When Anthropic gathered focus groups of students to better understand how they use AI, a colorful term kept cropping up: “brain rot.”

Students know that mindlessly plugging their assignments into AI tools isn’t exactly leading to a deep understanding, Anthropic’s education lead, Drew Bent, said. But the temptation to take these shortcuts is there nonetheless.

“If the tools make it so easy, and it’s late at night, and you just need to go to sleep, and you want to finish an assignment, and the tool just pushes you toward giving you the full answer, then that’s not conducive for your learning,” Bent told Tech Brew.

The onus, Bent said, is on the big AI labs to try to curb access to this “cognitive offloading,” as he more delicately puts it. Those focus groups were part of the research process for Claude’s Learning Mode, which is designed to help students think through problems rather than serve up answers.

Anthropic is far from the only big AI company with a product like this. Having roiled the education system with chatbots that readily spit out answers or whip up essays, companies like OpenAI, Google, and Microsoft are attempting to offer versions of their AI more conducive to learning, forge partnerships with universities and other educational institutions, and fund grants and initiatives that give students access to them.

“We have this general purpose tool that we built. It was not designed for education. It is being used in that way,” Bent said. “There’s also a lot of concerns around what it’s doing to education. Not all the AI labs are thinking about that as much as we would hope. And so we started to study it from a research perspective, and try to figure out, what can we do to improve our product, our model, our platform, to make it better for education?”

Google recently announced a $1 billion-value grant to give AI tools to college students for free, including its own Guided Learning mode. The search giant has already partnered with more than 100 colleges and universities.

“We’re still so early days about what’s possible. But I think the way to think about it is, a teaching assistant for every teacher, a personal tutor for every student,” Lisa Gevelber, Google’s global VP and managing director, told us of the company’s ultimate vision.

Between a White House executive order earlier this year, a flood of new state legislation, and industry initiatives like the above, schools, companies, and governments are beginning to think more about how generative AI tools can play a managed role in education and prepare students for a workforce where this technology has become widespread.

“It does seem like there’s a lot more wind beneath the sails for this right now because of actions at…the federal and state levels. In industry, I have definitely seen an uptick,” Victor Lee, an associate professor at Stanford University’s Graduate School of Education who studies AI and education, said.

Model students

Much of the attention on ChatGPT and its ilk in schools thus far has been around its use in cheating and the subsequent harm that could have on the critical faculties of students. How do you train a model that actively combats those tendencies?

Bent said the system prompt—the base instructions invisibly appended to every prompt—is part of it, and most education-focused models tweak this text to direct models to avoid easy answers. At Anthropic, Bent said the company also thinks about the data the model is trained on and how it’s presented to students.

Lee said education-focused models should also avoid some other common traits of general-purpose models.

“What I would like to see, ideally, would be responsible implementation for, one, matters of privacy and just the design of the experience to manage the duration and amount of use,” Lee said. “For example, [there’s] a sycophancy problem with some AI...to keep you talking to the AI. And that’s not necessarily in the users’ or the learners’ primary interest.”

Marc Watkins, director of University of Mississippi’s AI Institute for Teachers, said educators want AI that slows down the output of answers, challenges users with quizzes, and fosters more dialogue.

“Things like that, that involve a little bit of friction in the process of using an AI tool, that really does just sort of hyper-charge the experience, can be really helpful for learning,” Watkins told us.

Educator concerns

Many educators remain worried about the role of AI in classrooms, experts said, but it now goes beyond cheating: There’s now reason to fret about the mental health of students who see AI models as companions.

Watkins said it will take years of learning and figuring things out on the part of educators before AI will be a fully productive tool.

“I think it can have a positive impact on education…once it slows down and once you get intentional with it,” Watkins said. “But man, that takes time, that takes training, that takes a lot more work on the part of the teacher to do. So I think it’s going to be at least 10 years down the road before you can really get a handle on this in education.”