13 Powerful Studies That Tell the Story of AI

Technology has always had its skeptics.

In Plato’s Phaedrus, Socrates warned that the advent of writing—a disruptive new technology during an era of oral storytelling—would “create forgetfulness in the learners’ souls, because they will not use their memories.” His students, reading out answers already committed to parchment, would “appear to be omniscient and will generally know nothing… having the show of wisdom without the reality.”

More than two millennia later, with the subsequent threats of printed books, television, calculators, and computers behind us, generative AI is the intellectual catastrophe of the moment, poised to write essays, solve math problems, and “vibe code” with minimal effort from students.

For many teachers, it’s enough to justify banishing the technology from their classrooms.

But writing didn’t compromise memory, calculators didn’t destroy math, and computers haven’t made books obsolete. The question facing schools now, if history is a guide, is not whether AI is good or bad, but under what conditions it supports—or hinders—learning.

A handful of exemplary studies move beyond hypotheticals and offer a glimpse of what’s happening inside college and K–12 classrooms as students and teachers adopt different AI tools—from the ubiquitous large language models that provide lightning-fast answers to entire AI platforms that promise to ease the administrative and instructional burdens of teaching. Researchers from leading institutions like MIT, Harvard, and Wharton, for example, are probing how use of the software shapes memory, motivation, and higher-order thinking skills.

Some studies have pointed to the risk of “cognitive outsourcing”—when students hand off not just tasks, but also the mental effort required to complete them. Others are decidedly more optimistic about our future relationship with AI, discovering in it a tool that can push student thinking and help teachers produce high-quality materials quickly.

The ‘Assistance’ Illusion 

Ask AI for a little boost, and it’s only too happy to oblige. A student prompting the tool for a “good thesis statement about Lord of the Flies” or asking for a “hint about factoring polynomials” waits as the cursor blinks once or twice and then advances rapidly across the screen, generating blocks of polished, human-sounding text. Like the sleight of hand produced by the auto-completion of sentences in word processing software, the AI paragraphs are suddenly shimmering, finished products—and it’s not always clear where the machine’s thinking ends and the student’s begins.

But the research suggests that the convenience and speed comes at a steep cost: Answers that are effortlessly produced are also cognitively slippery, leading to poor recall and shallow understanding. Mirroring past research on the paradox of low-effort, low-impact techniques like highlighting, underlining, and rereading, AI’s nearly-magical fluency can trick students into thinking they’ve mastered a topic. That’s because AI generates polished answers that preempt further inquiry and writes with the unwavering confidence of an expert, researchers explain in a 2025 study—even when the software is wrong or guessing.

Cutting corners: In a large-scale 2024 study, researchers asked nearly 1,000 high school students to solve math problems under different conditions. Teens with access to ChatGPT trounced the competition, scoring 48 percent higher than their peers who completed the work using only course materials and their notes.

But what seemed like a clear victory for AI was short-lived. When asked to solve another set of similar math problems on their own, students who had previously used ChatGPT performed poorly.

“When technology automates a task, humans can miss out on valuable experience performing that task,” the researchers explained. An examination of ChatGPT logs revealed that most of the student interactions with the software were cursory, and a third of them consisted of students asking variations of “What is the answer?” By outsourcing the hardest cognitive work—analyzing the problem to determine the right steps to take, figuring out which formulas to apply, and then methodically working step-by-step to solve the problem—students denied themselves the opportunity to improve their mathematical skills.

Led astray: More recently, a 2025 study led by researchers from MIT reached a similar conclusion in another knowledge domain. When college students were asked to write thought-provoking essays on questions like, “Is a perfect society possible or even desirable?,” those who had access to ChatGPT tended to blindly “follow the thinking” of the tool, producing “statistically homogenous” essays that sounded canned. After wiring students up to EEG machines, the researchers caught a glimpse of the reasons why: “Brain connectivity systematically scaled down” the more students relied on ChatGPT to shape their answers.

Crucially, only 17 percent of the AI users could recall a single sentence from essays they claimed as their own. Students who used a search engine to inform their thinking, or wrote the essays on their own, had 83 and 89 percent recall, respectively.

Scope creep: As AI tools become more sophisticated—and more integrated into web browsers, search engines, and classroom platforms—the line between assistance and outright substitution is becoming increasingly blurred. In another 2025 study, students who learned about a topic through an AI-generated summary risked “developing shallower knowledge than when they learn through standard web search, even when the core facts in the results are the same.”

The biggest danger, perhaps—the one lurking behind all of the relevant studies—is what researchers recently dubbed “metacognitive laziness”—the “habitual avoidance of deliberate cognitive effort” by offloading not just the task itself, but also the monitoring and regulation of academic progress.

Wrestling with confusion is a natural part of learning, and the metacognitive exercise of checking one’s own work and correcting misconceptions is crucial for robust, long-term comprehension, the researchers argue. In the study, students who used AI in class to learn a new language had higher essay scores—largely because they would copy and paste ChatGPT’s answers—but also were prone to becoming “overly reliant on AI,” limiting their ability to monitor and reflect on their learning. When asked to revise their essays, the AI users “predominantly relied on consulting ChatGPT during the revision stage to refine and assess their essays,” unlike their peers who interrogated their writing on their own.

A (Challenging) Learning Partner

The same way human tutors are trained to avoid handing students the answers, well-designed AI tutors are engineered to stifle themselves, refusing to provide answers and guiding students without doing the mental work for them.

