Two recent surveys estimate that 70 to 80 percent of secondary students are using AI. While some use it to help them understand schoolwork, 63 percent of the users in one survey said they simply rely on AI to provide answers for their homework assignments.
That may lead to higher grades in the short term, but the evidence is mounting that the long-term effects are devastating. A recent large-scale study of Chinese secondary students found that when they stopped doing their own homework and used AI instead—as about 50 percent of those who used AI did—their scores on exams reached historic lows.
As for teachers, the proportion who use AI is smaller than that for students but growing rapidly. In surveys done by the EdWeek Research Center, the percentage who said they use AI to some extent in their teaching jumped from 34 to 61 percent between 2023 and 2025. According to another survey, teachers use it mostly for preparing lessons, creating worksheets, and modifying materials “to meet students’ needs.”
No doubt some of this AI use is beneficial, especially for saving teachers’ time. But neither they nor their students are getting adequate guidance on when AI is helpful and when it actually interferes with teaching and learning. Among teachers, 82 percent said they had received no formal guidance on using AI, according to a survey done earlier this year. Even when there is guidance on student use, teachers are given a lot of leeway in how they interpret it. And most secondary students say their schools aren’t teaching them much about AI either.
At the same time, the Trump administration is calling for more use of AI in schools, with a directive to the Secretary of Education to prioritize AI training for teachers through a federal grant program.
But the question of who provides that training and what it consists of is crucial. Last year, for example, the American Federation of Teachers launched an AI training center with $23 million in funding from companies that have AI platforms: OpenAI, Anthropic, and Microsoft.1 Will that training inform teachers about when it’s best not to use AI? Or will it encourage as much use of AI as possible?
Understanding How Learning Works
Fundamentally, in order to become wise users of AI, teachers—and, to some extent, students—need to understand the evidence from cognitive science on how learning works, and particularly the finding that there is no learning without some cognitive effort. If a student has ChatGPT or its equivalent write an essay for her, she’s not only outsourcing the writing, she’s also outsourcing the effort—and the learning that goes with it. She may not be able to quote anything from the essay just minutes after “writing” it.
That’s why cognitive psychologist Daniel Willingham proposed this principle in a recent Substack post: “Students should only use AI for things that they already know how to do well.” In the workplace, Willingham argues, it makes sense for people to use AI “as an associate” in performing their tasks, as long as they’re honest about AI’s contribution. But the point of education isn’t just to have students produce a paper or a problem set; it’s to enable them to learn from the mental processes involved in doing those things.
That framing suggests that schools could teach students how to use AI, as long as the lessons are embedded in material they already know well.2 Or students could use it for multiplying large numbers after they’ve solidified their knowledge of how to multiply. But they shouldn’t be using AI as a shortcut that prevents them from acquiring, reinforcing, or deepening their knowledge of material in the curriculum or mastering the skills the curriculum is designed to teach.
Most students aren’t in a position to make good judgments about when to use AI—or to resist the temptation to resort to it, even if they suspect it’s not great for their cognitive development. So it’s crucial that teachers acquire the understanding of cognitive science that will enable them to determine when AI is likely to interfere with learning. Unfortunately, teacher training programs rarely equip them with that understanding.
Frameworks for AI Use in Schools
Two efforts are currently underway to provide a “framework” for incorporating information about AI instruction into teacher prep curricula. One was released earlier this month by the American Association of Colleges of Teacher Education, which represents over 500 university-based programs. It identifies four “pillars” that should be included in AI teacher training, one of which is labeled “cognitive architecture and advocacy.”
That sounds promising, but the discussion of AI and cognition is disappointingly brief. There’s a mention of using AI to reduce “extraneous cognitive load” so that students can focus on “higher-order conceptual tasks,” but no specific guidance about what that means.
A prospective teacher might conclude that it’s fine for a student to use AI to, for example, create an outline for an essay as long as she writes the essay herself, on the assumption that outlining is just an “extraneous” menial task. In fact, though, creating an outline, if done well, is a powerful learning experience—one that the student will miss out on if AI does it for her.
A more promising framework is under construction by an organization called Deans for Impact, in collaboration with another organization called TeachingWorks, with a release planned for spring 2027. While the exact contours of the framework are still vague, CEO Valerie Sakimura told me the focus will be on “what novice teachers need to know about how learning happens in order to make good decisions” on using AI.
In formulating their framework, DFI and TeachingWorks will conduct surveys and interviews as well as visits to K-12 schools, teacher prep programs, and other organizations that are piloting programs on AI usage. When released, the framework will include recommendations for changes to teacher-prep coursework and student teaching experiences, along with case studies of early adoption efforts.
