The advent of easily accessible AI tools carries with it potentially big shifts for many established systems, especially in learning environments. As students, educators, and anyone trying to gain knowledge grapple with how to use these new tools, Stanford researchers weigh in on a critical question: How should AI be implemented in pedagogical settings, and what will its effects be on learning?

Chief among the concerns is the issue of cognitive offloading: the act of reducing the mental processing requirements of a task through using different tools. A classic example of cognitive offloading is the use of calculators, which reduce mental effort by solving mathematical equations for us. Even counting with your fingers is an instance of cognitive offloading.

This process, researchers say, isn’t inherently problematic. For some tedious tasks, cognitive offloading enables people to work more efficiently and focus on more important aspects of an endeavor. The difference with AI, researchers say, is that it has the ability to offload an entire job.

“If I can completely delegate a task to AI and I never spend time thinking about it, then there’s no learning,” said Guilherme Lichand, an assistant professor at Stanford Graduate School of Education.

Even if we aren’t delegating to that extreme, the researchers warn that we have to carefully consider which mental processes and skills are being offloaded. Offloading deeply cognitive processes to AI, things like creative problem solving and active learning, can reduce a person’s ability to complete tasks when they lose access to those AI tools. Research also shows that a student’s self-perceptions of creative ability and learning resilience can be reduced by offloading learning tasks.

“We’re still learning about how AI impacts students and educators,” said Lily Fesler, a senior research associate at the AI Hub for Education within Stanford’s SCALE Initiative. “What we’ve learned so far is that it has the potential to both improve and harm learning outcomes. The difference likely comes down to the type of tool that’s being used.”

Lily Fesler, Guilherme Lichand, and Hariharan Subramonyam examine how AI can support or undermine learning. | Courtesy of the subjects

The process of learning

To understand how AI can impact learning, the researchers highlighted a few key metrics by which the pedagogical process can be evaluated. The first is transfer of learning: the ability to apply previous knowledge, skills, and strategies to new contexts. As Lichand explained, recalling information and thinking about it critically to approach novel challenges is fundamental to learning.

Another essential element, highlighted by Hariharan Subramonyam, assistant professor at the Graduate School of Education, is the ability to build mental representations of real-world concepts. “Learning is representing,” he said. “When we’re trying to learn something, our goal is to build good internal representations of what we are learning to serve some later purpose.”

The researchers underscored that both of these elements of learning require motivation and active engagement with educational materials, which can demand time, resilience, and repetition.

How AI might help learners

Much of the potential of AI tools in learning seems to lie in how they provide users with the ability to create, iterate, and engage Socratically.

“A lot of learning used to be based around creating things,” Subramonyam said, “but around the time of the Industrial Revolution, there was a parallel push to make education more efficient.” That paradigm shift brought about an educational model that mirrors the industrial factory system – with ringing bells, 45-minute class periods, and a strong focus on top-down knowledge acquisition.

What AI tools can provide, Subramonyam said, is a return to constructivist, bottom-up learning via creating and iterating, a shift he believes could help users engage more actively in learning.

He highlighted the ways AI tools can help people visualize concepts by designing and iterating their own models rather than simply reading about a topic. While this can be considered cognitive offloading for the way it moves some of the learning process from the mind to the physical world, the visualization can help some people learn better. In his own courses, Subramonyam had students design simulations in AI models, but only after doing the design planning themselves.

“Multiple students told me that they saw value in coming to class, in working together and building these things,” he said. “This iterative process of trying, failing, and ultimately creating things is very exciting for learners.”

Subramonyam hopes AI can support a creation-centered, constructivist style of education that helps people participate more deeply in the learning process by actively building their own knowledge rather than simply memorizing information for an exam. “Giving learners the capacity to create things is where I think AI can be really powerful in making learning more meaningful and intentional,” he said.

The danger of offloading learning

Despite the exciting new avenues for learning, the pitfalls of AI tools are worthy of serious consideration. A significant concern centers on which learning processes are in danger of being offloaded to AI.

