Colleges Bet on AI Literacy Mandates as Brain Science Flags a Cost

Purdue University will require every incoming undergraduate to prove basic AI competency starting this fall, the first graduation mandate of its kind in the country. Ohio State, the 64-campus State University of New York system and DeVry University are close behind with mandates of their own.

A non-peer-reviewed study out of the MIT Media Lab, published earlier this year, complicates the timing. Researchers wired college students to EEG headsets, electroencephalography sensors that track electrical brain activity, and found that students who wrote essays using ChatGPT showed the weakest brain connectivity of any group tested, then struggled to recall their own words hours later. Colleges are building the literacy campaigns educators have been demanding. The early data on whether those campaigns work is thin.

A Brain Study Turns a Warning Into a Measurement

The study, led by MIT Media Lab research scientist Nataliya Kosmyna, split 54 participants into three groups: one wrote essays using only ChatGPT-4o, one used search engines but no AI, and one used nothing but their own knowledge. Each attended three sessions over four months, wearing a 32-electrode EEG headset while writing 20-minute essays built from SAT-style philosophy prompts, on topics like loyalty and whether a perfect society is possible.

The results scaled with how much outside help a person used. Brain-only writers showed the strongest and most distributed neural networks of the three groups, search engine users came in the middle, and ChatGPT users showed the weakest connectivity. Researchers call the pattern cognitive debt: speed now, at a cost that shows up later.

Time reported that ChatGPT users “consistently underperformed at neural, linguistic, and behavioral levels” and grew lazier with each essay, increasingly resorting to copy and paste by the study’s end. A fourth, optional session made the point sharper. Eighteen participants switched conditions, and those who had used ChatGPT for three straight sessions and then wrote unaided still showed underwhelming brain engagement, as if the deficit did not disappear the moment the tool was taken away. The paper has not yet cleared peer review, and its authors caution against generalizing beyond one model and one task.

Four Campuses Bet on Mandatory AI Literacy

That has not slowed the policy rollout. Purdue’s Board of Trustees approved an “AI working competency” graduation requirement in December, part of a broader AI@Purdue strategy, and it applies to every undergraduate entering this fall on the West Lafayette and Indianapolis campuses. SUNY folded AI ethics into its existing information literacy general education requirement, and Ohio State launched what it calls an AI Fluency initiative last autumn. DeVry is embedding AI literacy into every course by the end of this year.

Nobody involved is pretending the details are settled. Purdue provost Patrick Wolfe put it to the student newspaper days after trustees signed off.

Can we get 2,500 faculty to agree on one piece of guidance?

Wolfe told the Purdue Exponent the university still has to work out logistics without adding time to graduation or overloading an already crowded core curriculum. SUNY chancellor John B. King Jr. framed the change differently in the system’s announcement of the revised general education requirements, saying the university system is committed to academic excellence, which includes a robust general education curriculum. Ohio State’s rollout leans harder into hands-on workshops, folding generative AI basics into first-year orientation programming rather than a single new class.

Is Any of the Training Actually Landing?

A global survey of 45,398 students and faculty across 35 countries found 88 percent of students and 77 percent of faculty now use AI in the classroom, both up 16 percentage points in a single year. Only 29 percent of students believe their instructors are equipped to guide that use, and barely three in ten students see results from literacy training that most faculty say they already completed.

• 88 percent of students worldwide now use AI in their learning, and 77 percent of faculty use it in teaching, per the Digital Education Council’s 2026 global survey.
• 29 percent of students trust their instructors to guide AI use, a figure that drops to 17 percent among students specifically in the United States and Canada.
• 64 percent of faculty report completing AI literacy training, yet fewer than three in ten students say they can see it working.
• Faculty intent to keep using AI in the United States and Canada fell from 76 percent to 67 percent in a year, the steepest regional drop the survey tracked.

The mood gap runs deeper than trust. Just 26 percent of faculty in the United States and Canada said they felt excited about AI’s potential to make learning more effective, against 57 percent of faculty in the Asia-Pacific region answering the same question. Adoption is not the problem. Confidence in what adoption is doing to students is.

The Companies Writing Their Own Guardrail Curriculum

The organization measuring that gap is not a neutral referee. The Digital Education Council is an industry and university membership network founded in Singapore in 2024 by an edtech accelerator; its report carries a commercial partnership with the publisher Pearson. That does not make its numbers wrong. It does mean the industry building AI products and the industry measuring their classroom fallout increasingly overlap.

