The Night the Power Went Out: Six Hours Without AI
February 17, 2030
The failure cascaded at 10:47 PM on a Tuesday. A transformer fault at a substation in Southwark propagated through the grid in the specific manner that grid engineers call "sympathetic tripping" — a failure that recruits adjacent systems into its collapse, each tripping breaker triggering the next, the way a line of dominoes falls not because any single domino is weak but because they are connected.
St. Thomas' Hospital lost main power at 10:48 PM. The emergency generators activated at 10:48:03 — a three-second gap during which the hospital existed in the dark, sustained by battery backups and the residual charge in a thousand pieces of equipment. The generators were rated for essential systems: life support, surgical theatre lighting, emergency lighting in corridors and stairwells, the pharmacy refrigeration units.
The generators were not rated for the AI systems. The diagnostic AI, the scheduling AI, the bed-management AI, the drug-interaction AI, the triage-assistance AI, the imaging AI, the lab-results AI, and the seventeen other AI systems that had been integrated into the hospital's operations between 2026 and 2029 went dark at 10:48 PM and did not return until 4:52 AM.
Six hours and four minutes. In those six hours, St. Thomas' Hospital ran on humans.
The first hour
The immediate response was not panic. It was confusion — the specific confusion of people who reach for a tool that is not there. The A&E registrar, Dr. James Osei, reached for the triage-assistance system — a reflex, the way you reach for a light switch in a dark room — and found nothing. The screen was blank. The recommendations were absent. The patient in front of him — a sixty-year-old woman with chest pain, sweating, clutching her left arm — required a decision that the AI system would normally have informed within seconds: ECG interpretation, troponin risk stratification, disposition recommendation.
James had the training. He had graduated medical school in 2024, before the AI systems were fully integrated. He could read an ECG. He could calculate a HEART score manually. He could make the disposition decision — admit, observe, or discharge — using the clinical judgment that six years of medical training and four years of practice had developed.
He made the decision. He admitted the patient. The decision took him ninety seconds — ninety seconds of looking at the ECG, calculating the score in his head, assessing the patient's presentation, and arriving at a judgment. The AI system would have produced the same recommendation in four seconds.
The ninety seconds felt like a year. Not because James was slow. Because the ninety seconds contained something the four seconds did not: weight. The weight of responsibility. The weight of a decision that rested entirely on his judgment, his training, his assessment, with no algorithmic confirmation, no probability estimate, no safety net of machine intelligence cross-checking his reasoning.
He made the right call. The patient was having an NSTEMI. She was admitted to cardiology, treated, discharged five days later. The outcome would have been identical with the AI system active.
The experience was not identical. The experience was heavier.
The ward
On the general medical ward, Sister Patricia Mbeki — forty-three, twenty years of nursing, the kind of nurse who other nurses describe as "the real one," meaning the one you want at your bedside when the system fails — discovered that she knew more than she thought she knew.
The AI scheduling system had managed the ward's patient flow, medication timing, staff allocation, and discharge planning. Without it, Patricia stood at the nurses' station with a whiteboard and a marker and she rebuilt the ward's operations from her own knowledge — which patients needed observations at which intervals, which medications were due when, which beds could accept new admissions, which patients were approaching discharge.
She rebuilt it in twenty minutes. The rebuild was not as efficient as the AI system's allocation — she missed an optimal medication timing by fifteen minutes and she allocated one nurse to a bay that needed two. She corrected both errors within the hour.
The errors were interesting. They were the kind of errors that Priya Chakrabarti would later catalogue — errors that revealed the architecture of human attention. Patricia missed the medication timing because she was concentrating on a complex admission. She under-allocated the nurses because she was mentally modeling the entire ward and the model was slightly wrong in the corner she was paying least attention to.
The AI system did not make these errors because the AI system did not have attention. It did not concentrate. It did not model the ward in its mind. It processed all variables simultaneously, without priority, without fatigue, without the specific human limitation of being able to attend to only one thing at a time.
Patricia's errors were the errors of a mind. The AI's error-free performance was the performance of a system. The difference was not in the outcome — both produced adequate ward management. The difference was in the texture. Patricia's ward management had the texture of a person: attentive, fallible, adaptive, alive. The AI's ward management had the texture of a process: optimal, consistent, precise, impersonal.
The patients noticed. Several patients later told the review committee that the night the power went out was the first night they felt the ward was staffed by people who were making decisions rather than following instructions. "The nurse came to check on me," said one patient, "and she actually looked at me. Not at the screen. At me. I could tell she was thinking — really thinking — about whether I was all right. It was different. It was heavier."
The morning
At 4:52 AM, the main power was restored. The AI systems rebooted. The triage-assistance system resumed recommendations. The scheduling system recalculated the ward allocations. The drug-interaction system began cross-checking the prescriptions that had been written in the six-hour gap.
The resumption was seamless. The systems picked up where they had left off. Within minutes, the hospital was operating in its normal mode — AI-informed, AI-guided, the human staff receiving recommendations and confirmations and the ambient support of a dozen intelligent systems working in parallel.
James Osei was standing in A&E when the systems came back online. He described the experience to the review committee:
"The triage system booted and the first thing it did was retrospectively analyze the decisions I had made during the outage. It confirmed every decision. Every triage call, every disposition, every escalation — the system reviewed them and rated them as appropriate. I scored 100 percent."
"And I felt — this is hard to explain — I felt relieved and diminished at the same time. Relieved because the decisions were correct. Diminished because the confirmation came from a system, not from my own confidence in the decisions. For six hours, I had trusted my judgment. The trust was heavy. It was mine. When the system came back and confirmed the judgments, the trust transferred — it moved from my body to the screen. The weight lifted. And I missed it."
The report
The hospital's post-incident review was conducted over three weeks. The review found no patient harm, no clinical errors requiring remediation, no systemic failures beyond the power loss itself. The human staff had managed the hospital for six hours and four minutes without AI support and the outcomes were — the review's word — "adequate."
The word was debated. "Adequate" was technically correct. No harm occurred. But the staff objected to the word because it failed to capture what the six hours had felt like — the weight, the aliveness, the specific quality of running a hospital on human judgment alone.
Patricia Mbeki submitted a written addendum to the review:
"The word 'adequate' describes the outcomes. It does not describe the experience. For six hours, every decision in this hospital was made by a person. Not informed by a person, not reviewed by a person, not approved by a person — made by a person. The decisions were adequate. The experience was extraordinary."
"I have worked in this hospital for fourteen years. For the last four years, I have worked alongside AI systems that assist, recommend, and optimize. The systems are excellent. I am a better nurse with them than without them. This is not in question."
"What is in question — what the review should acknowledge — is what the six hours revealed: that we can still do this. That the training is still in our bodies. That the judgment is still in our minds. That the competence has not atrophied, not yet, not entirely. For six hours, we remembered what it felt like to be the system — not a component of the system but the system itself. We remembered the weight."
"The weight is what the AI removes. The weight of responsibility, of judgment, of the knowledge that the decision is yours and yours alone. The AI carries this weight for us, and we are grateful, and we should be grateful. But the weight is also the thing that makes us professionals. The weight is the practice. And for six hours, we practiced."
This is a supplemental entry in The Fracture Line. For the night that began the era this power outage briefly reversed, see The Last Diagnosis. For the protocols that were supposed to govern these systems, see The Orphan Protocols.