In our recent work with faculty, we often hear two questions: how are students using AI for academic work and how extensively are they using it? In this issue of the AI newsletter, we offer some answers to those questions by summarizing survey data from students at Princeton and in higher education more broadly. We know that faculty are also concerned about how AI affects learning. We review important findings from the research on this topic, and offer a few suggestions for how you might apply the research in your courses. We also share an interview with Tania Lombrozo, Arthur W. Marks ’19 Professor of Psychology, about how her expertise on learning, reasoning and understanding has influenced her own teaching in the age of AI.

How do Princeton students use AI, and what are their attitudes towards it? Although somewhat limited in scope, two recent surveys help answer these questions: The ALTA (Academic Life Total Assessment) survey, which was sent to all students in March 2026 on the initiative of the Undergraduate Student Government and the Senior Survey, administered to the Class of 2026 by The Daily Princetonian.   

The ALTA survey, which had a response rate of about 20 percent, found that a large majority of students, about 83 percent, have used AI in academics, and 64 percent use it on a weekly basis. Students most often ask AI to provide detailed explanations or broad summaries of a topic, or for help with studying. Students report that they less often use AI for more active tasks, like analyzing content for themes or help with data analysis. Most students (75 percent) in the ALTA survey say AI improves their understanding of course content and helps them learn. While students don’t report big improvements in academic performance from using AI, they do report that it helps reduce academic stress. 

Despite common usage, Princeton students are ambivalent about the technology. According to the ALTA survey, many students assume that their peers use AI more frequently than is likely the case. Students also assume those who use AI have an academic advantage over those who do not. Students are especially concerned about their peers using AI in take-home or in-person exams. 

This data aligns well with national data on student usage: students commonly use AI, though the amount varies among different disciplines (more common in STEM disciplines) and gender (more common among men). Students use AI most often to understand course material and to save time (for more, see the study by Gallup and the research by Associate Professor of English at University of Pittsburgh, Annette Vee). One of the main drivers of usage is academic stress.

National data also reveals that students have deeply conflicting feelings about AI—while they might seek help from it, they also recognize that it might be detrimental to their learning. Students report being upset by their peers’ use of it, and worried about how AI use can increase isolation, affect the social aspects of learning, and affect fairness in assessments and grading. AI also seems to harm the relationship between faculty and students, as faculty worry about scholarly integrity and students feel monitored or wrongly accused. A report from MIT similarly points out students’ ambivalence about AI: they are confused, curious, inspired, concerned, and anxious.

While still new and limited, research indicates that many uses of AI are detrimental to learning. There are some exceptions. A study from researchers at Harvard University suggests that students who engaged with a carefully designed AI tutoring chatbot, learned more than the students who engaged with an in-person active learning lesson. The tutoring chatbot had been instructed to incorporate many of the strategies we know promote learning: providing personalized and timely feedback, adjusting to a student’s pace, and managing their cognitive load. It is also worth noting that in this study, “learning” was measured by test scores, which is not necessarily equivalent to durable learning.

Most research, however, shows that the use of AI not only can have negative impacts on performance but also on long-term skill development. More specifically, the research points to the following interrelated, negative consequences:

Cognitive offloading is the outsourcing of intellectual work to an external tool or person. Not all cognitive offloading is negative. If I already understand multiplication, it is not necessarily detrimental to ask a chatbot to solve my math problem. But if my goal is to learn multiplication, asking AI to do it for me is counterproductive because I’m bypassing what is often referred to as “productive struggle” or “desirable difficulty.” To move beyond superficial memorization, a student has to actively engage deeply with a challenging problem or task. The use of AI can remove that necessary friction.

The use of AI can diminish critical reasoning. One study showed that participants tended to accept the response they got from AI even if those outputs were systematically incorrect, and even when they were incentivized to verify the accuracy of the information.

The use of AI can lead to a sense of false mastery or the mistaken belief that you have understood more than you do. In fact, the use of AI seems to make a person less able to reason about the information they gather, and more prone to illusions of understanding than searching the web does. This might be because AI is often used in a way that doesn’t involve deep cognitive engagement.

As Princeton professor of computer science Arvind Narayanan points out, if you use AI in areas you don’t understand, you’re tempted to treat the tool as an expert. Lacking content expertise, you don’t have the ability to evaluate the output AI generates. On the other hand, if you use AI in an area where you are an expert, you can critically evaluate the output of AI. You can also be strategic about the tasks you hand over to AI while freeing up cognitive capacity for other, more important, tasks. In fact, Narayanan argues that for an expert, the use of AI can help facilitate more learning and growth as more cognitive capacity is made available for developing higher level skills.

Interpreted cautiously, the research suggests that AI is not inherently detrimental to learning, but rather it depends on how it’s used. You might help build your students’ awareness and metacognitive skills by utilizing some of the following strategies: 

We suggest not only that you base your AI policy on the learning objectives for your course, but also discuss the ways AI might affect your students’ ability to achieve those goals. Help students reflect on questions like: Am I supposed to learn the skill I’m handing over to AI? Am I using AI to remove the friction necessary for deeper learning and critical engagement? Do I know enough to critically evaluate the AI output? You can find a number of other questions you might encourage your students to reflect on in Considerations for Using GAI for Engagement, Learning, Studying and Exam Prep at the McGraw Center’s website.

