The Archive of Errors: What Human Mistakes Reveal About Human Minds
September 2035
The collection began with a misdiagnosis.
Priya Chakrabarti was thirty-two, a data scientist at Imperial College London, and she had been hired to do something that felt, to her, like archaeology performed on the recently dead: she was cleaning the diagnostic records of the NHS's radiology department from 2022 to 2028 — the final years of human-primary diagnosis. The NHS had contracted a data hygiene project to sanitize the records for long-term archival, and Priya's job was to flag errors, categorize discrepancies, and prepare the dataset for a system that would use it to train future AI models on what correct diagnoses looked like.
The work was mechanical. She had processed eleven thousand records in her first month. Each record contained the human radiologist's reading alongside the AI's retrospective reading — the ground truth, as the dataset called it. Where the two diverged, Priya flagged the human reading as an error and categorized it: missed finding, false positive, incorrect characterization, wrong laterality, measurement discrepancy.
Eleven thousand records. Four hundred and twelve errors. A 3.7 percent error rate — respectable by the standards of the era, damning by the standards of the machine that had replaced the radiologists. The AI's retrospective error rate on the same dataset was 0.03 percent.
Priya flagged the errors. She categorized them. She prepared to move on. And then she made the mistake that changed her career: she read one.
The error
Record 7,419. A chest CT from February 2026. The human radiologist — identified only by an employee number, though Priya would later learn her name was Dr. Sarah Oluwole, fifty-one, twenty-three years of experience, Adaeze Nwosu's colleague at the same London hospital — had reported a 3mm ground-glass opacity in the right lower lobe as "likely benign, recommend follow-up in 12 months."
The AI's retrospective reading: "3.2mm ground-glass opacity with 2.7% annual malignancy risk. Recommend follow-up in 6 months per Fleischner Society guidelines."
The human had underestimated the size by 0.2 millimeters. The human had recommended 12-month follow-up instead of 6-month. The human had called it "likely benign" when the AI's probabilistic assessment placed it in a grey zone — not likely benign, not likely malignant, but ambiguous enough to warrant closer surveillance.
By the standards of the dataset, this was an error. Priya flagged it. Category: measurement discrepancy plus interval recommendation discrepancy.
Then she stopped. She looked at the record again. She looked at the timestamp: 02:47 AM. She looked at the case sequence: record 7,419 was the forty-third scan Dr. Oluwole had read that night. She looked at the preceding scan: record 7,418 was a trauma CT of a seventeen-year-old with a ruptured spleen — an emergency case, full of adrenaline and urgency, the kind of case that floods the body with cortisol and narrows attention to the acute findings.
And then Priya understood something that the error-flagging system had not been designed to detect: Dr. Oluwole's "error" was not random. It was a readable trace of a human mind in motion. The 0.2mm underestimate was the residue of attention — a mind still partially activated by the emergency case, its perceptual threshold slightly elevated by the adrenaline, causing it to minimize a non-urgent finding. The 12-month recommendation instead of 6-month was the same pattern: a mind calibrating risk against the background of what it had just seen, a mind for whom "urgency" had been temporarily redefined by a ruptured spleen.
The error was not noise. The error was a window into the architecture of Dr. Oluwole's attention at 2:47 AM after forty-three scans and one emergency.
Priya sat at her desk and she thought: What if all the errors are like this?
The collection
She went back to the beginning. All four hundred and twelve errors from her first eleven thousand records. She stopped categorizing them by type — missed finding, false positive — and started reading them as documents. As testimonials. As the involuntary autobiography of human minds under the specific pressure of professional performance.
What she found:
Errors clustered by time of day. Between 2 AM and 5 AM, the error rate tripled — not because the radiologists were incompetent but because the human circadian system degrades visual attention in predictable ways. The specific errors changed too: late-night errors were errors of omission (missing things), while early-morning errors were errors of commission (seeing things that weren't there). The human mind at 3 AM was a different instrument than the human mind at 10 AM, and the errors were the instrument's signature, the way a violin's overtones tell you about the wood.
Errors clustered by sequence. After an emotionally demanding case — a pediatric scan, a trauma case, a patient the radiologist knew personally — the next three to five readings showed a measurable shift. The shift was not uniform. Some radiologists became more cautious (overcalling findings), while others became less cautious (undercalling). The direction of the shift was consistent within individuals — each radiologist had a characteristic response to emotional load, a signature that appeared in their errors the way a handwriting style appears in letter formation.
Errors clustered by expertise. Junior radiologists made textbook errors — they missed what they hadn't learned to look for. Senior radiologists made creative errors — they saw what their experience told them should be there, even when it wasn't. The senior errors were more interesting: they were the traces of pattern recognition overshooting, of a mind so fluent in the visual language of pathology that it occasionally wrote the next word of the sentence before reading it. These were not failures of competence. They were the excess of competence — the cost of the same pattern recognition that allowed senior radiologists to catch what junior radiologists missed.
Priya assembled the errors into a database. Not a database of mistakes. A database of cognitive signatures. Each error was a data point that located the radiologist — their fatigue level, their emotional state, their expertise level, their characteristic response to stress — in a multidimensional space that the correct readings, in their uniform accuracy, could never reveal.
The correct readings told you what the scan showed. The errors told you what the mind was doing.
