuroscience, and how eager the public is to believe any claim that sounds scary or comes paired with a grainy image of a brain scan.

So I wanted to take a moment today to unpack exactly what the study authors did, what they actually found, and what the results of their work might mean for anyone concerned about the rise of AI — or the ongoing problem of irresponsible science reporting.

If you don’t have time for 4,000 lovingly crafted words, here’s the tl;dr.

The major caveats with this study are:

This paper has not been peer-reviewed, which is generally seen as an essential part of ensuring research quality in academia.

The researchers chose to get this paper into the public eye as quickly as possible because they are concerned about the use of LLMs, so their biases & professional motivations ought to be taken into account.

Its subject pool is incredibly small (N=54 total).

Subjects had no reason to care about the quality of the essays they wrote, so it’s hardly surprising the ones who were allowed to use AI tools didn’t try.

EEG scans only monitored brain function while writing the essays, not subjects’ overall cognitive abilities, or effort at tasks they actually cared about.

Google users were also found to utilize fewer cognitive resources and engage in less memory retrieval while writing their essays in this study, but nobody seems to hand-wring about search engines being used to augment writing anymore.

Cognitive ability & motivation were not measured in this study.

Changes in cognitive ability & motivation over time were not measured.

This was a laboratory study that cannot tell us how individuals actually use LLMs in their daily life, what the long-term effects of LLM use are, and if there are any differences in those who choose to use LLMs frequently and those who do not.

The researchers themselves used an AI model to analyze their data, so staunch anti-AI users don’t have support for there views here.

Brain-imaging research is seductive and authoritative-seeming to the public, making it more likely to get picked up (and misrepresented) by reporters.

Educators have multiple reasons to feel professionally and emotionally threatened by widespread LLM use, which influences the studies we design and the conclusions that we draw on the subject.

Students have very little reason to care about writing well right now, given the state of higher ed; if we want that to change, we have to reward slow, painstaking effort.

The stories we tell about our abilities matter. When individuals falsely believe they are “brain damaged” by using a technological tool, they will expect less of themselves and find it harder to adapt.

Head author Nataliya Kosmyna and her colleagues at the MIT Media Lab set out to study how the use of large language models (LLMs) like ChatGPT affects students’ critical engagement with writing tasks, using electroencephalogram scans to monitor their brains’ electrical activity as they were writing. They also evaluated the quality of participants’ papers on several dimensions, and questioned them after the fact about what they remembered of their essays.

Each of the study’s 54 research subjects were brought in for four separate writing sessions over a period of four months. It was only during these writing tasks that students’ brain activity was monitored.

Prior research has shown that when individuals rely upon an LLM to complete a cognitively demanding task, they devote fewer of their own cognitive resources to that task, and use less critical thinking in their approach to that task. Researchers call this process of handing over the burden of intellectually demanding activities to a large language model cognitive offloading, and there is a concern voiced frequently in the literature that repeated cognitive offloading could diminish a person’s actual cognitive abilities over time or create AI dependence.

Now, there is a big difference between deciding not to work very hard on an activity because technology has streamlined it, and actually losing the ability to engage in deeper thought, particularly since the tasks that people tend to offload to LLMs are repetitive, tedious, or unfulfilling ones that they’re required to complete for work and school and don’t otherwise value for themselves. It would be foolhardy to assume that simply because a person uses ChatGPT to summarize an assigned reading for a class that they have lost the ability to read, just as it would be wrong to assume that a person can’t add or subtract because they have used a calculator.

However, it’s unquestionable that LLM use has exploded across college campuses in recent years and rendered a great many introductory writing assignments irrelevant, and that educators are feeling the dread that their profession is no longer seen as important. I have written about this dread before — though I trace it back to government disinvestment in higher education and commodification of university degrees that dates back to Reagan, not to ChatGPT.

College educators have been treated like underpaid quiz-graders and degrees have been sold with very low barriers to completion for decades now, I have argued, and the rise of students submitting ChatGPT-written essays to be graded using ChatGPT-generated rubrics is really just a logical consequence of the profit motive that has already ravaged higher education. But I can’t say any of these longstanding economic developments have been positive for the quality of the education that we professors give out (or that it’s helped students remain motivated in their own learning process), so I do think it is fair that so many academics are concerned that widespread LLM use could lead to some kind of mental atrophy over time.

This study, however, is not evidence that any lasting cognitive atrophy has happened. It would take a far more robust, long-term study design tracking subjects’ cognitive engagement against a variety of tasks that they actually care about in order to test that.

Rather, Kosmyna and colleagues brought their 54 study participants into the lab four separate times, and assigned them SAT-style essays to write, in exchange for a $100 stipend. The study participants did not earn any grade, and having a high-quality essay did not earn them any additional compensation. There was, therefore, very little personal incentive to try very hard at the essay-writing task, beyond whatever the participant already found gratifying about it.

I wrote all about the viral study supposedly linking AI use to cognitive decline, and the problem of irresponsible, fear-mongering science reporting. You can read the full piece for free on my Substack.

[mischiefmanifold]

never underestimate how cathartic it is to be an absolute little weirdo. just a straight up freak

[mischiefmanifold]

misreading the vibes is my specialty

[mischiefmanifold]

i think we should talk about degendering more as a very real form of transphobia. you should not be calling a trans woman "they" when she's explained her pronouns to you. you should not be calling her a "person" instead of a woman. she's not too gnc, she's not too androgynous, you're not "confused" about her identity, you're degendering her. I fear we've gotten to a point we've forgotten the very basics of this movement is "trans women are women" and "trans men are men", and not just "trans people are someone who's pronouns you have to memorize so you don't offend them." you see her as a man in a dress and it pisses me off

[mischiefmanifold]

wait other post cancelled. mutuals. when r ur birthdays.

[mischiefmanifold]

desperately need someone to scream about the dsmp with actually

[mischiefmanifold]

gripping my hair, ripping it out AUGHHHH I LOVE.. MUSIC.. AND SOUNDS

[mischiefmanifold]

being abused or neglected really makes it clear how many things are skills that nobody really treats as. skills. exercising autonomy is a skill. listening to your body is a skill. resting is a skill. being liked and being loved are skills. nobody tells you how to do this shit because nobody even told me I was supposed to have learnt these things when I was a kid. I kind of just have to manually figure out what makes me freak out and work from there. unfair as shit

[mischiefmanifold]

i’ve decided i’m fine, and over it and i miss no one and i never will again

[mischiefmanifold]

I'll never forget when I mentioned offhand at work that I was groomed on a minecraft server at 13 and my coworker asked if dream was there 💀💀💀

[mischiefmanifold]

i need a bone crushing hug, an orgasm that'll make me see god, and $12,000

[mischiefmanifold]

i dont lose interest in favorite characters they just get added to my extra inventory slots

[mischiefmanifold]

Just wait untill my 6 tumblr mutuals hear about this

[mischiefmanifold]

If you haven’t changed your url in years tell me why as someone in your same boat it’s for science

[mischiefmanifold]

don't like auto-flush toilets. like excuse me I'll tell you when to swallow

[mischiefmanifold]

this makes me officially 0 for 5 on relationship success which I feel like warrants an award of some sort

discovered yesterday that I forgot that I dated someone in middle school and we were in the same server bc we were mutual friends with someone we went to HS with and they randomly started trying to get my friend group to drop me as a friend with false claims and hysterics

[mischiefmanifold]

discovered yesterday that I forgot that I dated someone in middle school and we were in the same server bc we were mutual friends with someone we went to HS with and they randomly started trying to get my friend group to drop me as a friend with false claims and hysterics