Discourage students from treating AI as an oracle, encyclopedia, or teacher. Instead, show them how they might use it as a “teachable novice” and support their own learning in the process. Instruct them to prompt the LLM to feign ignorance and then ask them questions, putting students in the role of providing explanations. This strategy creates the Protégé effect, which “forces the human learner to engage in the effortful act of explanation and reflection.” 

Tania Lombrozo is the Arthur W. Marks ’19 Professor of Psychology at Princeton University, as well as an Associate of the Department of Philosophy and the University Center for Human Values. She directs the Program in Cognitive Science and co-directs the Natural and Artificial Minds research initiative within the Princeton Laboratory for Artificial Intelligence. Her research aims to address foundational questions about cognition using the empirical tools of cognitive psychology and the conceptual tools of analytic philosophy. Her trade book on explanation, Why We Ask Why: The Science of Explanation and the Human Drive to Understand, will be released in October. She kindly agreed to talk to McGraw’s Mona Fixdal (MF) about AI and her Spring ’26 course Cognitive Psychology.

MF: You taught a large class on cognitive psychology this past spring. What was your AI policy? How did you seek to enforce it?

TL: For most assignments, AI use was not permitted at all. One exception was a short writing and presentation assignment for which students were allowed to use chatbots to brainstorm a topic. The instructions for the assignment included: “You are welcome to use generative AI tools to brainstorm topics that might relate to your interests and to cognitive psychology, just like you might chat with a friend or use a search engine. But you should not use AI tools for writing, analysis, or presentation preparation.” 



We did let students know that we would be on the lookout for AI use. But rather than enforcement, my main strategy was prevention through education (see the next question!) and the design of assessments (see the final question!).

MF: One of your areas of expertise is how we learn. Did you talk to your students about the way using generative AI might affect their learning?

TL: I changed and restructured my course content in three ways to educate students about the new realities of generative AI and its potential impacts on learning.

First, I moved the unit on learning (which the textbook covers about halfway through) to the second week of the course, so that students would learn about learning as soon as possible and be able to apply strategies for their own learning throughout the semester.

Second, I developed a new lecture for that week that I called “Learning in 2026.” The lecture covered studies on how using the internet or generative AI can generate illusions of understanding and impair learning. Because my course focuses on cognitive psychology, I was able to use this material to teach core course content related to memory encoding, memory retrieval, and metacognition.

Third, I incorporated some readings on generative AI throughout the course. For example, the week on problem solving and creativity included papers on how using generative AI can affect the creativity of individuals and groups.

My hope in covering this material was to empower students to make better decisions about their own AI use. The benefits of generative AI for students are often obvious: it can make things easy and efficient. But many people are less aware of the costs. Human learning and memory often benefit from friction rather than ease; we learn better when we engage in hard cognitive work, and not by making things as mentally easy for ourselves as possible. (For those who want to learn more, I discuss this in my forthcoming book, Why We Ask Why: The Science of Explanation and the Human Drive to Understand).

MF: Did you change any of your assignments or exams in response to AI, and if so how?

TL: I changed my assignments in several ways. I have mixed feelings about some of the changes—I think this will be an evolving experiment and I’m eager to learn from others’ successes.

First, I tried a writing assignment that also included a short oral presentation component (to be delivered without slides—just a notecard with notes if desired). The goal was not only to foster an important oral communication skill for presenters (and listening and question-asking skills for other students), but also to assess students in a way that’s harder to outsource to generative AI.

Second, I had more in-class assessments. This included short (low stakes) quizzes in precepts, as well as an in-class midterm and final exam.

Third, I weighted the course grade somewhat more heavily towards the in-class midterm and final exam than I have in the past. I will be reevaluating this moving forward—I think the process of writing is incredibly important for learning, and I don’t like the idea of reducing the quantity of student writing or downweighting its value relative to in-class assessments such as multiple choice tests. 

Faculty Workshop: Supporting Students to Think Critically about AI Use

How should we talk to students about their AI use–in our classes and beyond? How do we set realistic and precise AI policies for our assignments that align with our course objectives and expectations for student learning? In this workshop, Thea Goldring, Lecturer in the Writing Program, will take us through an exercise she uses in her classes and invite us to reflect on how we might adapt it to our own context.

September 29, 2026

12:15 pm – 1:00 pm

Faculty Discussion: Administering Exams in the Age of AI

Are you considering administering in-class, handwritten exams this semester? Thinking about instituting oral exams? How might such assessments align with your learning goals? What are some potential pitfalls of these formats? What new pedagogical possibilities might they offer? 

October 14, 2026

12:15 pm – 1:00 pm

Faculty Workshop: Designing Reading Interventions for Your Course(s) Part 1

What knowledge, strategies, habits of mind and skills do students need to read our assigned texts and learn from them effectively? How can we help students read in the ways, for the uses, and to the standards we expect? In this session, we will guide faculty through a research-based framework for developing their own pedagogical interventions.

October 27, 2026

3:30 pm – 4:45 pm

Faculty Workshop: Designing Reading Interventions for Your Course(s) Part 2

As instructors, we select texts, excerpts, and passages of particular importance to achieving our course aims. Students, however, are not always aligned with, sensitive to, or skilled at reading these texts analytically, critically or synthetically (with other assigned texts). In this session, we will workshop critical and analytical reading interventions for specific tasks and texts. Light refreshments will be served.

November 17, 2026

3:30 pm – 4:45 pm