Think about the last hard thing you worked out at work. Not looked up. Worked out.
You had to piece it together from a few conversations, some history, and a hunch you tested. It took a while. At the end you knew something the organisation had not known that morning.
Now ask where that went. Probably into the decision you needed it for, and then nowhere. The person who has the same problem in eight months will do the whole thing again from the start.
That has always been wasteful. Recent economic work argues it is now something sharper than waste.
The formal version of the problem
Daron Acemoglu, Dingwen Kong and Asuman Ozdaglar published AI, Human Cognition and Knowledge Collapse in February 2026. It is a working paper rather than a peer-reviewed publication, so the model is open to challenge, and challenge has already arrived. The structure is what matters here.
Their argument runs like this. Good decisions need two kinds of knowledge and they are complements, not substitutes. There is general knowledge, which the community holds. And there is context-specific knowledge, which the individual holds about their particular circumstances. Neither works well without the other.
When a person puts in the effort to learn something, that effort produces two things. It produces a private signal, useful to them immediately. And it produces a thin public signal, which accumulates over time into the community's stock of general knowledge.
The second output is a by-product. Nobody does the work in order to produce it. But it is the mechanism by which the general stock stays alive.
Agentic AI substitutes for the effort. And when the effort goes, the by-product goes with it.
The result is uncomfortable precisely because it is not a story about the technology being bad. Contemporaneous decision quality can improve. Each individual is better off. Meanwhile the learning that sustained the shared stock stops happening. Past a certain point the model tips into what the authors call a knowledge-collapse steady state. General knowledge eventually vanishes, while everyone continues to receive high-quality personalised advice.
Every individual decision to draw rather than add was the right one. The stock fell anyway. Leading to a knowledge-collapse steady state.
Why calling it a commons changes what you do about it
Maher Kallel and Mohamed El Louadi took that model apart in a paper posted in July 2026. It is a preprint and the authors themselves label it work in progress, so read it as a serious critique rather than a finished result.
They raise five structural objections. The model freezes the taxonomy of knowledge and ignores the new categories AI itself generates. Its assumption that AI only recycles rather than creates general knowledge is hard to square with systems like AlphaFold. The parameter that decides whether collapse happens, how sharply humans cut their effort as AI improves, is never measured and almost certainly varies by domain. There is a long history of cognitive alarm, from writing to printing to the calculator, in which societies adapted institutionally rather than collapsing. And the proposed remedy of deliberately capping AI accuracy requires international coordination no developer has any reason to accept.
Then they do something more useful than dismissing it. They keep the part that survives and rename it.
What Acemoglu and colleagues identified, they argue, is a credible negative externality on a shared resource. Which makes this a tragedy of the cognitive commons: a classic systems problem with a well-understood shape. Every individual actor behaves rationally. The shared resource depletes anyway. No malice, no error, no villain.
That reframing matters because commons problems have a known solution space. You do not fix a commons by asking people to be less rational. You fix it by changing what the structure rewards, or by building the mechanism that lets contribution accumulate. Kallel and El Louadi land on the second: the policy lever is better aggregation of validated human knowledge.
The same externality is running inside your organisation
Scale the model down from an economy to a company and every term still applies.
An analyst spends two days understanding why a market segment is behaving oddly. The private return is high: they answer their question. The public return is whatever they wrote in the deck, which is the conclusion, not the understanding. The reasoning, the discarded hypotheses, the thing that surprised them, the source they now trust and the one they no longer do, all of that stays with them and leaves when they do.
Now give that analyst a capable assistant. They get to the answer in two hours instead of two days. Better for them, better for the deadline, and the by-product that was already thin gets thinner.
Nobody in this story is doing anything wrong. The incentive structure is the problem. Individuals are measured on output. There is no line anywhere in a performance review for what you added to what the organisation collectively understands. So the private return is priced and the social return is not, which is the definition of a commons under pressure.