Deans for Impact is in an excellent position to address the AI issue thoroughly and accurately. Founded in 2015, the organization has worked with over 250 teacher prep programs to improve their offerings, largely by incorporating principles grounded in cognitive science into their courses. Earlier this year, it published the second edition of The Science of Learning, a user-friendly summary of cognitive science research and its implications for teaching.
In addition to understanding when it makes sense for students to use AI, the DFI framework also aims to help teachers use AI themselves. For example, Sakimura said, teachers might ask an AI tool to make a reading assignment “more rigorous,” but without an understanding of cognitive science they might not know what “rigor” looks like. They might think it just means a passage that’s longer or uses more complex vocabulary, rather than one that requires more productive, effortful thinking.
A Wedge Issue for Cognitive Science?
Sakimura said the DFI framework will also include policy recommendations, although it’s not yet clear what those might be. One that would make sense is to add requirements about AI-related knowledge to state teacher licensure exams. That could be an effective “wedge issue” for getting cognitive science content into those exams, and hence into teacher-prep curricula. Sakimura said that alongside the release of the AI framework, DFI would be continuing its efforts in that regard, citing Maryland and Texas as early leaders.3
In fact, if there’s a silver lining to the rapid spread of AI, which so far has had mostly negative consequences for education, it may be its role in spreading the news that learning requires a certain kind of cognitive effort.4 Most people, even if they don’t have a background in cognitive science, seem to have an intuitive sense that using AI to write an essay isn’t the same thing as writing one yourself. Even someone as immersed in technology as Bill Gates recently warned that AI could “lead to many people learning less.”
Perhaps, after being overshadowed by reading for so long, concerns about AI will at last lead to writing getting the attention it deserves from educators and policymakers. A recent New York Times article titled “What Do Students Lose When They Stop Writing?” explained that writing “builds our working memory, executive planning skills and metacognition.” That’s all true—and there are even more potential learning benefits from writing than that, as long as it’s taught explicitly and in a cognitively manageable way.
But as promising as the DFI framework is for equipping novice teachers with the knowledge they need to use AI wisely, it’s unlikely to be enough to ensure that AI use, and education in general, aligns with evidence from cognitive science. For one thing, the vast majority of teachers have already been through their pre-service training and therefore won’t benefit from whatever changes are made to ed school curricula going forward. We also need to reach the millions of teachers—and administrators—who are already on the job.
But even that is unlikely to be enough. It’s important that educators understand the basic principles of how learning works and unconscionable that most haven’t already been provided with that information. But asking them to translate those principles into classroom practice is a tall order—especially if the instructional materials they’re expected to use don’t align with cognitive science, as is often the case.
Reaching Curriculum Publishers—and Students
To make real progress, we need to approach the problem on multiple fronts. That should include providing the creators and publishers of curricula with training in the principles of cognitive science, so that they can produce materials that incorporate those principles rather than going against them.
Ideally, we’ll also provide students themselves with information about the importance of cognitive effort, in age-appropriate ways. It’s not surprising that students are using AI to bypass the hard work of writing or to simplify challenging reading assignments. We human beings are hard-wired to avoid effort, especially if we don’t understand its value. That may be true even if we see it as cheating.5
One problem is that, as a recent high school graduate told a reporter, many students don’t see why their schoolwork will benefit them once they graduate, so “they just want to skip through it as fast as they can”—and AI is a great way to do that. At the same time, teenagers believe that the skills that matter most for their future are the things AI “can’t or shouldn’t replace”—like “critical thinking” and “reading and understanding complex information.”
It might help to inform students that by relying on AI to do their assignments, they’re actually cheating themselves—and depriving themselves of the very abilities they say matter most for their future. Without acquiring the knowledge and skills covered in the curriculum, they’re likely to render themselves far less able to think critically or understand complex information. And instead of being the masters of AI, they will find themselves at its mercy.
AFT President Randi Weingarten expressed concerns about AI in schools earlier this year, even calling for a ban on having elementary students use it. But her change of heart apparently has not caused the AFT to backtrack from the establishment of its AI training center or the support it got from tech companies.
Actually, Willingham argues that students should learn how to use AI “at home and in the workforce,” not at school.
Maryland requires science of learning training for classroom teachers, and Texas has included an understanding of the science of learning as a requirement in its teacher certification code.
Of course, too much cognitive effort, or cognitive load, interferes with learning. But there is no learning without some cognitive effort—the kind that cognitive scientists call “desirable difficulties.”
At this point, some students don’t even see anything wrong with cheating. “Cheating is a creative expression and a form of problem solving,” one Ivy League student wrote in response to a survey on AI use.
Willingham's rule shifts the teacher's task from detecting AI to checking the skill was already there. Fluent output doesn't prove that, because the model predicts tokens rather than learning.
As a cautionary tale about avarice, theft, concentration of power, self-destruction on a hyper scale, and the dangers of human hubris, addiction to novelty and marketing hype masquerading as “innovation” and “intelligence.”