‘Friction-maxxing’ in school: Students should read primary literature, not AI summaries

Large language models have been designed, supposedly, to relieve friction and decrease the amount of effort needed to arrive at a desired outcome. In undergraduate classes, this aim poses a challenging question for both students and educators: Which LLM-aided shortcuts are helpful to the learning process, and which are harmful?

Many neuroscience and psychology researchers have made compelling arguments for incorporating LLMs into coding practice and instruction, but when it comes to reading and summarizing primary scientific literature, I believe undergraduate students and their instructors need to leave LLMs out of the process. Rather than asking ChatGPT to summarize a paper, students should encounter, dissect and interpret it themselves.

Outside of academia, the idea of “friction maxxing” has been gaining momentum as a way to build up tolerance to inconvenience, whether that means schlepping to the grocery store instead of ordering delivery or calling a friend instead of sending a quick text. I think this idea could play a helpful role in neuroscience education as well. Having to fight through a paper in a new field, encountering unfamiliar jargon at every turn, can be a slog, but it has both short-term and long-term benefits for young scientists. As a lecturer at the University of Pennsylvania, I teach seminars for first-year college students in which they choose a research topic, explore the literature and write a short paper about their findings. In my short time teaching these seminars, I have watched students reap the benefits of engaging directly with primary literature. 

My Ph.D. research on metacognition has given me a helpful lens through which to observe this learning process. Neuroscience is an inherently interdisciplinary field, and its practitioners must routinely engage with relevant work outside their area of expertise. A schizophrenia researcher, for example, may need to interact with research in neurochemistry, systems neuroscience and even sociology. They need to be able to navigate those papers—not to learn every term and technique but to integrate it into their understanding. And encountering unfamiliar jargon, methods and styles of writing is key to calibrating your metacognitive awareness around a new topic. This set of cognitive skills does not develop overnight, and it certainly doesn’t come from outsourcing the work to LLMs.

In the short term, struggling through a scientific paper can be a humbling experience. One of my students recently told me that he had experienced the Dunning-Kruger effect: He had started his research project feeling like he already knew quite a bit about the topic, but after reading one or two papers, he became aware of just how much he didn’t know. Hubris, or overconfidence, can negatively affect learning outcomes, whereas intellectual humility, or well-calibrated confidence, predicts better mastery. Had he used an LLM to translate the papers for him, he may not have had the same rude awakening that alerted him to his own lack of understanding. 

S

ome may argue that LLMs can support academic reading in more productive forms than a straight summary. I agree that LLMs could be useful for unpacking a complex concept, but educators need to thoughtfully scaffold LLM use in their classrooms if they are going to allow it. First, LLM outputs are still factually unreliable, so students need to be able to fact-check them. But students won’t be able to fact-check if they don’t build their confidence by reading primary literature or finding other trustworthy sources first. We owe it to students to equip them with the tools and skills they’ll use long into the future, even if LLMs are part of the equation.

In my class, I show my students where to find accessible yet reliable sources of information, including textbooks and popular science articles in trustworthy outlets, that they can use to decipher complicated text. I have seen how students in my class who switch between reading popular science pieces and primary literature tend to walk away with a broad understanding of the field as a whole and the ability to dive deep into specific findings. I also introduce my students to the expert-written articles, such as those in Nature’s News & Views or Science’s Commentary sections, which journals publish to add historical context often not included in the paper itself. These pieces give students a sense of the importance of the findings. These resources have stood the test of time and are more dependable than LLMs. 

Teaching students to read scientific literature can also be a helpful way to teach them about the scientific process itself. In my class, I use primary literature as a window for discussing the norms and conventions of knowledge attribution in science. Deciphering the original wording of a scientific paper opens up a conversation about how and why authors build an argument around scientific consensus rather than citing individual findings. This exchange offers students a glimpse into scientific culture, welcomes them in and helps to steer them away from consuming science through a predigested, and potentially inaccurate, third party.