The Other Cognitive Cost of Convenience
Same Prompt, Same Output: The Homogenization of Human Ideas
One of the more interesting things I ever heard suggested about the potential impacts of an emerging technology came from a writer who was probably best known at the time (ten-ish years ago) for magazine profiles of pop celebrities like Britney Spears and 50 Cent. Chuck Klosterman, musing on a podcast about the still relatively early days of smartphone ubiquity, remarked that we were all embarking on a grand experiment that could reveal what’s lost collectively when everyone stops daydreaming.
Perhaps not surprising: Klosterman would go on to spend much of the decade since that podcast writing about thought experiments, philosophy, and ethics. And while not exactly measurable, I think his provocation about the lost imaginational value of daydreaming was pretty prescient, and it’s something I’ve been thinking about lately in the context of another emerging technology and the impact it might have on our collective imaginative capacity.
If you follow this sort of thing, you might have seen some heated discussion online of a recent paper from the MIT Media Lab describing a study of “the cognitive cost of using an LLM in the educational context of writing an essay.” Most of the conversation has been focused on the learning loss (or “cognitive debt”) found to be experienced by participants who used an LLM to write the essay versus those who had access to a search engine but no LLM and those relying on “brain only.”
The MIT folks have been careful to emphasize the importance of context and to discourage use of terms like (no joke) “brain rot” in discussing the implications of the study. But the upshot for learning and education is both pretty stark and – to my mind – pretty intuitive: When you do less of the work, you derive less of the learning benefit. That seems pretty uncontroversial.
Interestingly, it’s also probably less relevant to LLM usage outside of the education/learning context and in the larger context of the future of knowledge work. Most knowledge work isn’t done for the cognitive benefit of the worker. Rather, it’s performed to generate some informational or analytical value external to the worker herself. If the work is of good quality, it might not matter who does it – or how.
But there’s actually another aspect of the MIT study that’s more relevant and concerning here. LLM usage in the study tended to have a homogenizing effect on the essays produced by that participant group. As the lead researcher explained to Kyle Chayka of the New Yorker, “The output was very, very similar for all of these different people, coming in on different days, talking about high-level personal, societal topics, and it was skewed in some specific directions.”
Chayka discusses the MIT study and several others that point in a similar direction in an article bluntly titled AI is Homogenizing Our Thoughts. This, I think, is the bigger concern for imaginative work – and for innovation, and it has me wondering à la Klosterman whether we’ve embarked on a grand experiment that could reveal what’s lost when we overrely on tools that quickly and easily supply us with superficially decent but narrow/blandly undifferentiated outputs.
Preaching the gospel of decentralized innovation years ago, Pascal was fond of quoting the legendary chemist and Nobel prize winner Linus Pauling: “The best way to have good ideas is to have lots of ideas.” The unspoken assumption there was that the ideas would be differentiated, which historically was a pretty safe bet when those ideas would be coming from a diverse group of humans with a range of different experiences and influences – different sets of training data, if you will.
What happens to that bet should all of those marvelously differentiated humans come to rely on an uber-convenient tool that tends to homogenize output?
Ethan Mollick will tell us that “[T]he best way to make sure that AI doesn’t make you intellectually lazy is to not use it in an intellectually lazy way.” Fair, I suppose, but I’m not sure how much of the workforce will have the time, training, support, or incentive to identify and explore intellectually robust ways of LLMs in their work.
We’ll see…
@Jeffrey