Back to you: In a study published last year, Harvard researchers tested the impact of an AI tutor they carefully designed to adopt “best practices from pedagogy and educational psychology to promote learning.” With the right guardrails in place—the tool provided hints instead of answers, broke down difficult problems into more manageable chunks, and followed up with probing questions—the AI-tutored group more than doubled their learning gains compared with students who participated in in-class lessons.

“An AI tutor should not replace in-person teaching—rather, it should be used to bring all students up to a level where they can achieve the maximum benefit from their time in class,” the researchers suggest.

This approach echoes the one taken by the researchers in the 2024 study on math learning, mentioned above, who also exposed students to a GPT tutor whose responses were governed by “detailed instructions asking GPT-4 to avoid giving away the full solution, and instead provide incremental hints to help the student solve the problem.” Students who worked with an AI tutor performed just as well—though not appreciably better—on a quiz of knowledge than those who worked with traditional pencil and paper.

Metacognitive nudges: AI can also assist in planning activities that scaffold academic work—without doing the work itself. By breaking large projects into smaller tasks, scheduling review sessions over several days and weeks, and setting intermediate deadlines that prevent last-minute scrambling, AI can assist students as they manage their workloads. 

In a 2025 study, researchers at Kent State University found that organization and planning tools were nearly as popular as AI tutors, with 29 percent of college students reporting regular use of AI study and scheduling apps, resulting in a “significant reduction in study hours alongside an increase in GPA.” AI tools, then, can help students develop stronger study habits, leading to “more efficient research and information gathering” in addition to “better organization of study materials.”

Upping Your Teaching Game

Not long ago, teachers were skeptical of AI, with just one in four using it for planning or instruction, according to a RAND report. A year later, the number has more than doubled, with a 2025 Gallup/Walton Family Foundation poll revealing that 60 percent of teachers used AI at least once during the school year, signaling a rapid shift from hesitation to hands-on experimentation.

Mirroring larger trends in the workforce—a Harvard Business School study found that employees who used AI for common workplace tasks were 25 percent faster—teachers are increasingly using AI to handle time-intensive tasks. Today, nearly a third of teachers report using AI on a weekly basis to draft lesson plans, adapt materials for different learners, generate quizzes and exit tickets, and analyze student data, with an estimated savings of about six hours per week—the equivalent of reclaiming six full weeks over the course of a school year.

“When using any AI tool to create material, it’s important that you review the material yourself to check for any errors,” write high school teachers Tyler Chance and Kristin Petrowske. “Still, the amount of time saved overall makes it easy to do a quick accuracy check.”

That’s a good question: The quality of academic materials created by AI often outstrips expectations. In a large-scale 2025 study involving nearly 1,700 college students, a Harvard-led team of researchers used ChatGPT to generate multiple-choice practice questions across 91 classes ranging from introductory statistics to advanced biochemistry. Mixed into each set were expert-written questions, allowing the researchers to compare any differences between the two.

Not only did the AI-generated questions perform “on par with those created by experts,” the researchers note, but also they have “the potential to substantially reduce instructor workloads, increase access to high-quality assessments, and improve learning outcomes through more tailored instruction.”

More time for humans: For Danielle Waterfield, a University of Virginia researcher and former special education teacher, AI could play a major role in easing administrative paperwork, letting teachers spend more “face-to-face time with their students.” In a 2025 study, Waterfield and her colleagues compared individualized education program goals written by experienced special education teachers with those written by ChatGPT and found virtually no difference in quality—despite the teacher-written goals taking much longer to produce.

Waterfield’s findings are echoed by a 2023 study that found that AI could provide “time-saving assistance” by handling common administrative tasks such as scheduling, drafting letters to parents, and writing announcements, allowing educators to focus more on instructional planning and student engagement. When used thoughtfully, AI can also help teachers “augment their instruction” by creating quiz materials that “align precisely with the targeted learning objectives and success criteria of the instructional unit,” for example.

Schrödinger’s Cat

Perhaps the most controversial question surrounding AI in schools isn’t whether it’s helpful or harmful, but how it complicates the issue of cheating. In many classrooms, AI exists in a state of ambiguity, acting as both a legitimate study aid and a partner-in-crime, with the line between the two becoming increasingly easier to cross.

“Like many teachers, I sometimes notice my unease growing as I read a piece of student work,” writes high school English teacher Jen Roberts. “Word by word my suspicion builds, and then all at once I understand. This is an artificial intelligence–generated text and not something my student actually wrote. Now what do I do?”

While cheating in the past was moderately cut-and-dried—there’s no doubt that copying answers or plagiarizing passages falls under the category—some AI interactions fall into a murkier gray area. Is reading an informational paragraph produced by AI any different than learning from a lecture or a textbook? Probably not. The problem, writes Wharton professor Ethan Mollick, is that students are prone to “illusionary knowledge” and may not realize that they’ve sometimes removed themselves from the equation entirely. “After all, they are getting advice and answers from the AI that help them solve problems, which feels like fluency.”

The risks of use—and overuse—are real, and can be hard for students to spot. New research, however, may ease our minds, at least temporarily. In an anonymous study designed to elicit honest answers, over 300 college students said they rarely turned to AI to write reflective essays—answering prompts about their personal values, for example, or making connections between a lesson and their own experiences. While a third said they used large language models like ChatGPT for help in those circumstances, most used it only to edit a draft or help them brainstorm topic ideas.

Perhaps, then, motivated students—like motivated students in every era—will eschew the tool in favor of owning their ideas and mastering subjects that interest them. When confronted with personal essays, only a paltry 1.3 percent of the students said they turned in an essay that was “mostly written by generative AI.”