Blog

Over the past several months, we’ve been inundated with news about AI in education. OpenAI, Anthropic, and Google have all rolled out products for students and teachers. Entire new high schools and colleges have sprung into existence on the premise that AI can accelerate learning. Harvard Dean David J. Deming has told the college’s almost 7,000 undergraduates that they not only can but should incorporate AI into their studies.

It’s easy to understand the enthusiasm. What could be better for learning than superintelligence on tap? And with AI infiltrating all manner of work, shouldn’t students learn to use it now?

Unfortunately, by and large, the reality in education has not lived up to this hype. In fact, it’s becoming clearer every day that AI-assisted study can be harmful to fundamental cognitive skill development.

Of course even well-intentioned and carefully designed study tools can and do carry students into cheating territory. But the more subtle and pervasive problem with AI-assisted study is something else, a phenomenon often called the “fluency illusion.” Familiar to tutors and educational researchers alike, the fluency illusion describes the experience—one we’ve all had—of feeling sure we know something when we don’t. It’s an “uh huh” in lieu of an “aha,” information in lieu of knowledge.

The risk of false fluency is not only getting surprised at the difficulty of an assessment. It’s also the loss of opportunities to struggle, and with them, opportunities to develop metacognition—the ability to know, with accuracy, when you understand something and when you don’t. Now research findings and institutional perspectives are coalescing around this concern. A few stories have stood out to me:

• A recent large, well-designed, and longitudinal study found that students who used general-purpose AI tools scored about 20% lower than their un-AI-assisted peers on high-stakes entrance exams.
• A Brown economics professor reported the worst final exam grades he’d ever seen after his class held its final in the lecture hall. And that’s after the midterm, taken at home and likely with AI help, showed the highest grades he’d ever seen.
• MIT’s committee on the use of AI in teaching and learning found that AI use is “undermining student mastery and confidence.”
• A similar committee at Cornell found that “AI can produce the appearance of learning without the intellectual formation that makes learning meaningful and durable.”

To be sure, there are areas where this new technology is helpful, and we’ve seen some of AI’s benefits firsthand when used skillfully within a tutoring program.

But most of all, this flood of AI educational tools has put in stark relief the tried-and-true ways to cultivate students’ long-term success. Above all, that means directing and motivating students’ own good, hard work—and imbuing the learning process itself with joy and meaning so that shortcuts lose their luster.

So as we head into the heart of the fall term and its major assessments, and as chatbots vie to become our study buddies, we’re sharing an analogy with our students: You wouldn’t bring a forklift to the gym; the point was never simply to move the weight. It was to become someone who can.