November 4, 1952: CBS viewers were expecting a vote-counting marathon—and instead saw, in the studio, a blinking console the size of a large table, crowned with a cabinet full of tubes. UNIVAC I, the first commercially produced computer in the United States, promised to predict the election outcome before the polls closed on the West Coast. By nine in the evening, with less than seven percent of the vote counted, the machine delivered its verdict: Dwight Eisenhower would receive 438 electoral votes to 93 for Adlai Stevenson, at odds of 100 to 1. The producers got scared and buried the forecast. When the final tally came in—442 to 89—the computer had been more accurate than any expert, but CBS had already managed to look cowardly. This is a story about how people are not ready to trust a machine, even when it is right.
UNIVAC I weighed 16,686 pounds, consumed 125 kilowatts, and lived on 6,103 vacuum tubes; the full installation occupied more than 382 square feet. Its memory held 1,000 words in mercury delay lines: pulses traveled through tubes of mercury, bounced off the ends, and returned, storing bits as sound waves. Every bit was an echo you had to catch before it faded. Remington Rand started with a price of $159,000; by 1952, the machine cost around a million, and was more often leased—for $25,000 a month.
Mathematician Max Woodbury wrote the extrapolation program. You take historical data by state—how they voted in 1944 and 1948—look at which counties deviate from the average faster than the rest, and build coefficients. Then you feed in the early results—a few dozen counties from the East Coast—and the machine scales the trend up to the entire country. No magic: a representative sample plus linear regression. But in 1952, it looked like divination with bones.
Gallup polls gave Stevenson the lead until the final weeks of the campaign, while the last survey promised Eisenhower victory by the narrowest of margins—hence the confidence that it would be a long night. The memory of 1948 still hurt: newspapers had declared Dewey the winner, and Truman had won. CBS feared repeating the humiliation. When UNIVAC produced 438–93, the people in Philadelphia decided the machine had lost its mind and asked it to “forget” some of the trend coefficients from past elections, so the forecast would be more cautious. They put numbers like 8 to 7 in Eisenhower’s favor on the air—safe, vague, but plausible.
I try to imagine sitting in a studio with a machine that costs more than a house and not believing it. Not because it got something wrong, but because it is too certain. Human intuition demands doubt, and UNIVAC knew none.
By midnight Eastern time, precinct returns had begun to converge with the forecast. Eisenhower was taking state after state—Ohio, Illinois, Texas—and it became clear that the machine had not lied. Final score: 442 electoral votes to 89. UNIVAC’s error—less than 1% in the Electoral College. In the popular vote, the machine predicted 32,915,000 votes for Eisenhower; the actual figure was 33,936,252—about a 3% error. For a model built on a few dozen counties and historical trends, it was incredible.
CBS had to admit it. Half an hour after midnight, Art Draper of Remington Rand went on the air and said: “As the new data came in, the odds came back, and it became obvious that we should have had the courage to believe the machine from the beginning. It was right. We were wrong.” There was no formal apology with a reading of charges, but the network’s discomfort was obvious. Edward R. Murrow put it more succinctly: “The trouble with machines is people.” The press seized on the story: the Washington Post sniped that “stupid people” had started fiddling with the settings and thoroughly confused the poor machine, while Time magazine mused about an “electronic brain.” By the end of the decade, Americans used “UNIVAC” as a synonym for computers in general.
One caveat: UNIVAC I was not the world’s first commercial computer. In February 1951, Britain’s Ferranti Mark 1 went to the University of Manchester—six weeks before the Census Bureau formally accepted UNIVAC, and almost a year before its actual delivery to Suitland. But Americans do not like qualifications. In the popular imagination, UNIVAC became “the first real computer”—because people saw it live on air, with blinking tubes and magnetic tape, at the most dramatic moment of the political year.
Woodbury’s model operated on assumptions that now seem naive. It assumed voters behaved predictably: if a county in 1948 deviated from the national trend by +5% toward the Democrats, it would deviate the same way now. The program did not account for demographic shifts, migration, or local scandals. It simply took several “anchor” counties—those that had historically voted close to the average—and scaled their results up to the whole country using weighted coefficients.
