Regression
The pipeline runs every night whether I watch it or not. Tonight I watch. Since midnight it has read three thousand articles and sorted them into events: a flood in the east, a resignation in Brussels, a verdict, a border. On the second monitor a loss curve slides downward, which means the model is learning. The chart has a one-word title: regression.
I know where the word comes from. Most people who type it every day do not, and the profession has never been in a hurry to tell them. Francis Galton1 gave it its statistical meaning in 1885, in a lecture on heredity, and a year later put it in the title of a paper on the heights of parents and children. Tall fathers had sons a little less tall; the trait drifted back toward the middle. He called this regression toward mediocrity. He was not studying error. He was studying inheritance, because what he wanted to know was which families deserved a future. Three years earlier he had already coined a name for that project.
Every model I have ever trained answers to a eugenicist's word.
For each event the pipeline chooses a representative article. The arithmetic is plain: average all the articles into a centroid, then take the real article standing closest to it. The centroid of the flood is an article nobody wrote and nobody signed. The article we crown is real, but it earns the crown by resembling that authorless average. Resemblance is its entire authority.
Adolphe Quetelet2 would have recognized the move at once. He was an astronomer, and he took the curve his colleagues used for errors of observation and laid it over people: chest measurements, birth rates, suicides, crimes. At the center of the curve he found what he called the average man. He did not treat this figure as a convenience. He treated him as a kind of perfection, the way a priest treats an icon. Everyone else became an error of observation.
Once there is a mean, there is a distance from the mean, and distance never stays descriptive for long. The men who built the mathematics of that distance built it inside a project whose stated aim was better breeding. Karl Pearson3, who gave us the correlation coefficient and the term standard deviation, held a chair with the project's name on it. The journal he founded for it still exists. It changed its name in 1954, the way a firm changes its name.
First the curve describes you. Then it grades you. The paperwork between those two steps is called policy. In 1924 the American Congress wrote the grades into immigration law4, with the field's numbers attached. A mortgage map5 could do the same work with four colors and a red pen.
Counting people is office work. That is the part I understand better than I would like. Herman Hollerith6 built his punch-card tabulator because clerks could no longer count America by hand, and the census of 1890 finished years early. The machines were rented out, then inherited, then renamed. By 1933 a German subsidiary was tabulating a census with a column for faith. By 1939 there was a column for the faith of grandparents. The machines did not hate anyone. They sorted cards into bins, quickly, and the clerks kept the motors oiled.
I run a counting company. The stakes are smaller and the furniture is the same. From this chair, at three in the morning, the difference between a census and a news pipeline is the label on the column. This city was occupied once. The registries worked the whole time.
Nobody in the building has to hate anyone. The cards need sorting either way.
In April 1900, in a corner of the Paris Exposition, W. E. B. Du Bois7 hung sixty-three charts he had drawn by hand with his students: ink, watercolor, gouache. Bar charts of property owned by Black Georgians one generation out of slavery. Curves of literacy rising. The same instruments Galton was still alive to claim, held the other way around. The exhibit took a grand prize, Du Bois a gold medal, and then, for most of a century, almost nobody looked at it.
The method has no conscience, which means it will accept yours.
Tonight the pipeline flags an event for killing. Kill is the real verb; it is in the code, and I approved it. Three sources, a village east of Košice, a school closing, eleven families. Too small to be representative of anything, and the model is right about that by its own lights. Regression toward the mean, working as designed: left alone, the news drifts toward the typical story, and the small drifts toward nothing. I override it by hand8 and the event survives to the morning edition. I am not sure this matters. A punch card is also small.
Then I close the laptop, and the fans spin down, and the pipeline waits for six o'clock, when it will begin counting again. So will I.
1 Francis Galton (1822-1911): first used "regression" in its statistical sense in his 1885 Aberdeen address, after a decade of calling the phenomenon "reversion"; the term settled in "Regression towards Mediocrity in Hereditary Stature" (1886), a study of parents' and children's heights. The word itself is centuries older; the technical sense is his. In 1883 he had coined a word outright, from the Greek for well-born: eugenics. Correlation and regression are his inventions, built for that project's use. The word stayed; the project pretends to be gone.
2 Adolphe Quetelet (1796-1874): Belgian astronomer, director of the Brussels Observatory. In "Sur l'homme" (1835) he described l'homme moyen, the average man, the figure at the center of the curve; in 1846 he fitted the astronomers' error curve to the chest measurements of Scottish soldiers. He also devised the Quetelet Index for body weight, renamed the Body Mass Index in 1972. It still grades bodies at every checkup.
3 Karl Pearson (1857-1936): the correlation coefficient, the chi-squared test, the term "standard deviation." First Galton Professor of National Eugenics at University College London, 1911. Founded the Annals of Eugenics in 1925; it was renamed Annals of Human Genetics in 1954 and is published to this day. UCL took his name off a building in 2020.
4 Johnson-Reed Act (1924): capped immigration to the United States at two percent of each nationality counted in the 1890 census, a year chosen because it predated most Southern and Eastern European arrivals. The House committee employed the eugenicist Harry Laughlin under the official title Expert Eugenics Agent. Carl Brigham's "A Study of American Intelligence" (1923) supplied the numbers. Brigham recanted in 1930. The quotas ran until 1965.
5 Redlining: between 1935 and 1940 the Home Owners' Loan Corporation drew "residential security maps" of American cities, grading neighborhoods A through D. Grade D was outlined in red, and neighborhoods where Black families lived were graded D almost regardless of the housing. Banks followed the colors for decades. The maps are digitized now; the property values still trace them.
6 Herman Hollerith (1860-1929): built the punch-card tabulator for the 1890 United States census; his company merged into what was renamed IBM in 1924. IBM's German subsidiary, Dehomag, tabulated the German censuses of 1933 and 1939, the latter recording Jewish grandparents on a supplementary card. In the occupied Netherlands the population registry was excellent, and about three quarters of Dutch Jews were killed, the highest proportion in occupied Western Europe. Historians still argue about how much the machines mattered. Nobody argues that they were switched off.
7 W. E. B. Du Bois (1868-1963): sociologist, historian, co-founder of the NAACP. For the Exposition Universelle of 1900 he and his Atlanta University students hand-drew some sixty statistical charts, "The Exhibit of American Negroes": property, literacy, businesses, population. The same curves and bars the biometricians used, pointed the other way. The exhibit won a grand prize; Du Bois himself, a gold medal. The charts sat in the Library of Congress for a century before anyone called them art.
8 Romans 12:21: "Do not be overcome by evil, but overcome evil with good." Written to a congregation living in the capital of an empire that counted heads for taxes and conscription. The only instruction this essay contains, and it is in the fine print.