The overlap runs further into the classroom itself. Anthropic has partnered with the nonprofit Teach For All to train more than 100,000 teachers across 63 countries in what it calls AI fluency, using its own Claude chatbot as the teaching tool. The company whose product prompts the literacy concern is also funding and shaping the curriculum meant to address it, a structure that echoes how tobacco and food companies once helped write the health guidance aimed at their own products.

Jim Wentworth, associate director of educational innovation at the University of Illinois Urbana-Champaign’s Center for Innovation in Teaching and Learning, made the food industry comparison explicit in an essay arguing that AI literacy needs to become a campus-wide public health campaign. His own acknowledgments section carries a small, telling detail: he thanks a colleague for reviewing an early draft and credits Google Gemini for helping refine the essay’s style and clarity, while noting the human author remains solely responsible for any errors. An essay warning against outsourcing thought was partly polished by the tool it warns about.

The Food Industry Needed Thirteen Years for One Label

The comparison Wentworth draws holds up reasonably well on the historical record. The bipartisan Senate Select Committee on Nutrition and Human Needs, chaired by Senator George McGovern, published Dietary Goals for the United States in January 1977, calling for Americans to raise carbohydrate intake and cut fat, sugar and cholesterol. Meat, dairy and egg producers objected at a second hearing later that year, and the committee released a softened second edition within months. The Department of Agriculture never adopted the original goals outright; instead it partnered with federal health officials to issue the first Dietary Guidelines for Americans in 1980. A peer-reviewed history of the committee’s clash with meat industry lobbyists describes senators facing a scientific debate they were not equipped to referee, producing two editions of the same report in one year.

Mandatory nutrition labels did not arrive until the Nutrition Labeling and Education Act of 1990, thirteen years after McGovern’s report and a full decade after the low-fat food boom it triggered. Purdue’s mandate arrives roughly three years after ChatGPT’s public launch in November 2022. Higher education is moving structurally faster than the food system ever did. Whether faster also means better understood is the open question the DEC numbers raise.

The Case for Augmentation Over Displacement

Wentworth’s essay pushes back against the word “artificial” itself, arguing it frames the technology as a rival rather than a partner, and makes a case for augmentation instead. He lays out three concrete swaps for classroom practice.

• Amplifying critical thinking: use AI to debate an argument or stress-test a theory’s assumptions instead of asking it to summarize a book.
• Driving scientific discovery: let AI sift millions of data points so a researcher can spend more time on the hypothesis, not less time thinking altogether.
• Enhancing creativity: keep the human as the visionary and editor, using AI to speed up iteration before a final creative decision.

Some of that model is already shipping. Google has tied Gemini, NotebookLM and Classroom into a single study-notebook and test-prep loop for students. Zoho is chasing a similar bet abroad with free AI classes aimed at government schools across India. Five academics interviewed on the Channel Islands made the human-centered argument even more directly, concluding that technology alone falls short without the people who apply it. Wentworth’s own framing rests on the same bet: that the tool is judged by what it does to the thinking around it, not by how fast it produces an answer.

Frequently Asked Questions

What does “cognitive debt” mean in AI research?

Cognitive debt is the term MIT Media Lab researchers use for a pattern they measured in their 2025 essay-writing study. Relying on a chatbot lowers immediate effort but appears to weaken memory, neural engagement and a person’s sense of ownership over the work, a cost that surfaced even after participants stopped using the tool.

Has the MIT ChatGPT brain study been peer reviewed?

Not yet. As of publication the paper remains a preprint on arXiv and has not cleared peer review. Its authors caution that the sample, 54 participants mostly from five Boston-area universities, is small and limited to one model, GPT-4o, tested on a single short essay task.

Do the new college AI literacy requirements apply to students already enrolled?

Mostly not yet. Purdue’s AI working competency and SUNY’s revised information literacy requirement both begin with students entering in fall 2026, not currently enrolled undergraduates. Ohio State’s AI Fluency initiative and DeVry’s course-by-course rollout are already reaching students this year.

Does using AI for schoolwork always weaken critical thinking?

The evidence so far cuts both ways. The same global survey that found trust and training gaps also found 61 percent of students say AI frees them to spend more time thinking through ideas, and 31 percent say they are attempting harder work than they would have tried on their own.