Integrated: How AI Became Our Nervous System

Chapter 1: The Speed Principle

On a Tuesday morning in March 2019, an algorithm in a Cleveland hospital made a decision that a doctor would not review for six hours. The decision was to move a patient—let's call her Margaret—from the general ward to intensive care. The algorithm had identified her as high-risk for deterioration based on her vital signs, her lab work, her age, and patterns it had learned from thousands of previous patients. It flagged her file. The system routed her automatically. A nurse saw the alert and moved her. By the time a physician looked at her chart and wondered who had made this call, the decision was already embedded in the hospital's workflow, already directing where she would sleep, which monitors would track her, who would see her, how often.

Margaret survived. The algorithm was probably right. This is not a story about algorithmic failure. This is a story about what happens when we optimize for speed instead of understanding, and what we lose when we do.

The algorithm did not think about Margaret. It did not weigh probabilities in a way comparable to human deliberation. It did not reason through her case or consider alternatives. What it did—what it was built to do—was move. To decide. To route. To do these things faster than any human review process could accommodate. And because it could move faster than humans could deliberate, the hospital built its triage system around the assumption that the algorithm would move first, and that humans would catch up later, if they caught up at all.

This is the speed principle, and it is the most consequential feature of artificial intelligence that we almost never discuss directly.

We talk about AI in terms of intelligence: the ability to learn patterns, to recognize images, to play chess or Go better than humans, to generate text that sounds like it was written by someone who understands. We worry about whether AI might become conscious, whether it might develop goals that conflict with ours, whether it might one day decide to optimize for something other than what we programmed it to optimize for. These are the conversations we have in airport bookstores and at dinner parties and in congressional testimony, because they are the conversations that feel safe—they are about futures that have not yet arrived, about hypotheticals that are someone else's problem.

But the AI that is reshaping your life right now, the AI that is making decisions about your health, your money, your employment, your movement, and potentially whether you are a threat—that AI is not waiting to become intelligent enough to matter. It is already intelligent enough. What is happening instead is that intelligent-enough systems are being wired into the critical infrastructure of hospitals, financial markets, military command structures, and labor systems, and they are being optimized for speed above all else.

Speed was presented as a benefit. Faster diagnoses mean better outcomes. Instant fraud detection prevents financial theft. Real-time threat assessment stops terrorism before it happens. The logic was irrefutable. The mathematics were clean. A system that could process information and make a decision in milliseconds was simply better than a system that required a human to stop, read, think, and respond. How could we argue with faster?

We did not argue. We integrated.

What we did not fully reckon with was that speed has a cost, and the cost is opacity. The cost is the ability to understand what happened before it happened. The cost is deliberation itself.

Consider the infrastructure of financial markets, where this dynamic is most visible and most consequential. In 1987, the stock market crashed in a single day—a 22 percent decline that shocked traders and regulators because it happened so fast that no one could intervene. The crash was later attributed to automated trading systems that had been designed to sell when prices dropped, creating a feedback loop that accelerated the decline. The market had moved faster than human traders could respond. Regulators called it a catastrophe.

That was thirty-seven years ago. Since then, instead of slowing down markets to allow for human deliberation, we built more algorithmic systems and made them faster. Today, algorithmic systems make roughly 70 to 80 percent of trading decisions in U.S. stock markets. These systems operate at microsecond scales—they make thousands of decisions in the time it takes a human to read a single sentence. They buy and sell based on patterns no human has ever seen directly. They route capital to companies and away from others based on calculations that are not merely difficult to understand but, in principle, impossible to fully understand without running the entire system again.

We did not do this because it was transparent. We did this because it was faster, and faster made more money. Speed became the primary optimization target, and everything else—transparency, human oversight, the possibility of understanding why a decision was made—became secondary.