The science version of this has just been described in detail. Fulvio Castellacci and seven co-authors, in a paper posted in July 2026 from a roundtable at the ‘AI for Science and Innovation Workshop’ at IMT Lucca in April 2026, examine generative AI in scientific research. It is a preprint and not peer reviewed. Their headline observation is that the productivity gains do not survive disaggregation: AI-assisted work shows measurable gains in publication volume and citation share, while the evidence on novelty, disruption and breakthrough output remains ambiguous or negative.
More of the countable thing. Not more of the thing that mattered.
They identify three mechanisms driving the gap between private and social returns, and each one has an exact organisational twin. Information asymmetry, where others cannot tell how a result was produced or how much to trust it. Negative externalities on a shared knowledge base, where individually sensible use degrades what everyone draws on. And depletion of research capacity, where the skills that were built by doing the work stop being built.
To see what your organisation can come to understand on a live question, rather than what it has filed, talk to us.
The remedy the research points to is not restraint
Here is the part most commentary on knowledge collapse misses.
Acemoglu, Kong and Ozdaglar find that welfare is non-monotone in AI accuracy, which implies some interior optimum and leads to the awkward suggestion of capping precision. Kallel and El Louadi are right that this is politically unworkable.
But the same model contains a second result that is not awkward at all. Greater aggregation capacity for general knowledge unambiguously raises both welfare and resilience. Not conditionally. Not up to a threshold. Unambiguously.
That is a different instruction. It does not ask anyone to use less AI, or to work more slowly, or to coordinate with competitors. It says: build the mechanism that captures human understanding into a shared stock, and the collapse dynamic weakens regardless of how good the models get.
Castellacci and colleagues arrive somewhere compatible from the governance side. Their proposed framework, Responsible Research with AI, rests on four principles. Disclosure, meaning declaring where and how intensively AI was used. Differentiation, meaning that guidance should vary by career stage so that people still learning are protected from skipping the learning. Narrative, meaning an accurate shared account that AI redistributes work and expands verification rather than reducing total labour. And proportionality, meaning requirements should scale with what is at stake.
Translate those four into an organisation and they are practical immediately. Say where the machine did the work. Protect the people whose judgement is still forming. Stop telling a story about AI that your own evidence does not support. Put the effort where the consequences are.
What an aggregation mechanism has to do
Aggregation is the word the research keeps arriving at, and it is easy to hear as storage. It is not. A document repository is aggregation of files. What is needed is aggregation of understanding, which has harder requirements.
It has to capture the ground, not just the conclusion. A finding is worth very little without knowing whether it rests on research, on expert fact, on lived experience, on values or on gut feel. Two people can state the same thing from completely different places and only one of them is evidence for what you are about to do.
It has to hold the relationships. The most valuable object is rarely a single contribution. It is one person's input changing another's, and that chain has to survive rather than being summarised into a paragraph.
It has to keep disagreement. A shared stock that has been cleaned into consensus has thrown away the information about where the question is still open.
It has to make thinness visible. Knowing where the collective picture rests on one person's word is worth more than any individual entry in it.
And it has to be somewhere people are actually contributing to, which means the act of adding has to be part of doing the work rather than an administrative task afterwards.
That is what Hunome is. Each Spark carries its knowtype and other characterisations. A SparkMap holds the perspectives, their grounds and the connections between them, so understanding accumulates across rounds instead of resetting. The Shared Understanding Index measures whether people are building understanding together rather than whether they agree. The output is deliberative intelligence: what the organisation now understands, how it got there, and where it is still forming.
The stock is not a metaphor
Most organisations do not have a shared knowledge stock. They have documents, which are conclusions with the reasoning removed, and they have people, who hold the reasoning and will leave.
That was survivable when learning by doing quietly replenished the stock in the background. The recent work is a warning that the background process is weakening, and that it weakens fastest exactly where AI is most useful.
Nobody has to argue against the technology to take this seriously. The argument is about what happens to the shared thing while everyone rationally optimises the private thing. The research says the lever is aggregation. Aggregation, however, is not enough. A pile of contributions is not understanding. What matters is what happens to them once they are together: whether the reasoning stays attached, whether people can see where the others stand, and whether the whole can be revisited when conditions change.