On April 17, 2026, Daron Acemoglu delivered a keynote as part of Universitas Sebelas Maret's 50th Anniversary international webinar, an event that drew more than 1,800 participants across 31 countries. His message was direct. Technology is not neutral. Its impact depends on the systems that deploy it. When those systems are built to include the people, whose work makes the gains possible, innovation reaches more people. When they are built to leave those people out, the gains collect in fewer hands.
That argument is not new in his work. What changed is where it was placed. Artificial intelligence now sits inside that same framework. The question is no longer what the technology can do. The question is who it is designed to benefit and what it draws from the people operating around it.
Two months earlier, Acemoglu and his co-authors formalized a related dynamic in AI, Human Cognition and Knowledge Collapse. The paper shows that when AI substitutes for the human effort that produces knowledge, short-term performance can improve while the underlying knowledge base begins to erode. In plain terms, the organization gets faster at doing the work while losing the understanding that made the work possible in the first place.
The gains are real. The loss is also real. It is harder to see. As AI gets more accurate, the model predicts a point where the outputs still look good but the knowledge that made them possible is already disappearing. The idea has been named by researchers and policy leaders. What is still missing is the version that works inside an actual organization.
The frame the workforce conversation is still running on
Since generative AI entered the workforce conversation, it has been shaped by a simple premise. Younger workers are fluent. More experienced workers are behind. The organizations that win will be the ones that invest in the cohort that grew up with the tools. The strongest available data does not support that premise.
Stanford's Digital Economy Lab found that since the widespread adoption of generative AI, workers aged 22 to 25 in the most AI-exposed occupations experienced a measurable decline in employment, while more experienced workers in those same roles remained stable or continued to grow. Separate studies show a similar split. Younger workers report higher rates of AI use. Mid-career professionals report stronger effectiveness when applying it in real work contexts. Adoption is not the same as execution. Frequency is not the same as value.
The premise has persisted anyway. Not because it is accurate. Because it is useful. It supports training models, hiring strategies, and adoption narratives that are easy to fund and easy to explain. It gives organizations a visible place to direct investment. It gives vendors a clear audience to target. It gives internal leaders a defensible story when adoption numbers are under pressure. The cost of that story is not being measured where it is actually being paid.
The layer the current conversation does not account for
Underneath the generational narrative, organizations have been relying on something that has not been formally named in their AI strategies. It is not absent from their operations. It is simply not visible inside their frameworks. Every functional deployment of AI depends on a layer of human knowledge that already exists inside the organization. Not documented process steps. Not formal training materials. The actual decision logic that determines how work gets done.
It shows up in the parts of the work that never made it into a handbook. The judgment call that keeps a client relationship intact. The reason a process works even though the written version of it is wrong. The decision someone made years ago that everyone still quietly relies on. That kind of knowledge is carried by the people who have been inside the work long enough to recognize it. It does not show up on any org chart. It holds the organization up anyway.
It is not in the model, the system or in the prompt. It is not being measured as part of the organization's AI strategy. This is where the current approach begins to break down.
Most organizations are deploying AI into workflows without first identifying where this layer of knowledge sits, how it is being used, and what happens when it is bypassed. The result is a pattern that is easy to miss in the early stages. Output quality appears to improve. Speed increases. Adoption metrics begin to move.
At the same time, something else begins to shift. The effort required to produce knowledge starts to decline. The system becomes more dependent on what has already been built and less capable of generating new understanding inside the work itself. Decisions continue to be made. The reasoning behind them becomes harder to trace. The organization becomes more efficient at execution and less aware of how its own work actually functions.
This is not a failure of the technology. It is a misalignment in how it is being introduced.
The research explains what happens when knowledge production declines. What it does not yet specify is how organizations can recognize this layer before it erodes, how they can surface it without slowing down adoption, and how they can build AI into it rather than around it. That is the work most organizations have not started. It does not begin with another tool. It begins with a different starting point.
What this means right now, inside organizations already deploying AI, is that the gap between what the adoption metrics are measuring and what the organization is actually losing is growing every quarter. The dashboards look healthy. The outputs are faster. The reports show progress. None of those signals are capturing the drain happening underneath. The organizations that will look back on this period as the one that cost them the most are the ones whose numbers looked best along the way.
If the pattern continues without adjustment, the direction is predictable. The organization becomes more efficient at producing decisions and less capable of explaining why those decisions work. The professionals who carried the judgment begin to move on, retire, or disengage. What they carried is not reproduced because the conditions that produced it are no longer being supported. What remains is an operation that runs fast and cannot be repaired, because the people who knew how to repair it were never part of the strategy.
This is where the research and the operational work meet. The adjustment that changes this direction is specific and it has a sequence. It begins with closing the formation gap. It happens before tools, before training pipelines, and before adoption metrics.
The formation gap is the distance between what the organization is asking its people to do with AI and what those people have actually been given the conditions to become ready for. That gap does not close on its own and it does not close by applying more of what caused it. It closes when the organization meets its people exactly where they are, names what they already carry, and builds the AI strategy forward from that recognition. Every step after that works. No step before that does.
The organizations that will adjust do so because of one recognition, not because of budget, technology stack, or leadership philosophy. It is a single recognition. The human layer is not a cultural concern or a change-management problem. It is infrastructure. It holds weight. It can be mapped, measured and built around. The ones that do not see it will continue to treat it as sentiment and will continue to be surprised when their most capable professionals quietly stop carrying work that was never named as work.
The principle is public. The research is published. The operational translation is now available to any organization that chooses to do the work. What happens from here is a question of recognition, not a question of capability.
To learn more, visit my website at inclus-ai.com.