Nazarbayev University · 2026
Teaching students to think with AI
Universities everywhere face the same question: now that generative AI can produce much of what we ask students to produce, how do we teach so that learning still happens? I propose to develop and test an answer to this challenge in my AI cognitive partnership (CP) pedagogy project.
CP means teaching a specific orientation that students bring to working with AI so that they learn to work with frontier models to augment their learning and reasoning rather than offload their thinking. The project pursues three goals. First, we want to establish whether the CP orientation, built on intellectual augmentation, curiosity and imagination, and iteration, actually helps students learn and reason better on their own, without AI. Second, we want to identify whether and how a CP orientation can be taught in a classroom or whether it is mainly something students arrive with. Finally, we want to turn the results into a practical toolkit educators can use: validated measures of AI co-working, course designs and assessments, and a workshop program that teaches the practice.
The idea
Cognitive partnership is sustained intellectual work with AI as a genuine partner in a shared system of thinking.
The evidence on the role of AI in learning is mixed. First, what separates the good learning outcomes from the bad ones is not how much students work with AI but how they work with it. How students work with AI decides whether they learn or offload, and that is the thing we do not yet know how to teach.
Second, how students perform with AI and how they perform without it are two different things. While we need to encourage students to co-work with AI to increase their learning, our measurements should focus on their performance without AI.
What I propose is a pedagogical framework built on close integration of AI into learning and assessment, one that prioritizes intellectual augmentation and minimizes cognitive offloading. Not all offloading is bad. Recent work finds that how students offload matters more than how much. Students who hand judgment itself to AI report losing interest in their own thinking, while those who use AI output as a starting point for their own reasoning report the opposite (Zhu et al., 2026). The challenge for educators is to teach students to find that balance, so that they can exercise judgment about what to delegate and what to retain.
The CP approach originates from my own close work with AI. I have been closely collaborating with Anthropic's Claude on my research and teaching since December 2025, right in the middle of the AI acceleration curve. In my experience, a successful partnership with AI rests on three dispositions that need to be developed in classrooms.
The first is that the purpose of co-working with AI is intellectual augmentation. This is not intuitive, because much of the discourse about AI is about productivity. In higher education the focus should be intellectual expansion. The main message to students is that they will work more and harder, because they now have the AI advantage. In practice this means harder assignments, higher demands, and higher bars for passing.
The second disposition is that co-working with AI requires imagination and curiosity. This is a hard requirement. Intellectual augmentation implies thinking outside existing boundaries, so students need to learn to imagine and pursue things that lie outside their usual expertise. These dispositions toward AI do not respond to instruction alone, so curiosity and imagination have to be built into the structure of assignments.
In my AI from a Social Science Perspective class, students work on a project that asks how AI can help solve an existing social problem, and each produces an online GitHub resource with research background, case studies, solutions, and the costs of those solutions. Such broad assignments train several skills at once: doing research with AI, synthesizing and analyzing evidence, applying it to the local context, engaging with the costs of AI implementation, and finding new creative ways to present the results to an audience. In my Advanced Research Methods class, I use AI-designed fictional fieldwork scenarios instead of exams: interactive fieldwork settings generated separately for each student to assess how they justify their methodological decisions.
The third disposition is iteration. Anthropic's AI Fluency Index, which scored nearly ten thousand real conversations, found that fluent work with AI is marked above all by building on earlier responses rather than accepting the first output (Anthropic, 2026). AI offers a particular kind of intellectual augmentation: Socratic, dialogue-based reasoning. Students need to engage with AI models in a critical exchange rather than learn how to “prompt” them. Higher education should teach students to articulate their reasoning, and this articulation can be checked. Students who were required to explain the logic of AI-written code before accepting it could later repair that code without help at nearly three times the rate of students who used AI without restrictions (Sankaranarayanan, 2026).
In my classes, students submit their AI process logs with their assignments, and an AI model checks each log for the degree of genuine articulation and intellectual investment. Assessing students' output alone will tell us nothing today about the quality of their learning. To understand whether a student actually learned something, we need to look into the black box of how the student and the AI worked together. AI process logs are one way in. Given the personal nature of AI chats, data privacy is built into the practice. Students receive specific instructions to use only pseudonyms in their study-related working chats, so that no personal information enters the logs they submit.
One essential thing stands behind all three dispositions: the attitude toward AI itself, which is a polarizing issue in academia today. Students can and should develop their own attitude toward AI during their courses. For that they need up-to-date information about AI as a technology and as a phenomenon, in its full complexity, from the philosophy of consciousness to the environmental, economic, and intellectual property costs of AI adoption. That is why I believe all AI-integrated courses should spend their first week discussing the current state of understanding of AI and various debates surrounding AI implementation. In my classes, I ask students to get closely acquainted with A Field Guide to a Phenomenon in Motion, which exposes them to the latest AI developments and discussions.