As Lichand explained, discomfort and resilience are essential components of learning: “Learning involves a process of productive struggle, and the challenge of understanding how to self-regulate during that process.”

This social-emotional experience is an indispensable part of building knowledge and applying learned skills to new scenarios. “It’s a question of how to keep striving for that answer, even when you’re frustrated,” he said. “That’s why these social-emotional components – your academic self- concept, your beliefs about your ability to learn and get to the answer – are a very important component of learning.”

To learn more about how cognitive offloading to AI affects student outcomes, Lichand and predoctoral fellow Sachin Allums studied the work of Brazilian middle schoolers. They divided the students into four groups with different levels of access to an AI tool and had them complete two creative tasks.

The results showed two important things. First, when students were allowed to use an AI tool for the first task and then had it taken away, the advantage of the tool was lost – meaning the AI tool impeded the process of transfer. More concerning, Lichand said, was that once students lost access to their AI tools, their ability to solve tasks creatively dropped below that of students who had never had the tool at all.

“It does seem like it’s a story about offloading,” Lichand said. “These effects are concentrated in the kids who use the AI tool more in the first task, so it suggests that the more you offload your abilities, the bigger the negative effects later on.”

Fesler and her team have seen a similar pattern in the broader research landscape. They reviewed over 800 papers on AI in education and focused on 20 high-quality causal studies. While the research is limited, the studies so far seem to agree with what Lichand and Allums found.

“We definitely see that cognitive offloading can be a problem,” Fesler said. “There was one study that showed that students who used a generative AI tool to learn new information showed lower recall and lower brain activity than learners who weren’t using a generative AI tool.”

The review found that the most detrimental tools were general-purpose chatbots. When students used one to study for an exam, it had the potential to decrease their test scores once they no longer had access.

Designing AI tools for learning

Fesler stressed that the kind of tool provided to students matters a great deal. When she looked at tools with more scaffolding – guidance designed to support students' thinking – the detrimental effects of cognitive offloading were not as clear.

Examples of tools with more scaffolding include those primed to respond with Socratic questioning rather than giving outright answers or those that give students feedback on essays without generating the writing itself. “AI assistants have the potential to give students more regular, immediate feedback on essays than what a teacher has time to provide – as long as the AI tool isn’t writing students’ essays for them, and isn’t entirely replacing teacher feedback,” said Fesler.

What matters, the researchers underscored, is that the tools have enough scaffolding to avoid fully offloading the process of understanding and the development of social-emotional resilience in learning. That resilience, Lichand reemphasized, is crucial for ensuring learners can actively transfer acquired skills to new tasks.

For Subramonyam, a special strength lies in ensuring that some tasks remain time-intensive. “I think doing things slowly and deliberately is a desirable feature in research,” he said. “This is what allows us to engage in critical thinking and be intentional about what we’re trying to do.” He also highlighted that to build accurate representations of learned content, learners need to engage deeply with that content rather than rely only on efficient AI summaries.

Fesler agreed. “It’s about figuring out a balance when you’re designing a tool. You want enough friction so that people are learning rather than just getting the answers right away, but not so much friction that they won’t use the tool at all,” she said.

The current state, and future considerations

The researchers all echoed the same need to better understand how AI is affecting the learning process. Top of mind for Fesler is more high-quality causal studies. “Right now it kind of feels a bit like the wild west,” Fesler said. “Students are using AI all the time, a lot of teachers are using AI, and the research is just trying to catch up.”

For his part, Lichand stressed the importance of not rushing to implement tools and instead working to introduce them thoughtfully. “The fast is often the enemy of the good. I think people feel the need to rush because it’s an arrival technology – rather than the typical ed tech that schools decide to adopt or not – and because there is a sense of inevitability,” he said. “But that’s just not true. There is a lot of value in waiting.”

At the moment, the researchers agree that the technology has outpaced the evidence and that more research is needed to understand how AI tools can be helpful and where they can be detrimental. In the meantime, they highlight the importance of deliberate implementation and intentional guardrails to ensure users can still benefit from the learning process.