The paper
She published in 2036. "Error as Signal: What Human Professional Mistakes Reveal About the Architecture of Expert Cognition." The paper was forty-seven pages. The abstract was one sentence: Human errors are the fossil record of human thought.
The paper argued that the AI systems that had replaced human professionals were, by design, error-free within their confidence bounds — and that this error-freedom, while clinically superior, was also cognitively opaque. An AI that never made errors could not be read. Its internal states did not leak into its outputs. It was a perfect surface — smooth, correct, impenetrable.
Human professionals, by contrast, leaked constantly. Their errors were the leaks. Every mistake was a window into a process — attention, memory, emotion, circadian rhythm, expertise, fatigue, the entire embodied context of a mind at work. The errors were the only places where the process was visible from the outside.
And the process, Priya argued, was the thing that mattered — not for clinical accuracy, where the AI had won conclusively, but for understanding what human cognition actually was. If you wanted to map the territory of the human mind — its architecture, its rhythms, its characteristic modes of failure and recovery — you did not study the correct answers. You studied the errors. The errors were the coastline. The correct answers were the open sea — vast, navigable, featureless.
The foundation
The paper was read by twelve people in its first year. One of them was a young researcher named Solène Diarra, who was twenty-six years old in 2036, finishing a master's degree in cognitive science at the Sorbonne, and beginning to think about the boundary between human and machine cognition not as a problem to solve but as a territory to explore.
Solène read Priya's paper in a café in the 5th arrondissement, on a Tuesday afternoon in October, and she underlined a single passage:
The errors AI never makes are as informative as the errors humans do make. Every class of error that disappears when AI assumes a task defines a dimension of human cognition that the task required. Fatigue errors define the dimension of embodiment. Emotional-carryover errors define the dimension of affect. Expertise-overshoot errors define the dimension of pattern recognition. When we catalogue the errors that AI eliminates, we are not cataloguing human weakness. We are cataloguing human presence — the specific, irreducible presence of a body and a mind and a history, working together, in time, under pressure, in the dark reading room at 3 AM.
Solène wrote in the margin: This is the map.
She did not know yet what she meant. She would not know for eight more years — not until 2044, when she stood at the boundary for the first time and recognized in the territory's deep structure the same thing Priya had found in the errors: the signature of human cognition, not in its accuracy but in its characteristic ways of being wrong.
Priya's archive of errors became one of the foundational datasets for the Cartography Institute, established in 2047. The errors were the first map — not of the territory itself, but of the mind that would explore it. Before the cartographers could map the boundary between human and machine thought, they needed to know what human thought looked like from the outside. Priya had shown them: it looked like its mistakes.
Priya, 2037
Two years after she began the collection, Priya visited Dr. Sarah Oluwole — the radiologist from record 7,419, the error that had started everything. Sarah was fifty-seven now, retired from diagnostic work, teaching medical history at King's College.
They met in Sarah's office, which was lined with textbooks she no longer assigned. Sarah had agreed to the meeting because Priya's paper had described her error — anonymized but recognizable to Sarah, who remembered the night, who remembered the ruptured spleen, who remembered the forty-third scan at 2:47 AM.
"You described my error as a 'cognitive signature,'" Sarah said. "A trace of my mental state at that moment."
"That's what the data shows."
"The data shows a wrong answer. I underestimated a measurement and recommended the wrong follow-up interval. In clinical terms, that's a mistake."
"In clinical terms, yes."
"In your terms?"
Priya thought about this. She had rehearsed many versions of this answer, and none of them felt adequate, because the answer required a kind of double vision — the ability to see the error as both a mistake and a message, both a failure and a finding.
"In my terms," Priya said, "your error is the most informative reading in the entire dataset. The forty-two correct readings before it tell me what the scans showed. Your error tells me who you were at 2:47 AM. The emergency case had flooded your system with adrenaline. Your attention threshold was elevated. Your risk calibration was shifted toward the acute. You were still, neurologically, in the emergency — your body hadn't finished processing the seventeen-year-old's ruptured spleen — and that unfinished processing bled into your reading of the next scan. The error is the bleed. The error is the evidence that you are a body as well as a mind, that your diagnostic skill was not a disembodied algorithm but a process rooted in a nervous system that was tired and activated and human."
Sarah was quiet for a while.
"The AI would not have made that error," she said.
"The AI does not have a body."
"The AI does not have a 2:47 AM."
"No. The AI does not have a 2:47 AM."
Sarah looked at her hands — the hands that had adjusted images on screens for twenty-three years, the hands that still made the scrolling gesture when she described a finding to her students, the hands that carried the muscle memory of four hundred thousand scans.
"Then the errors are the proof," she said. "The proof that a human was here. The proof that the readings were made by someone who could be tired, who could be affected, who could carry one patient's emergency into the next patient's scan. The errors are the fingerprints."
"The errors are the fingerprints," Priya said.
Sarah nodded. She opened a drawer and took out a stack of printed scans — teaching materials, annotated in her handwriting. She held them for a moment. The hands that held them were the same hands that had read the forty-third scan at 2:47 AM, the hands that had made the error that was not a failure but a fossil, a trace, a record of what it was like to be human and reading and tired and haunted by a ruptured spleen in a dark room in the middle of the night.
"Keep the errors," she said. "They're the best thing we ever made."
This is the second entry in The Inheritance. For the night that preceded the era of errors this archive documents, see The Last Diagnosis. For the atlas that eventually included Priya's work, see The Atlas of Disappearances.