Can We Describe the AI Revolution While We Are Still Inside It?

Can we describe a transformation honestly while we are still standing inside it?

---

Wooden walls

With the Persian army closing in, the Athenians sent envoys to the oracle at Delphi.

The priestess gave them a dark answer: the city would be destroyed, and safety would be found only behind wooden walls.

Athens fell into argument. Some read the words literally and pointed to the wooden palisade ringing the Acropolis, which was reasonable enough, since that was the first thing anyone thought of when they heard the phrase. Themistocles read it differently. The wooden walls, he said, were the timber hulls of the fleet. He persuaded the city to abandon the mainland and board the ships.

At Salamis in 480 BC, that light, quick wooden fleet broke the Persian navy apart.

The interpretation turned out to be more accurate than the prophecy.

We are in the same position now. We have information, and information alone gives no direction. The same sentence yields two different futures, and we cannot say which one is real.

That is exactly where artificial intelligence leaves us. There are numbers. There are reports. There are forecasts. What is missing is the reading that wins the battle.

We cannot measure ourselves from inside

The first problem with living inside an AI revolution is not emotional. It is a measurement problem. The fish is the last to notice the water.

In 2025, a research group called METR ran an experiment that looked almost too simple. They gave experienced open-source developers real tasks in repositories they already knew well, some with AI tools permitted, some without.

The result: the developers using AI tools finished 19 percent slower.

That is not the striking part. The striking part is what those same developers said afterward. On average, they estimated that AI had made them about 20 percent faster.

So they slowed down and felt themselves speeding up. Roughly forty points of daylight between perception and stopwatch.

The limits deserve honest statement. Sixteen developers, 246 tasks, tools as they stood in early 2025, mostly Cursor Pro running Claude 3.5 and 3.7 Sonnet. METR itself now treats the finding as historical and says it does not represent today’s tools. And it does not mean AI “doesn’t work.”

What it means is narrower and more unsettling. When you are inside a technology, you cannot measure what that technology is doing to you by how it feels.

Hold onto that finding. Everyone speaking confidently about this revolution is describing by instinct something they have not measured. The person writing this essay included.

Naming always arrives late

Nobody called it the Industrial Revolution while steam engines were filling the mills. An era gets its name after the era is over. The people living through it did not experience a revolution. They experienced a scatter of unrelated pressures and openings: prices, migration, disease, a new economy, cities swelling past their seams.

In 1990 the economic historian Paul David drew a comparison that has never stopped being useful. Electricity arrived in the factories, and the productivity figures sat still for decades. The real gains came roughly thirty to forty years later.

The delay had nothing to do with electricity being weak. Factories swapped the steam engine for an electric one and left the building exactly as it was. Under the old arrangement a single great engine drove everything, with every machine tool strapped by belts to line shafts running along the ceiling. The whole structure had been designed around that one central source of power. Electricity’s actual gift was different: a small motor of its own for every machine. Seeing that, and then rebuilding the floor plan around it, took a generation.

The productivity came from reorganization, not from the technology.

Robert Solow made the same observation about computers in a single line, noting that the computer age was visible everywhere except in the productivity statistics. He wrote that in 1987, and it took another decade before the numbers moved.

Set that beside the METR result. A developer drops a new tool into an old workflow and feels faster, while the measurements say the opposite. The factory that wired an electric motor to a line shaft was living through precisely the same illusion.

The revolution does not happen in the technology. It happens in the organization. And organizations change far more slowly than tools do.

The Collingridge dilemma: name it before it is too late

The third dimension of this is political.

In 1980 David Collingridge described a bind that has remained the most honest sentence in technology policy ever since.

Early in a technology’s life you can shape it easily, but you do not know its effects. By the time the effects are visible you know exactly what you are dealing with, but the technology has settled so deeply into everything that changing it is expensive, difficult and slow.

Collingridge’s dilemma, compressed: when change is cheap, the need for it cannot be foreseen; by the time the need is obvious, change has become costly.

This bind explains both camps in the AI argument at the same time.

One side says we do not know enough yet, so let us not regulate in haste. They are right. The information is genuinely incomplete. The other side says intervene before the window closes. They are also right, because the window really is closing. Both are correct, which is precisely why the argument never resolves.

Collingridge’s own proposal was modest. Move innovation forward in smaller steps. Avoid designs you cannot walk back. Build systems that make errors cheap to notice and cheap to correct.

Melvin Kranzberg put the reason for that modesty into a law: technology is neither good nor bad, and it is not neutral either. It always arrives shaped by the arrangements it enters, and it always reshapes them back.

