← The canon · AItopiaOrAImageddon?

UNIVAC on election night, CBS, 4 November 1952

moment · Remington Rand's Eckert–Mauchly Division in Philadelphia, which built the machine and ran the forecast — Max Woodbury of the University of Pennsylvania and Herbert F. Mitchell Jr. of the UNIVAC Applications Department on the method, Herman Lukoff keeping the hardware alive, Arthur F. Draper and Philip Vincent as the executives who decided what left the building; and CBS Television, which put it on the air — Walter Cronkite anchoring his first national election night, Charles Collingwood at a mock-up console in New York talking to a machine ninety miles away · 1952

Something that happened and changed what people expected next.

Descends from The Hollerith electric tabulating system, and the 1890 census.

Filed under idea in canon/proposals.md, and idea is wrong. The kind is moment. This is not a close call and the correction is the least interesting thing this file does, so it is dealt with first and quickly, and then the length goes where it belongs: into the claim the proposal attached to the entry, which is wrong in four separate places, and into what the record actually supports instead.

build.py's gloss for idea is "A framing later work is built out of rather than argued about. Cited when the thing being watched descends from it and the descent explains its shape." Nothing here fits. There is no mechanism in this entry that later work builds on. The statistical method — extrapolating a national result from early returns by comparing them against the same precincts in prior elections — was not new in 1952 and Chinoy's central historical finding is precisely that it "did not represent a sharp break with pre-computer approaches." The newspapers had been doing it by hand for sixty years. What election night 1952 contributed was not a technique but an evening, and the gloss for moment — "Something that happened and changed what people expected next" — is the one that lands.

Why not prediction, which is the rival that nearly wins. It has an unusually good case. There is a dated claim in its own words, printed on paper; there is a date it was made and a due date a few hours later; there is a self-grade delivered by the predictor on live television at half past midnight ("It was right, we were wrong"); and there is an independent grade, assembled from the broadcast tapes fifty-eight years later, that differs sharply from the self-grade in the direction that flattered the predictor. That is the exact shape the prediction kind was built to hold, down to the requirement to record both grades and say they differ.

It loses on subject. UNIVAC's forecast was a claim about an election, not a claim about computing, and the prediction kind earns its place in this canon by accumulating a base rate for how wrong forecasting about this field runs. A correct forecast about Ohio does not deposit into that base rate. But the discipline of the kind is imported wholesale into section 3 below, because it is the honest way to handle an entry whose central artifact is a printout with a number on it, and because the self-grade-versus-independent-grade structure is, as it turns out, the most useful thing in the entry.

Why not interpretation. The strongest rival on substance, and it is worth being explicit about why it loses, because the argument recurs. What actually travels from 4 November 1952 into 2026 is not the broadcast — almost nobody has seen the broadcast — but a parable: the machine was right, the humans were scared, the humans suppressed it, the machine was vindicated. That parable is a reading of events, and an interpretation entry on it would be a legitimate and useful thing to hold. But deep-blue-1997 settled this: the id names an event, events are moments, and the genre that grew out of the event deserves its own id if it deserves one at all. The same ruling applies here. If the owner wants the parable as its own entry — the vendor-authored postmortem that becomes the historical record — it should be proposed separately, and this file does not invent an id for it.

The proposal's claim, which is wrong four times

The proposal line reads:

> univac-election-1952 — UNIVAC I (Sperry Rand) predicted an Eisenhower > landslide from one percent of returns on election night 1952; CBS did not > believe it and did not broadcast the real forecast until the result was in. > Cite when: automation bias is discussed — this is the mirror image, a correct > machine prediction suppressed because the humans did not trust it, and the > honest counterweight to every story about people deferring to a model too > readily.

Taken in order:

"Sperry Rand" is an anachronism. In November 1952 the company was Remington Rand, which had bought the Eckert–Mauchly Computer Corporation in 1950. Sperry Rand did not exist until Sperry Corporation and Remington Rand merged on 30 June 1955, two years and eight months later. The error is small but it matters for this canon specifically, because hollerith-census-1890 already traces the lineage that produces the name — James Legrand Powers out of the Census Office, the Powers Accounting Machine Company of 1911, the 1927 merger into Remington Rand, and only then Sperry Rand — and an entry that garbles the company at the moment it appears breaks the chain the canon has already recorded.

"One percent of returns" is wrong by roughly a factor of five, and the figure it is derived from is itself contested. The printout that survives in several archives is headed:

> IT'S AWFULLY EARLY, BUT I'LL GO OUT ON A LIMB. > UNIVAC PREDICTS -- with 3,398,745 votes in --

Against a final national turnout of 61,751,942, that is 5.5 per cent, not one. And Chinoy's reconstruction complicates even that: an International News Service dispatch printed in the Hartford Courant the next morning attached the identical figure — 3,398,745 — to the aired 314-electoral-vote forecast rather than to the withheld landslide one. The number is real; which forecast it belongs to is not settled.