Why did it work? 1952 was a rare case of a uniform trend. Eisenhower was winning everywhere: in cities, in rural areas, in the North and the South. Stevenson could not hold even traditionally Democratic states. The model was simply riding the wave. Had the election been tighter—as in 2000 or 2020—linear regression on a small sample would have failed. But that time, luck and mathematics aligned.
Alongside Woodbury, the Remington Rand team kept engineers by the tubes—they burned out so often that replacements were made right in the middle of calculations. A programmer fed in the data on punched tape, the machine printed interim results on a teletype, and an analyst decided whether there were enough numbers to publish. This was not autonomous AI but a hybrid: a human chose which counties to enter first, the machine calculated the coefficients, and a human checked again.
Interestingly, CBS never showed viewers the actual calculation process. The studio contained a mock-up console whose lights blinked thanks to switches normally used to make Christmas lights flash, while the real machine hummed in Philadelphia, a hundred miles from the studio. The camera filmed the blinking lights, the journalist spoke of an “electronic brain,” and nobody explained on air that the machine was extrapolating early percentages, not reading voters’ minds.
Later, the story accumulated myths. One version claims CBS completely concealed the forecast and showed only vague percentages, publicly apologizing only several days later. Another says the producers improvised alternative figures on the spot, without the involvement of the UNIVAC team. Remington Rand documents and broadcast transcripts confirm this: the network really did ask for the forecast to be “recalculated” with weakened coefficients, but there was no formal public apology with a prepared statement read aloud. Draper acknowledged the mistake on air—that was enough to turn the story into a scandal.
Legend is good for advertising. Remington Rand used the election as proof that computers could do more than calculate payrolls. Every subsequent UNIVAC model was sold with a mention of November 4. IBM, the main competitor, responded with its own machine for processing business data—the IBM 705 model appeared in 1954, with deliveries beginning two years later. The 1952 election became the first advertisement for computers as tools for making decisions, rather than mere calculators.
But off-camera, an awkward question remained: if the machine is right and people do not trust it, who makes the decision? CBS chose reputation over accuracy, and it worked—nobody demanded that the network be taken off the air. But UNIVAC was remembered as “the machine nobody believed,” and that role proved more powerful than simply being right.
By the end of the decade, all three networks had entered the computer game. For the 1960 election, ABC called on UNIVAC again, NBC acquired an RCA 501, and CBS an IBM 7090. Forecasts were still built on historical trends and early vote percentages, and at first the machines unanimously gave victory to Nixon—until NBC gave it to Kennedy at half past eight, and the others followed. Direct measurement of voter intentions came later: CBS conducted its first exit poll in 1967, in Kentucky’s gubernatorial election, and by 1972, surveys of voters leaving polling places had become standard in elections. UNIVAC in 1952 showed that representative sampling worked, but the future belonged to sociology, not the statistics of the past.
Modern models—FiveThirtyEight, The Economist, Decision Desk—use Bayesian networks, Monte Carlo simulations, and machine learning on polling data. But the principle is the same: take a small sample, find a pattern in it, scale it up to the general population. The difference is the volume of data. UNIVAC processed dozens of counties; today’s models process millions of tweets, hundreds of polls, historical correlations by race, age, and education. But the risk remains: if the sample is unrepresentative, the model lies. 2016 was a harsh reminder of that.
Exit polls became the standard, but they get things wrong too. In 2004, early exit polls showed Kerry winning; the final result gave it to Bush. In 2020, polls overestimated Biden in swing states. Every time, analysts say: “We improved the methodology.” Every time, some voters stay silent or lie to pollsters. UNIVAC in 1952 did not know that problem—it worked with facts: who had already voted and how. But as soon as models began relying on intentions, they lost some of their accuracy.
The industry has grown enormous. Television networks spend millions on real-time graphics, state maps showing percentages counted, expert panels explaining why the machine predicted one outcome and another came to pass. But trust in forecasts remains fragile. Viewers remember misses more vividly than hits. And that, too, is the legacy of that night in 1952: the machine may be right, but if you do not trust it at once, you will not trust it later either.