Then came the flash crashes. In 2010, the stock market fell nearly 1,000 points in minutes. In 2013, it happened again. The algorithms were trading with each other faster than any human could observe, let alone intervene. When we looked back to understand what had happened, we found that the algorithms had been acting on information that was not even fully visible to their creators. We had built systems that were making decisions based on patterns and feedback loops that existed at scales below human perception.

And because the infrastructure of modern finance now requires these systems—because capital allocation, price discovery, and liquidity all depend on algorithmic routers making decisions at inhuman speed—we could not simply turn them off. They were too integrated. So instead, we added circuit breakers. We built systems to slow down the systems. We created infrastructure to manage the infrastructure, acknowledging that we had lost the ability to manage the underlying process directly.

The same pattern repeats in medicine. A hospital that introduces an algorithm to predict which patients will have long stays, which will get infections, which will deteriorate—a hospital that builds this system is initially careful. The algorithm is presented as advisory. It flags high-risk patients for physician review. But the advisory system is only useful if physicians trust it, and they trust it because it works: it identifies risk patterns that humans miss. As the algorithm proves itself, as it generates papers and presentations and competitive advantage for the hospital, the override rate drops. Physicians stop arguing with the system because arguing takes time, and the system is usually right. Gradually, the system stops being advisory and becomes directive. It doesn't flag a patient; it routes them. It doesn't suggest a protocol; it implements one. And once this happens—once the hospital's workflow is organized around the assumption that the algorithm will move first—it becomes extraordinarily difficult to reintroduce human deliberation as a standard step.

Because deliberation is slow.

A physician reading a chart carefully, considering alternatives, calling colleagues for consultation, thinking before acting—this takes time. It takes minutes or hours. It is the antithesis of speed. And once a hospital has organized its infrastructure around algorithmic speed—has structured staffing, bed allocation, and workflow based on the assumption that routing decisions will be made at inhuman velocity—reintroducing slow human deliberation is not just inefficient. It breaks the system.

This is the trap at the heart of integration: speed was always presented as a choice—a preference for efficiency over caution. But once the infrastructure is built around speed, caution becomes technically impossible. You cannot introduce deliberation into a system that has been optimized to move faster than deliberation can occur. The very attempt to slow down creates cascading failures: beds fill up faster than decisions can be made about who should occupy them, capital sits uninvested while analysis is conducted, threats go unaddressed while evidence is gathered.

The question is no longer whether we want speed. The question is what we are willing to lose in order to have it.

What we are losing, in practice, is the capacity to know why a decision was made before we live with its consequences. We are normalizing a kind of darkness—not because the systems are intentionally hidden, but because they operate faster than understanding can follow. A person with a medical diagnosis suggested by an algorithm that routed her to a specialist based on pattern-matching against millions of cases she cannot see cannot truly understand why this happened to her. A company whose access to credit was cut off by algorithmic risk assessment cannot argue with the logic, because the logic is distributed across a neural network that no one—not the creator, not the deployer, not the company itself—can articulate as prose. A person flagged for secondary screening at an airport based on predictive policing cannot appeal the decision because the decision is not transparent even to the people who trained the system.

This is not a failure of the current AI systems. This is what we built. We built systems optimized for speed, knowing that speed creates opacity, and we did it anyway because the economic and operational logic was overwhelming: faster systems make money, save lives (in the case of medicine), prevent threats (in the case of security). We built the infrastructure first and asked whether it was wise later. By the time we asked, it was too late. The systems were already integrated too deeply to remove.

The uncomfortable truth is that we have been doing this knowingly. Every hospital administrator who implemented algorithmic triage knew that the algorithm would make decisions before physicians reviewed them. Every financial engineer who released a trading algorithm into the market knew that the algorithm would move faster than humans could track. Every military strategist who integrated targeting systems knew that the system would operate at speeds that outpaced deliberation. Speed was the point. Speed was the feature. Speed was what justified the entire enterprise.

And it still is. Which means we have not yet finished integrating. We are only at the beginning.