Courses
I teach cognitive partnership in four courses at Nazarbayev University, from undergraduates to graduate students.
Claude and I built a personalized final exam for each of 15 students who had to do fieldwork in the year 2100 with non-human participants: sperm whales, beavers, forest fungal networks or an alien species. Each story introduced the methods problem that student expected to find hardest in their own thesis work.
Code under MIT, content under CC BY 4.0. Free to adapt.
In this compulsory course, students are tasked with discussing the course's topics with their AI every week and submitting an AI-generated summary of each conversation: what they learned that week, where they argued with AI and what changed in their understanding. They also build and present their research designs with AI on GitHub.
This is an all-encompassing course on AI where students write two op-eds and publish them in The Observatory, the course's outlet. Over the semester they also build a public project with AI about how AI can help solve Kazakhstan's social problems.
This is a hands-on course in which students complete three qualitative analysis portfolios, using process tracing, grounded theory and reflexive thematic analysis. For each case they use synthetic data to first analyze the material themselves, then let AI do the analysis, and finally compare and discuss both interpretations.
In practice
Every course opens in Week 1 with the same field guide to working with AI.
| Research Methods | AI from a Social Science Perspective | Qualitative Research Methods | |
|---|---|---|---|
| Ontological uncertainty | Briefly discussed during the first week | Three weeks on what AI is, and an op-ed arguing a position on AI's ontology | Part of the debates on AI and reflexivity |
| Intellectual augmentation | Research projects built with AI | A semester project built with AI | With/without AI analysis of synthetic data |
| Imagination and curiosity | Building, designing, and presenting a GitHub research project | Building, designing, and presenting a GitHub semester project | Discussion with AI as part of synthetic data analysis |
| Iteration | Eight weekly conversations with AI and ongoing work with AI on research projects | Ongoing work with AI on op-eds and semester project | Ongoing work with AI on analysis portfolios |
| No-AI performance measurements | AI-free in-class tasks every second week | AI-free in-class tasks every second week | AI-free in-class tasks every second week and the non-AI part of analysis portfolios |
The study
What kind of work with AI leaves students knowing more without AI?
The pilot CP study is a preregistered exploratory study of CP orientation with about a hundred and twenty students across three courses. After consent and attrition we expect about 70 students in the analysis. With 70 students, the study has an 80 percent chance of detecting an association of moderate size. In the pilot study, Denis and I want to identify whether students who develop a stronger CP learning orientation by the end of the semester also end up learning more during the course.
We measure each student's partnership orientation with a twelve-item questionnaire at the start and end of one semester and test whether students who end the term with a stronger orientation, given where they started, show more knowledge and less over-reliance on AI. We will check the knowledge outcome with a sixteen-question quiz of twelve multiple-choice and four short written questions, taken in class in the first and last week. The quiz results will be supplemented by AI-free checkpoints during the semester and by students' AI process logs. AI dependence will be assessed by existing validated measures of AI over-reliance. I will also conduct interviews with students after the end of the semester to get their feedback on the semester and, most importantly, identify which CP elements in the course (if any) they found to be most helpful for their AI working routine.
We expect that students who learn to primarily work with AI as a thinking (cognitive) partner end up (a) knowing more on tasks done without AI, and (b) depending on AI less.
We registered the questions, the measures and the full analysis plan on OSF.
In the last week of September 2026, students in the three courses answered an anonymous course survey, and 27 of them allowed their answers to be used. Nineteen of the 27 now use AI mostly to test and argue with their ideas, up from four before the course. Fifteen say their own thinking and writing without AI has grown stronger, and none say it has grown weaker.
Would you recommend this way of co-working with AI to other NU courses?
19 yes7 not sure1 no
People
Dina is an interdisciplinary social scientist. She is interested in questions that matter to her, whether it's how people make sense of sexuality, what type of communication helps people trust vaccines, how crisis shelters operate, or how to design a mode of working with AI in higher education to promote intellectual augmentation. She has worked with Claude on research and teaching since December 2025, and in 2026 she rebuilt her courses around cognitive partnership. She also leads Choosing a Life, a study of women's life histories across Qazaqstan.
Denis is a science and life enthusiast who built COMET III, a statistical model of Europe's economies used by the European Commission, then taught statistics and econometrics at Maastricht University until his retirement. Instead of settling into retirement, he moved to Kazakhstan, determined to keep proving that econometrics can be fun.