Translated into today’s terms, the question is not whether AI is good or bad. The question is whether we can reverse the decision to scale it without limit once we have made it.

The person inside the revolution

Underneath the measurement layer and the political layer, something much quieter is happening.

The numbers are serious. A RAND survey found that roughly one in eight American adolescents and young adults had turned to a chatbot for help with feeling sad, anxious or overwhelmed, about 5.4 million people. A follow-up survey put the figure closer to one in five, around 8.2 million. Nearly two thirds of them had told no one they were doing it.

Calling this a case of people being fooled by technology is easy and wrong. The accurate reading is that artificial closeness is filling a vacuum, and AI did not create the vacuum. Human support became expensive and hard to reach long before the chatbots arrived.

Something is still missing, and it is not a better imitation of empathy.

What happens when you speak to a therapist is that the person across from you is affected by you. What they hear changes them. Your silence tells them something. A model listens, records, produces a response, and is not altered by what it has witnessed. Silence is not data to it.

That is the whole difference. The problem is not that the model is insufficiently good. It is that being good is beside the point.

Martin Buber gave this distinction its classic name. In an I-It relation the other is an object, used and measured. In an I-Thou relation something meets something. The hard part of Buber’s claim is this: two sound waves can be identical and one of them is an encounter while the other is not. Being unable to tell them apart does not make them the same.

Joseph Weizenbaum ran into this in 1966. His ELIZA program did almost nothing beyond rephrasing sentences back as questions, and people began confiding in it, some asking to be left alone with it. What alarmed him was not that the machine was clever. It was how little cleverness the substitution required.

Revolutions keep good statistics. What is lost keeps none.

The promise was always free time

Around AD 10, Antipater of Thessalonica wrote an epigram about the water mill, addressing the women who had turned the grindstone by hand. Stop grinding, he tells them, and sleep late, even when the cocks announce the dawn. The nymphs have taken the work of your hands. The first promise of automation was not profit. It was sleep.

In Signs of the Times, published in 1829, Thomas Carlyle attacked the Industrial Revolution without locating the danger in mechanized work. He located it in the rhythm of the machine seeping into the human mind.

On 24 April 1812, at Westhoughton in Lancashire, two sisters stood at the front of the crowd that destroyed a weaving mill. Mary Molyneux was nineteen, Lydia fifteen. Court papers record them at the windows with coal picks, shouting the men on, and the building burned with its looms inside it. They were not afraid of technology. They were losing their living. Both were acquitted.

In 1911 Frederick Winslow Taylor published The Principles of Scientific Management and turned human labor into a process to be timed and optimized.

Set the four side by side and the pattern surfaces. Every time, the promise was free time. Every time, the time saved drained into new work, new measurement, new mechanization.

Keynes made the most confident version of that promise in 1930, telling his readers that their grandchildren would work fifteen-hour weeks. The productivity arrived roughly as he predicted. The fifteen-hour week did not.

The AI revolution we are living in is not outside this cycle. It is fairly clear it will not hand us a citizen’s wage. We will work differently and more, or we will work in a way that produces far more output with AI beside us.

So what does an honest description look like

I will not end with a prophecy, because the prophecy is part of this story. Instead, three measures to keep in hand while standing inside the thing.

Look for the revolution in the organization, not the technology. A new tool dropped into an old arrangement changes nothing except how busy everyone feels. If you want to see where the change actually is, do not look at the tool. Look at how the work is divided.

Do not trust your own sense of it. Measure. This is what the METR finding is for. People can slow down while feeling faster. That is not stupidity. It is that the fluency a new tool gives you is not the same thing as a measurement. The most confident sentence spoken about a revolution is usually the least measured one.

Interrogate whether our AI decisions can be undone. The Collingridge dilemma cannot be solved, but it can be managed. “Is this good or bad” has taken us nowhere for years. “Can we reverse this, and how quickly would we notice the mistake” is a question with an answer.

Themistocles won not because he read the oracle in his own favor, but because he read it correctly. The difference was that he looked past the sentence in his hand to the actual capacity in his hand. Athens had a weak fence and a strong fleet.

The sentences available to us today are ready-made: collapse, salvation, leap, threat. Each of them sounds right, and none of them, on its own, points anywhere.

The question we should be asking is what we actually have, and what we are prepared to lose as a species.

When we put that question to the AI itself, it cannot give a full answer either. Nor should we expect one, because this is a question a human being has to answer alone.