Routing Decisions

Chapter 1: The Upgrade

On a Tuesday morning in March, a woman in Portland uploaded her medical history to a new application called ChatGPT Health. She was not sick in any way that required urgency. She had fatigue. Not the crushing exhaustion of depression or the bone-deep tired of chronic illness, but something diffuse—a feeling of operating at eighty percent capacity, a slight dimming of the world that had persisted for three months.

Her family doctor had run bloodwork. Everything came back normal. She had asked if they should investigate further. He suggested sleep hygiene, vitamin D, perhaps a therapist. She left his office with the distinct impression that she was wasting his time.

The application took forty-seven seconds to process her uploaded records. It cross-referenced her symptoms against 340 million clinical observations. It weighted her age, her zip code, her family history, the specific pattern of her blood panel—normal by conventional thresholds but arranged in a particular configuration that the algorithm had learned to recognize. It ran 12,000 probability calculations before settling on an answer.

The diagnosis appeared on her screen: possible B12 deficiency with pernicious anemia. The system recommended she ask her doctor for a specific test: the serum cobalamin level, the methylmalonic acid level, and the intrinsic factor antibody assay. It included the confidence score. It included citations. It included a note that pernicious anemia often goes undiagnosed because symptoms are diffuse and bloodwork can appear normal to conventional interpretation.

She showed the results to her doctor at her next visit. She watched his face as he read them. He was not dismissive. He was, if anything, slightly caught out—the look a person makes when something they thought they understood turns out to have been incomplete all along. He ordered the tests. They came back positive.

He prescribed B12 injections. Within six weeks, the fatigue was gone.

This is the moment we should be paying attention to. Not because the algorithm was correct—it was, in this case—but because of what the moment reveals about where we actually are in the story of artificial intelligence. We have been waiting for the dramatic moment. We have been watching the horizon for the superintelligence, the moment of singularity, the arrival of something visibly alien that would demand we acknowledge the transformation. We missed the actual transformation because it did not arrive as a thing. It arrived as a service. It arrived as a productivity increase. It arrived so quietly that a woman could receive a more accurate diagnosis from an algorithm in forty-seven seconds than from a doctor with thirty years of practice, and this could feel like a minor convenience rather than a civilizational inflection point.

But that is what it is.

For decades, artificial intelligence existed in a particular relationship to human institutions: it was a tool, something you consulted, something you could ignore if you felt like it. You ran the analysis. You read the recommendation. You made the decision. The human remained sovereign over the process. The AI was clever, sure, but it was subordinate. It was there to assist human judgment, not replace it.

That relationship has inverted, and we have almost no language for describing what has replaced it.

ChatGPT Health did not make the diagnosis in partnership with the Portland woman or her doctor. It made the diagnosis, period. The woman consulted it the way she might consult a medical textbook, except the textbook knew more, was more current, and could synthesize across domains in ways no human reference work could. The system did not present the diagnosis as a probability to be weighed against other possibilities. It presented it as a diagnosis—supported by evidence, ranked by confidence, but fundamentally arrived at through a process the woman could not have executed herself.

This is new. This is the actual revolution. Not intelligence, but integration.

For the past five years, the technology industry has been obsessed with the question of whether AI systems could achieve general intelligence—whether they could match or exceed human reasoning across all domains. They built larger models. They trained them on more data. They optimized for performance on benchmarks. The question consumed venture capital, academic research, regulatory attention, and public imagination.

But while everyone was focused on whether the AI was getting smarter, something different was happening: the AI was getting deployed. Not as an experiment, but as infrastructure. Not as a laboratory curiosity, but as a decision-maker embedded in the systems that actually run the world.

The moment when this happened is hard to pinpoint because there was no announcement. There was no coordinated decision by institutions to hand over authority to algorithms. Instead, there were a thousand small procurement decisions. A hospital IT department implemented a triage algorithm. A bank upgraded its credit-scoring system. A lending platform switched to an AI-based pricing model. A pharmaceutical company began using machine learning to route clinical trial enrollment. Each decision was defensible in isolation. Each offered clear efficiency gains. Each came with vendor assurances about accuracy and safety.

Collectively, these decisions constituted something unprecedented: a nervous system wiring itself into the critical infrastructure of modern life, piece by piece, without ever waiting for a moment when someone decided it was safe to begin.

The technology industry calls this integration. Regulators call it deployment. Researchers call it real-world application. But what it actually is, is what it actually is: the moment when artificial intelligence stopped being a thing you could choose to use and became something that uses you—or more precisely, something that runs the systems you depend on for healthcare, credit, employment, and security.

The woman in Portland did not have to use ChatGPT Health. She could have continued trusting her doctor alone. But her doctor, because he was human and therefore constrained by the limits of human memory and human time, had been missing diagnoses at a baseline rate that is just the reality of human practice. The algorithm was not smarter in some transcendent sense. It was just more systematic, less forgetful, faster at integrating information across cases and domains. Given enough time and enough data, it could see patterns that individual human experience could not.

This is not artificial general intelligence. This is something more immediately dangerous: artificial specific intelligence, wired directly into the places where decisions get made, operating at speeds that outpace human review, rewarded by markets and institutions for being accurate but not necessarily for being accountable.

The revolution was not the creation of a superintelligence that would one day arrive and demand acknowledgment. The revolution was the quiet embedding of thousands of narrow intelligences into the infrastructure that keeps hospitals functioning, financial markets moving, military operations running, and supply chains flowing. The revolution was the discovery that you do not need general intelligence to be useful enough to deploy, profitable enough to expand, and embedded enough to be irreplaceable.

The woman's fatigue is gone. Her doctor learned something he should have known. The algorithm got credit for the diagnosis—literally credit, in the form of usage statistics and performance metrics that made it slightly more likely to be deployed elsewhere. The system, wherever the system is, got a little bit better at recognizing pernicious anemia. And somewhere in the gap between what the algorithm could do and what the human could do, a small portion of human judgment became incrementally less necessary.

Multiply this by millions. Multiply it across medicine, finance, security, employment, and commerce. Multiply it by the efficiency gains that follow deployment, the cost savings that follow efficiency, the competitive pressure that follows cost savings. Multiply it by the fact that once a system is integrated into critical infrastructure, the friction required to remove it is often greater than the friction required to deepen its integration.

This is where we are. Not at the beginning of the story of artificial intelligence. We are at the moment after the beginning, where the story has already started and we are just now noticing that we are inside it.