Smarter, not harder
MIT researchers recently discovered that ChatGPT users showed dramatically reduced brain activity compared to unassisted writers, which of course was the news outlets' cue to pounce on the narrative, "AI is making us dumber". I'm not convinced that's what the data is telling us, though.
The setup matters here. Short, twenty minute, formulaic SAT essays, participants encountering ChatGPT for the first time, restricted solely to that tool. What the researchers caught wasn't some slow cognitive decline, it was novices handing the job straight over to a powerful tool. These weren't necessarily aspiring writers chasing mastery either, they were paid participants getting through a narrow task. The study design all but guaranteed shallow use of the tools.
The bit that bothers me is that neural activity is being treated as the same thing as cognitive engagement, and I don't think it is.
Neuroscience has understood for a long time that expertise shows up as neural efficiency rather than excess. Chess grandmasters show less brain activity than novices during play. Professional pianists exhibit minimal motor cortex activity where beginners show scattered activation. A lot of activation quite often means someone is struggling and adapting, not that they're any good.
The study's own citations tell a more mixed story as well. Yang et al. found that "higher-competence learners engaged more deeply with content through reading-intensive behaviours" when using AI. Strategic users stayed engaged whilst leaning on the tool. They knew what to offload and what to hang on to. The MIT participants had no reason to reach that level, and no route to becoming high-competence learners either.
The worrying patterns in the MIT study, the fragmented sense of ownership, people unable to quote their own work back, look to me like the shallow engagement of first contact and low incentives. Hand someone a cognitive tool with no grounding and no reason to care, and they'll usually let it do the work. Even so, how quickly it happened says something about how readily we hand over mental effort when a tool turns up.
Calculators freed engineers from basic arithmetic, high level languages freed programmers to tackle harder algorithmic problems, and I'd say AI tools are promising much the same, freeing us from the mechanical work that stops us reaching the harder problems rather than from thinking itself. So the question for me isn't whether we'll think less, it's whether we'll get to solve more ambitious problems.
The study has caught one end of a spectrum that runs from mental offloading to mental efficiency. Yang's strategic users and the chess grandmasters sit at the other end, where less neural activity can be a sign of someone who has adapted rather than given up. The difference, as far as I can tell, is whether you know what you're delegating and why.
Of course, the study's limitations don't invalidate its findings. What they show is that we're at the messy start of working alongside these tools, and we've barely begun to see what that could look like. A crude measurement of neural activity as a proxy for intelligence doesn't mean we're all going to, or indeed should have to, surrender our mental faculties to AI tools as we use them. Less brain activity might just mean you've stopped doing the arithmetic by hand.