"CBS did not believe it" is, on the available record, backwards. Every first-hand account from the people who were in the room places the decision in Philadelphia, at Remington Rand, among the machine's own operators. Woodbury and Mitchell's post-election report: "there was a hurried consultation among the top Remington Rand people present and Dr. Woodbury. Could this prediction be correct? In view of the obvious state of fatigue of all of us who had programmed the job and the lack of time to test the entire procedure on prior election data, it was decided that it would be too risky to release this prediction." Lukoff's memoir has the officials putting their heads together and saying "We can't let this go out. The risk is too great." Chinoy states the position plainly: "The various accounts by those working on the UNIVAC predictions in Philadelphia contain no suggestion that CBS was informed of that first prediction," and Collingwood's and Cronkite's on-air behaviour — Collingwood visibly frustrated at getting nothing from the machine until 10:30 — is consistent with not knowing it existed. Collingwood's own radio script the next day put the blame on the "miserable statisticians who were handling him." Chinoy is careful to say no document has surfaced that settles it either way. But the proposal asserts as fact the one version the surviving evidence does not support.

"Did not broadcast the real forecast until the result was in" is wrong on the clock. A UNIVAC forecast of an Eisenhower landslide at 100-to-1 odds was read on the air by Cronkite shortly after 11:45 p.m. EST, and elaborated by Collingwood a few minutes later at 424 electoral votes to 107. Eisenhower did not address the country until well after 1:30 a.m.; several states were still being called past 2 a.m.; the broadcast ran until about 3 a.m. and the landslide forecast aired with more than three hours of it left to run. What CBS never aired was the earlier forecast. That is a real and interesting fact, and it is not the fact the proposal states.

What survives of the proposal — and it survives in better shape than the four errors suggest — is the instinct that this belongs in the canon. It does. Just not for the reason given. Section 2 makes the case that holds.

The descent, and the ancestors that are missing

descends_from: [hollerith-census-1890], and the link is load-bearing rather than decorative. The machine on election night was built by a division of Remington Rand, and Remington Rand is one of the two corporate lineages the Hollerith entry already traces out of the 1890 census — Powers' mechanical tabulator, the Powers Accounting Machine Company, the 1927 merger. More to the point, the activity is the same activity: aggregating a national count of the American population by machine, on a deadline, because doing it by hand takes too long. The 1890 census machine counted the population; the 1952 machine counted the population's votes. The Hollerith entry's argument that the capability to enumerate a population is a capability with consequences is the general case of which this is a very mild instance.

Two genuine ancestors the canon does not hold, named here rather than invented as ids:

A sibling, not an ancestor: rosenblatt-nyt-1958. Six years later, the New York Times turned a perceptron demonstration into a durable public overclaim that outlived every correction. This entry is the earlier and cleaner instance of the same mechanism, with one difference that makes it more useful: in 1958 the press did the amplifying, and in 1952 the vendor wrote the story itself, on the air, in the small hours, and the press merely repeated it. A reading that wants to argue about how claims about machines get fixed in public memory should hold both.

What it is

The night

Tuesday 4 November 1952. Dwight Eisenhower against Adlai Stevenson, in the first presidential election in which television was the medium most people watched it on. The final polls said close. Gallup's last national figure had Eisenhower ahead but within a few points, and the entire profession was four years past the humiliation of 1948 and determined not to repeat it.

CBS opened its television coverage at 8 p.m. EST from Grand Central Terminal in New York, Walter Cronkite anchoring, with Edward R. Murrow and Eric Sevareid analysing and Robert Trout on the radio side. NBC started radio at 8 p.m. and television at 9, out of Studio 8-H, with Merrill "Red" Mueller and John Cameron Swayze. Both networks had a computer. This is the first fact the standard story loses: there were two machines on the air that night, and the smaller one went first.

The two machines

CBS had UNIVAC I, serial 5 — a Remington Rand machine that had been built for the Atomic Energy Commission and had not yet been delivered. It stayed at the Eckert–Mauchly plant in Philadelphia, roughly ninety miles from the studio, because moving it was out of the question. Vote counts reached it by teletype from CBS's election headquarters in New York, which was itself taking them from the wire services. The forecasts came back the same way, onto a video monitor mounted on a studio wall. What the audience saw in New York was Charles Collingwood standing at a mock-up console — a prop — and periodically addressing the real machine as though it were in the room. Collingwood introduced it as "a fabulous electronic machine, which we have borrowed to help us predict this election from the basis of early returns as they come in," pointed at the tape drives and said "See those round things over there, looks kind of like candy mints? Well, those are reels of metallic tape," and — in a line the standard telling almost never quotes, though it is the most honest sentence anyone said about the exercise all night — told the audience: "This is not a joke or a trick. It's an experiment. We don't know. We think it'll work. We hope it will work."

NBC had the Monrobot, from the Monroe Calculating Machine Company: small enough to sit in Studio 8-H against a wall, and described on air by George Hicks in terms that give a fair sense of what a 1952 audience was being asked to imagine — "This is a huge machine. It will add, subtract, give odds, multiply. All you do is feed in the new figure and out on a square card comes the printed result you want." When NBC opened the top so viewers could see the lights blink, solder fell onto the drive.

What the audience actually saw, in order

This is the part the legend replaces with a single sentence, and it is worth laying out with the clock attached. Times are EST, from Chinoy's reconstruction against the CBS kinescopes held at the Paley Center for Media.

