When MIT is Also Beating Dead AI Horses

My quick and unplanned "react" article to a new study on AI in education that will probably make a splash.

There are few institutions with the scientific cachet of MIT. So when a team from the MIT Media Lab publishes a study called “Your Brain on ChatGPT”, complete with EEG brain scans and a host of researchers, you pay attention. The study sets out to measure the cognitive cost of using an AI assistant for an essay writing task. Their conclusion, after wiring up 54 university students and analyzing their brain connectivity, is that using an LLM results in a measurable decrease in cognitive engagement.

On the surface, it’s a worrisome and important finding. But as I dug into the methodology, my initial interest curdled into a familiar frustration. For all its impressive scientific apparatus, the study arrives at a conclusion that was not only entirely predictable, but profoundly uninteresting.

The reason is simple: the task itself was practically designed to produce this exact result. Participants were given a choice of standard SAT-style essay prompts and a mere 20 minutes to write about them. Think about that. A timed, formulaic, low-stakes writing assignment on a topic you have no personal connection to. This isn’t a task that inspires deep thought; it’s a chore. It’s precisely the kind of cognitive grunt work that tools like ChatGPT are marketed to eliminate. Of course the students who were handed a powerful cognitive offloading tool used it to offload the cognition. What else would we expect?

Figure 6, page 31. Percentage of participants within each group who struggled to quote anything from their essays.

In my most successful article to date: The Good, the Bad, and the Ugly Science of AI in Education, I called this kind of study the “dead horse studies”, and they represent a frightening portion of the literature out there. In it I showed that there’s a widespread failure in the research community to integrate AI into studies in a way that is meaningful or thoughtful. This MIT paper, for all its technical brilliance, feels like another example. An expensive, high-tech setup was deployed to confirm a hypothesis that was obvious from the outset because of the pedagogical weakness of the task.

This brings me to the core of the issue, which I laid out in Students Are Using AI, Now What? For any student assignment involving AI to be effective, I argue that it must be built on four pillars: it must be designed to ensure cognitive investment, it must have assessment validity, meaning it actually measures the student’s skills and not the AI’s. It must also enhance, or at least reinforce, the student's AI literacy, and finally, it should ideally foster human-to-human interaction, not isolate the student behind a screen.

This is mine, it’s not evidence based, more common sense based IMO.

The task in the MIT study gets a disastrous zero on all four counts. It doesn’t demand cognitive investment; it invites cognitive laziness. It doesn’t teach students how to use AI more effectively. It’s a solitary activity devoid of interaction. And it fails to assess any meaningful skill in the experimental group, a fact proven by their inability to remember or quote from the essays they had just written.

This is what makes the study so frustrating. Imagine what we could have learned if that impressive EEG setup had been applied to a well-designed task. In The Personalization Paradox, I made the case, echoing an amazing Isaac Asimov interview, that cognitive investment is guaranteed when you use technology to pursue goals you’ve set for yourself. What if the study had been built on that principle?

What if participants had been given a full hour—or more! to write a chapter of their autobiography? Or tasked with learning a new coding language of their choice? Or asked to research and plan a project they were genuinely passionate about? No human on the planet wakes up wanting to write an SAT-style essay in 20 minutes for fun. But when the task is personal and meaningful, the drive to learn and create is intrinsic. That’s a condition where—I think—AI could become a truly powerful partner, and the resulting brain activity would be fascinating to see. Would we still see cognitive offloading, or… something else?

That’s the kind of research we desperately need. For once, we have brave scientists not scared of publishing worrisome results about AI, and that should be applauded. But these results were a foregone conclusion. Their expensive, sophisticated experiment could have yielded far more interesting insights with a few simple tweaks to the design. We need more creative and pedagogically sound uses of AI in these studies to give educators and students real ideas about what to do with these powerful new tools.

There are people out there who are coming up with good ideas, if you’re one of them or you know someone, let me know! As far as I’m concerned, you’ll have to wait till September before I put forward a few ideas. In the meantime, and if you’re interested, you should subscribe to Mike Kentz and sit tight. Tschüss!