We are weeks away from our next academic session.
One of our main purposes is to ensure that learning happens effectively. Classroom preparation, participation in classrooms, and the student engagement process all need to maintain the sanctity of learning. The use of AI has to be guided; it cannot be left to students alone.
Before we set the question papers and design the internal continuous assessments, it is important that we pause on a question that sits underneath the very purpose of education. When a student finishes an examination or submits their ICA, what exactly are we measuring? Their learning, or something that only looks like learning?
A study published in Science on 21 May 2026 makes this question difficult to avoid. Igor Chirikov of the University of California, Berkeley, and his colleagues analysed survey responses from more than 95,000 students across twenty major public research universities in the United States, gathered during the 2023 to 2024 academic year. Their finding is direct.
- Roughly two-thirds of students reported using generative AI over the study period, and 37 per cent reported using it regularly.
- Usage varied sharply by discipline. 62 per cent of computer science students reported regular use, against 24 per cent in the arts.
- Business and economics showed high adoption. Patterns of use varied too, with economics at 17 per cent and journalism at 16 per cent among the higher rates, and biology among the lowest at 5 per cent.
Like the authors, we should not treat this as a passing, curious statistic. They argue that “Generative AI calls for assessment reform in higher education”, and they observe that the technology is “making common forms of evaluation, such as tests, projects, or term papers, less reliable as a measure of student capability”.
This is a measurement issue and not cheating. And rightly so, in my view. So this changes the frame of reference. The instrument we have trusted for decades, the term paper, the take-home assignment, the project report, may no longer be measuring the thing we believe it measures.
In countries like India, we have a larger issue of demographic differences and unequal access to AI tools and to AI literacy. So this is not one problem with one answer. It is a different problem in every classroom and institution across India.
Long-time readers of my notes will recognise that we have been circling this question for years, well before generative AI arrived. In one of my notes, I remember discussing two empirical distributions of assignment marks side by side, Course A and Course B. In Course A, every student scored between 44 and 57. The marks bunched tightly near the top. In Course B, the marks spread from 22 to 58. We asked them what this difference suggests.
The answer was uncomfortable. A distribution that collapses towards the top is rarely evidence that everyone learned well. More often, it is evidence that the examination tested only recall and recognition, or that some students carried an unfair advantage of using AI, and the assessment quietly stopped discriminating between the student who understood and the student who did not. When that happens, our assurance of the learning process will cheerfully report that every course outcome was achieved. But the impact is not favourable.
Generative AI is the same threat wearing new clothes. If our ICA asks a student to summarise, to describe, to define, or to reproduce a standard answer, a capable model will now do that work in seconds, and the resulting distribution will look reassuringly healthy. We will have measured access to a tool, not the learning.
We have a name for this gap already. It is the gap between an examination that reveals learning and one that merely produces marks.
There is a second aspect we should not ignore, and that is the question of the relevance of rubrics when AI is being used. A student who cannot see the basis of a judgement cannot learn from it. In an age where the student can ask a machine for an answer instantly, an opaque assessment that returns only a number is the least defensible position we can hold.
So what can we actually do on Monday morning?
I do not believe the answer is to hunt for an AI-proof examination. The Berkeley authors themselves conclude that “there is no single AI-proof assessment model” and recommend reforms tailored to individual disciplines. I would put it this way. The defence is not detection. The defence is design.
Begin with one course. Take its end-term paper and its ICA and ask a single question of every component: if a student used a capable AI model, would this task still distinguish the one who learned from the one who did not? The honest answer for many of our knowledge-and-comprehension questions will be no. Then redesign towards the higher dimensions of Bloom’s Taxonomy, towards application, analysis, evaluation, and creation. Ask students to defend a judgement (perhaps by making a viva mandatory), to apply a concept to a context they have not seen, to critique a flawed argument, to build something from their own reasoning. These are the dimensions a model cannot fake on the student’s behalf without the student understanding the work.
You cannot strengthen student learning with AI if the faculty are not equipped. The first step, therefore, is to form an Anchor Group that conducts an AI readiness survey and draws up a faculty development plan. No initiative can be implemented without first building the competency to carry it out. Once that is in place, we form faculty groups to build an inventory of use cases, pilot them, and share them so that sustainable practices spread across the institution.
What we are dealing with here is a design problem. While developing the inventory of use cases, keep the following in mind.
- Begin from the first principle: learning is a change in the mind produced by productive struggle. Do not let AI remove the very effort that creates the learning.
- Bring AI into a task only where it acts as a more knowledgeable partner that the student grows beyond. If a teacher cannot design that scaffolded environment, the student is better served by not using the tool.
- Have students formulate and scope the problem first, then select and evaluate the tool, interact with it critically, and reflect. Problem formulation before prompting keeps the student thinking.
- Train students to tell apart results that are useful, accurate, fabricated or inaccurate, and to watch for the bias a tool inherits from its training data. Information literacy, critical thinking, problem-solving, adaptability, and data and AI literacy all matter here.
- Move ICAs and examinations up Bloom’s levels, to application, analysis, creation and judgement, so that the paper still reveals learning. Ask one question every time: would this assessment still tell apart a student who learned from one who leaned on AI?
- AI tools trained on a hierarchical reading of Bloom may impose rigid hierarchies. In real situations, thinking does not work that way. The thinking skills fire together and feed one another in an interrelated, cognitive-web fashion. Design your critical-thinking tasks with that web in view rather than a ladder.
I would like to hear how you are dealing with this in your own courses and programmes. And if you would like to know how to put any of these points into practice, do write to me.