Resistance of data centers is now a fixture of American politics. A March 2026 Gallup poll found that 71 percent of Americans opposed having one built near them, and Brookings reports that local opposition blocked or delayed 75 projects worth $130 billion in planned construction during the first three months of this year.
Locally, those fights are mostly about what people say they’re about, including water, noise, transmission lines, power bills, and what happens to the house next door.
But most Americans don’t have a data center going up nearby. The broader hostility seems to reflect a desire to slow AI down or maybe even make it go away. Data centers give people a visible place to do that because they have street addresses, power meters, and permitting hearings they can attend.
I don’t think compute is the most effective constraint.
In AI Wisdom Volume 1: Meta-Principles of Thinking and Learning, I wrote about the odd relationship between constraints and creativity. If you have a huge closet, you don’t have to think very hard about organization. If space is at a premium you get creative about each square inch. A 2022 meta-analysis of 111 studies found an overall positive relationship between constraints and creative performance, although the effect depended on the circumstances.
Compute is a type of constraint like closet space. Every AI company has an incentive to become more efficient because efficiency saves money. Yet a severe constraint creates much more pressure to try something very different. U.S. restrictions on China’s access to advanced AI chips have been porous, but they have still imposed a substantial limit to approaches.
I think there is a more important constraint to AI progress. AI has learned almost everything it knows about physical and social life from records people made of the world. Cameras, microphones, smart glasses, and eventually robots can give it access to much richer experience. If we want some control over how quickly AI acquires certain capabilities, who gets to collect that experience and what they are allowed to do with it afterward may matter more in the long run than how many megawatts are being pulled from a power grid.
What AI Has Never Had
Writing is a compressed description of experience, and a lot disappears in the compression.
Think about a meeting. Someone makes a point that lands wrong and the room shifts. The boss walks in already annoyed, and everyone quietly changes how they behave. Three people agree to something, but only two of the agreements are real. The quiet person may be bored, lost, intimidated, or the only person who has actually understood the problem.
Management books, case studies, and novels have covered all of those things at length. Reading them doesn’t give you the same judgment as repeatedly being in rooms where the outcome mattered to you.
Text also usually describes an event after it happened. Being in the negotiation means deciding what to say while you still don’t know what will happen, watching the other person react, and deciding what to do next. If you push too hard, you may learn that from the expression across the table, the offer that disappears, or the conversation with your boss afterward.
Experience includes that sequence of decisions and consequences. A model can be told that pushing harder was a bad choice and rewarded during training for giving better advice the next time. It did not make the choice itself and then have to continue in the situation its choice helped create.
That is a substantial part of what people mean when they say AI lacks lived experience or common sense. Human judgment contains a great deal of knowledge that is difficult to put into words because we didn’t acquire it through words in the first place.
We Are Recording More of the Missing Pieces
Microphones and cameras are changing AI’s naivete. Zooms and meetings that once disappeared when everyone left the room increasingly are records that can be stored, searched, summarized, and potentially learned from.
Video adds information that a transcript misses. First-person recording is especially interesting because it captures the world from roughly the place where someone is acting in it. Meta’s Ego4D project collected more than 3,000 hours of first-person video showing ordinary activities in homes, workplaces, and other settings, specifically to help AI systems learn from events unfolding over time.
Watching still isn’t the same as deciding. A video usually doesn’t tell you which options someone considered and rejected, what people were trying to accomplish, or how much the outcome mattered. A camera doesn’t have to decide what to do next, while a robot or a person does.
If a warehouse robot reaches for one box and knocks three others off a stack, it now has to operate in the situation its own action created. Google DeepMind’s AutoRT project showed how quickly this can become a data-collection process at scale, gathering 77,000 real-world robot trials across 6,650 tasks.
Robots eventually need experience in places that don’t look like robotics laboratories. A recent San Francisco lawsuit made that unusually visible. A homeowner alleges that employees of The Bot Company rented his Airbnb without disclosing that they intended to test household robots there and left more than $12,000 in damage. The San Francisco Standard found negative reviews from at least 12 other Airbnb hosts connected to three of the guests, including reports of damage to cabinets, furniture, walls, and floors. A household robot eventually has to deal with actual households, which have different cabinets, furniture, clutter, appliances, floor plans, lighting, and all the other peculiarities a lab can’t reproduce indefinitely.
Once people buy these machines, access gets much easier. A robot may be there to fold laundry, load a dishwasher, or carry something from one room to another. Sensing the environment is necessary for it to do the job, and in the process it can collect much more information than it needs for the immediate task.
The Hive Mind
Most of what you learn through experience stays inside you.
That is part of why I got involved in AI and education and why I keep writing books in an era when fewer people read them. I’ve accumulated things over the years that I think are useful, and I want to get as much of them out of my head as I can. A book is one way of leaving some of that behind, but it’s a pretty lousy transfer mechanism compared with what digital intelligence can do.
Nobel Laureate and AI pioneer Geoffrey Hinton has made a great deal of this advantage of digital intelligence. Many copies of the same model can have different experiences and then share what they learn. As Hinton points out, if he learns quantum mechanics and wants someone else to know it, he has to teach them. He can’t copy what changed in his brain into theirs. Copies of a digital model can effectively do something much closer to that.
That hive-mind property changes the value of experience. I can’t add your forty years of experience to mine and walk away with eighty years. A fleet of machines can collect failures, corrections, demonstrations, and successes across many places and use them to improve later versions of all of them.
Human intervention makes some of the failures especially valuable. When a robot gets stuck and a person has to take over, the weakness is readily identified and the human supplies training data to shore up the knowledge. Multiply that across thousands or eventually millions of machines encountering different homes, warehouses, offices, people, objects, accidents, and weird exceptions, and the experience gap can close much faster than human experience ever spreads.
Compute restrictions can slow AI development, but compute is also something companies have enormous financial incentives to control. Tighten the constraint too much and you also increase the payoff for finding another approach.
Real-world experience has a different bottleneck. If an AI system is going to learn from what happens inside homes, factories, offices, hospitals, or schools, it needs access to those places. Someone has to permit the recording, wear the glasses, buy the robot, or allow the resulting information to be kept.
Then we get to decide whether data collected for one reason can quietly be used for another. A meeting transcript exists because somebody wanted meeting notes. A factory robot observes a worker because the worker is showing it what to do. That doesn’t mean the resulting information should be allowed to become part of a general training set used to make every future AI system more capable.
Most of the political fight is aimed at the giant buildings drawing loads of electricity. I think we should pay at least as much attention to who gets to collect real-world experience, how long they get to keep it, and what they’re allowed to do with it.
AI has never lived in the world. It has mostly been trained on records people made of it. We’re now building machines that can make much richer records of their own, act on what they encounter, and pool what they learn.
Whether all of that becomes training data is still something we can control. But the opportunity window is probably measured in months.
©2026 Dasey Consulting LLC
This article was written with the assistance of AI. The typical way in which I apply AI to the process is detailed in a prior article.