Op-Ed: The New Academic Bargain in the Age of AI

Inside Berkeley IEOR’s emerging classroom responses

By Alberto Gennaro (PhD IEOR ’26), Grace He (3rd Year PhD), Ricky Huang (3rd Year PhD), and Jessica Zhao (1st Year PhD).

For generations, college has relied on a simple premise: the work students turn in is evidence of their efforts and thinking. Essays show how they argue, problem sets show how they reason, and coding assignments show how they translate ideas into working computational tools. Artificial intelligence (AI) has unsettled this premise almost overnight. Across domains and assignment types, students can now generate polished responses without wrestling with and grasping core topics, and the results can be indistinguishable from real understanding. This is the central problem AI poses to higher education: not simply that some students might “cheat” by passing AI-generated content as their own work, but that traditional assignments no longer reliably demonstrate what students know. This problem is especially sharp in fields like Industrial Engineering and Operations Research (IEOR), where the goal is not just to compute an answer, but to decide how a messy real-world system should be modeled in the first place. And critically, this is not a problem that universities can prepare for at their own pace. AI is already woven into how students study, write, code, debug, summarize readings, and check their answers, whether or not a syllabus explicitly permits it. Recent U.S. survey data suggest that this shift is already widespread: a 2026 Lumina Foundation-Gallup study found that 57% of U.S. college students use AI in their coursework at least weekly, including about one in five who use it daily. The reality is that AI has already become part of the academic workflow; as such, it cannot be treated by higher education as an outside threat.

This leaves universities at an uncomfortable middle ground. Banning AI outright is too simple: it assumes students will not use tools that are already built into how they study, write, code, debug, translate, and check their work, and it prepares them poorly for workplaces where AI use is quickly becoming standard. But unrestricted use is just as naive. If students can rely on AI to formulate arguments, write code, solve problems, or explain results, then the tools can replace the very struggles through which learning happens. UC Berkeley’s Research, Teaching, and Learning guidance reflects this tension. Rather than treating AI as a problem that can be solved by detection alone, it encourages instructors to clarify expectations, connect AI policies to course goals, emphasize reasoning and course-specific thinking, and consider formats such as open-resource assessments, oral or written follow-ups, and in-person work. UC Berkeley also notes that no single technical approach can fully prevent AI use or access to outside resources, which makes the broader point clear: the goal should not be to pretend AI can be kept out of college, nor to treat every AI-generated answer as acceptable, but to draw a sharper line between using AI to support thinking and using AI to avoid thinking.

In practice, however, this nuance is difficult to pinpoint and implement because much of the burden falls on individual instructors. Professors are now expected not only to teach their subject, but also to decide what kinds of AI use count as learning, what kinds count as misconduct, how to rewrite assignments, how to detect misleading or fabricated AI output, and how to handle disputes when the boundary is unclear. The result is a patchwork of policies that can vary widely from one course to another, even within the same department. Berkeley IEOR offers a useful example of this emerging landscape, where faculty are not responding to AI with one unified rule, but with a range of approaches shaped by what each course is trying to teach.

In one graduate IEOR course, the AI policy takes a deliberately permissive but structured approach. Students are allowed to use AI tools for homework, coding, and report writing, but the syllabus makes clear that AI should supplement, rather than replace, their own reasoning. The policy requires students to critically evaluate AI-generated content, verify its accuracy, and disclose how AI was used, including the tool, its purpose, and any significant modifications made to the output. This reflects a broader educational philosophy: AI use is not treated as inherently dishonest, but students are still responsible for understanding, checking, and owning the work they submit.

In another graduate course, the final project asks students to build a personal knowledge-base system using a large language model (LLM), making AI not only a tool students may use, but also an object of study and design. The course policies permit AI use on assignments and projects, but with disclosure: students must state when and how they used AI through a link to the full AI chat, and are still expected to write their own submissions. The syllabus also includes generative AI, language models, and image generation models as course topics, placing AI inside the technical content of simulation and data modeling. Instead of treating AI as a fundamental threat to academic integrity, the course folds it into both the methods students are learning and the systems they are asked to create.

Another IEOR example goes even further by changing the role AI plays in the classroom. Rather than using AI as a tutor that gives students answers, Assistant Teaching Professor Daniel Pirutinsky is piloting PupilBot, a custom AI agent designed to act like a confused novice learner. The idea is based on the “protégé effect”: students often understand material more deeply when they have to teach it to someone else. In practice, PupilBot asks clarifying questions, expresses confusion, and resists giving direct solutions, forcing students to slow down and explain their reasoning step by step. This flips the usual fear about classroom AI. Instead of letting students outsource thinking to a chatbot, the tool is designed to make students do more of the thinking themselves: defining assumptions, articulating logic, and confronting gaps in their understanding. Pirutinsky’s pilot suggests one possible future for AI in education, where the goal is not to make assignments easier, but to use AI to create new forms of active learning that are harder to fake.

These different approaches do not resolve every problem AI brings to college education. They instead show that colleges can no longer rely on traditional assignments alone to prove what students know. Even thoughtful AI policies must confront questions of equity, privacy, dependence, and consistency. Some students can afford stronger models and better AI-enhanced study tools, while others cannot; at the same time, banning AI entirely can punish students who use it for accessibility support, language assistance, or individualized tutoring. There are also harder questions about what happens when students become too dependent on AI, when private course materials are entered into commercial platforms, and when policies vary so much across classes that students are left guessing what counts as acceptable help. But these challenges are exactly why universities need a more serious response than either panic or surrender. AI has made answers easier to produce, but that only makes judgment, explanation, and intellectual responsibility more important. The task for college education is not to preserve a world before AI, but to build one in which students learn to use powerful tools without losing the habits of thought those tools are supposed to extend.