A new MIT paper should make every MBA student rethink how they’re using AI — and what they’re quietly losing in the process.
There is a particular kind of confidence that fills a business school classroom. Students who have read the case, run the numbers, stress-tested the assumptions. It feels like knowledge. And for now, increasingly, some of it is borrowed.
A working paper from three MIT economists — Daron Acemoglu, Dingwen Kong, and Asuman Ozdaglar — published in February 2026 doesn’t target MBA students specifically. It targets everyone. But its central warning lands with unusual sharpness for a generation being educated at the precise moment when agentic AI can draft your memo, model your valuation, summarise your case pack, and generate your talking points before the professor has finished calling the roll.
The paper’s argument, stripped of its mathematics, goes like this: human beings learn two kinds of things simultaneously. They learn general knowledge — frameworks, principles, accumulated wisdom about how the world works — and they learn context-specific knowledge, the particulars of their situation. These two types of knowledge are complements, not substitutes. Neither is worth much without the other. A consultant who knows every restructuring framework ever devised but cannot read a specific client’s political dynamics will give textbook advice that destroys value. A consultant who reads the room brilliantly but has no frameworks is just an expensive conversationalist.
What makes the paper dangerous reading is what it says about effort. When humans grapple with hard problems, they produce both types of knowledge at once. The struggle is the point. And crucially, the general knowledge produced in that struggle doesn’t just stay with the individual — it leaks out. It becomes the Stack Overflow answer someone else reads at 2am. The peer-reviewed insight that reshapes a field. The institutional memory that onboards the next cohort. It is, in the economists’ language, an externality that the individual doesn’t capture but the community needs.
Agentic AI — the kind that gives you a personalised recommendation rather than a general explanation — short-circuits this process neatly. It hands you the context-specific answer directly. You stop struggling. The externality stops flowing. And over time, in the paper’s formal model, the stock of general knowledge that the whole community depends on begins to erode. In extreme cases it collapses entirely, rendering even the AI’s recommendations useless, because personalised advice requires a foundation of shared knowledge to mean anything at all.
The authors call this “knowledge collapse.” The math is rigorous. The real-world evidence they cite is suggestive: Stack Overflow has seen declining engagement since the widespread adoption of generative AI tools. Wikipedia contributions have fallen in areas where ChatGPT is an effective substitute. These are not trivial data points. They are early readings on what the paper models as a tipping-point dynamic.
Now, where should a thoughtful MBA student push back? The paper has genuine limitations. It treats AI as purely substituting for human effort, but powerful tools have historically *expanded* what humans attempt — the spreadsheet didn’t make analysts lazy, it made sophisticated modelling worth doing. The model also assumes relatively weak institutions, but professional licensing, accreditation, and formal education exist precisely to mandate learning even when personal incentives erode. And the paper’s own extension acknowledges that AI-generated synthetic data can prevent full collapse in domains where outputs are verifiable. The apocalyptic scenario requires a specific set of conditions that may not hold universally.
But here is what should concern you regardless: the paper’s predictions about *relative value* are robust even if full collapse never arrives. As agentic AI handles more context-specific execution, the people who understand the underlying principles — who can interrogate an AI’s output, identify where the model is wrong, synthesise across domains, and make judgments in genuinely novel situations — become scarcer and therefore more valuable. The people who have outsourced that understanding become dependent and interchangeable.
This is not a technological prediction. It is a structural one, and it applies directly to how you are spending the next two years.
Three things follow. First, the tasks most worth doing in business school are the ones that feel inefficient: arguing through a case without looking at the model answer, drafting a memo before asking AI to improve it, sitting with an ambiguous problem long enough to develop an actual view. Not because AI assistance is cheating, but because the struggle is where the general knowledge forms. Second, the most defensible career positions are those of aggregator and interpreter — roles that connect knowledge across disciplines, geographies, and functions in ways that require human judgment at the synthesis layer. Third, and most practically, if you cannot explain why your AI-generated output is correct, you do not actually know what you know.
The paper’s formal recommendation is that regulators should sometimes deliberately degrade AI’s precision to force human learning. You do not need to wait for a regulator. You can degrade it yourself — selectively, strategically, and in full awareness that the cognitive infrastructure you’re maintaining is not just a private asset. It belongs to the room you will one day be the smartest person in.
*Based on “AI, Human Cognition and Knowledge Collapse” — Acemoglu, Kong & Ozdaglar, MIT, February 2026.*
Stay connected with us to explore endless opportunities at Praxis Business School!
Visit our website at https://praxis.ac.in/ to learn more about our programs, admissions, and campus life. For any queries, feel free to reach out to us at https://praxis.ac.in/contact-us.
Follow us for the latest updates, insights, and success stories.
We look forward to connecting with you!