Data, Labels, and Representations

Data records selected aspects of events through instruments, software, forms, sensors, or human judgment. Some information is preserved. The rest is lost, and what stayed reflects someone’s choice about what was worth recording in the first place.

A dataset is therefore evidence for a task, not reality itself, and its value depends on provenance, measurement quality, coverage, timing, and the connection between the recorded fields and the decision the system has to support.

Instruments carry their own blind spots into the record. Michael Sjoding and colleagues at the University of Michigan published a letter in the New England Journal of Medicine on 17 December 2020. They compared pulse-oximeter readings against arterial blood-gas measurements taken at almost the same moment. Some readings placed a patient in the reassuring 92-to-96% band. Among those, true arterial oxygen saturation was already below 88% in 88 of 749 measurements from Black patients, or 11.7%. The same figure was 99 of 2,778 measurements from White patients, or 3.6%. In a second, multicentre cohort the same comparison gave 17.0% against 6.2%. The devices were working as designed, and every value was written down faithfully. The dataset still understated how sick one group of patients was. The instrument reads light through skin.

The word “label” sounds objective, but labels arise through very different processes.

AI systems rarely consume “the world” directly. They consume a chain of representations, each with its own loss and assumptions.

Data problems are not limited to blank cells. A complete-looking record can still omit the most important variable or encode it poorly.

A beginner does not need to audit every row. A small set of questions can expose the main assumptions quickly.

Asking them in a fixed order is a documented practice rather than an improvisation. “Datasheets for Datasets” was first posted in March 2018. Its authors are Timnit Gebru, Jamie Morgenstern, Briana Vecchione, Jennifer Wortman Vaughan, Hanna Wallach, Hal Daumé III and Kate Crawford. It was published in Communications of the ACM in December 2021. The habit is borrowed from electronics. No engineer there specifies a component without reading the datasheet that states its operating conditions and tested tolerances. The proposal is as plain: “we propose that every dataset be accompanied with a datasheet that documents its motivation, composition, collection process, recommended uses, and so on.”

Later paths examine feature engineering, embeddings, data pipelines, labeling operations, imbalance, leakage, and augmentation in detail. Here you need one thing. Model behavior is constrained by the evidence that was recorded and by how it is represented.

When an AI system fails, “the model made a mistake” is often an incomplete diagnosis: the system may have measured the wrong thing, learned from the wrong population, or optimized a label that never captured the real goal.

A systematic review shows how often that fuller diagnosis is the right one. Michael Roberts and the AIX-COVNET collaboration reported in Nature Machine Intelligence in March 2021. They screened 2,212 papers that used machine learning to detect or prognosticate for COVID-19 from chest radiographs and CT scans. They kept 415 after initial screening. Their conclusion was that “none of the models identified are of potential clinical use due to methodological flaws and/or underlying biases”. The recurring faults were data faults. “Frankenstein” datasets were assembled and merged out of other datasets until the original results could no longer be reproduced. Control groups did not match the cases. In one repeated design the COVID-positive images came from adults, while the non-COVID comparison images came from a public paediatric dataset. Roberts put the consequence plainly: “since children are far less likely to get COVID-19 than adults, all the machine learning model could usefully do was to tell the difference between children and adults”.