What happened in Philadelphia

Three things, in sequence, and none of them is the thing the legend describes.

First, disbelief — inside the vendor. The landslide forecast printed out and collided with everyone's prior. Draper afterwards listed the reasons: the race was expected to be close; the forecast rested on a small fraction of votes cast; it handed Eisenhower southern states with barely any returns in; and whole states had reported nothing at all. Under those conditions, Draper wrote, "it was beyond comprehension that the machine could predict with such a degree of certainty that the odds would be greater than 100 to 1." Vincent's account is blunter and more damning, because it names the actual competing authority: they trusted Gallup more than they trusted their own machine.

Second, a search for a bug, and then — finding none — a deliberate degradation of the model. This is the part that never makes it into the retellings and it is the technically interesting part. The team looked for programming errors, found none, and then, in Draper's on-air words, "we asked UNIVAC to forget a lot of the trend information that we had put into it, assuming that that was wrong. So UNIVAC worked on a smaller margin of knowledge. This won't give a wrong answer, but it'll throw the odds to the extent that you saw." Concretely: the component that had detected a large swing away from historical voting patterns and propagated that swing into areas that had not yet reported — the component that was correct, and that was the entire reason for having a model rather than a tally — was watered down until the output stopped being alarming. The 314–217 forecast at 10:30 and the 270–261 toss-up at 11:30 are both products of the degraded model. As raw votes accumulated, they eventually overwhelmed the handicap and the odds swung back to 100 to 1 on their own.

Third, the accounts of all this immediately began to disagree with each other. Chinoy documents the divergence at length and it is worth recording, because it is the mechanism by which a manufactured story sets. Draper's written account compressed the whole evening into a single mistaken prediction corrected after ten minutes. Lukoff put the eight-to-seven forecast on air at 9:15 p.m. Mickelson had the eight-to-seven going to CBS at 9 p.m. and being corrected fifteen minutes later, and attributed the error to an extra zero on Stevenson's New York total. Wulforst, a later Sperry Rand publicist who was not present, described an arbitrary change to the extrapolation factor, a 317-vote forecast, a further tweak, and Draper going on at midnight. Chinoy's verdict: "A number of accounts of behind-the-scenes activity and how that translated into results released to the CBS audience are at variance — in the number, chronology, and timing of various predictions — both with each other and with what can be seen in the broadcast itself." Even the time of the famous forecast is unresolved: the Woodbury–Mitchell report survives in two versions identical but for that one detail, one saying 8:30 p.m. and one saying 9:15 p.m., and the archived printouts carry a handwritten "8:30 PM" in different handwriting on different copies and no machine-generated timestamp at all. Chinoy notes, dryly, that there do not appear to have been 3.4 million votes available to analyse until well after 8:30, before you allow for preparing the data and the five minutes the program took to run.

The confession, and the sentence that became the legend

Cronkite set it up as a joke — "perhaps, it is suggested, Univac and Charlie Collingwood can tell us who's going to win the World Series — next year." Collingwood handed it over with an apology of his own: "Now I've got a lot of faith in this machine, and I was sadly shaken by this aberration that it had. So let's go down to Philadelphia and see whether we can get an explanation of what happened to UNIVAC from Mr. Arthur Draper… Art, what happened there when we came out with that funny prediction?"

Draper, on camera, seated by the console with a headset and a pleasant smile:

> Well, we had a lot of troubles tonight. Strangely enough they were all human > and not the machine. When UNIVAC made its first prediction with only three > million votes in, it gave five states for Stevenson, 43 for Eisenhower, 93 > electoral votes for Stevenson, 438 for Eisenhower. We just plain didn't believe > it. […] As more votes came in, the odds came back, and it was obviously evident > that we should have had nerve enough to believe the machine in the first place. > It was right, we were wrong. Next year we'll believe it.

Collingwood: "You mean four years from now, Art?" Draper: "Well, yes."

Then Collingwood, in New York: "I think that was very handsome of Mr. Draper and the people down there to take the blame for UNIVAC's error. It just goes to show that you can't outsmart the machine." And Cronkite, closing the loop: "I also like that report of Mr. Draper in Philadelphia, proving that humans shouldn't tamper with the thinking of electronic machines."

Everything the story became is in those four minutes. Draper was speaking for the manufacturer of a product with almost no commercial track record, in a moment where the alternative narrative available to the audience was the machine we rented spent the evening being wrong on television. He converted a bad night into a parable in which his product was the hero and his own staff the villains, and he did it live, and it worked so completely that seventy-four years later the parable is what people mean when they say "UNIVAC." It was carried the next day on CBS radio affiliates, picked up by the Associated Press, and printed in the Baltimore Sun under a headline — "Electric Brain's Aberration Is All Its Masters' Fault" — that shows how fast it set. Remington Rand's publicist Arch Hancock was already writing it up for the trade press within weeks. Chinoy's term for the result is exact: a "standard, incomplete, and, sometimes, inaccurate narrative."

Why a reading would cite it

The proposal's cite case does not survive the record, and the honest thing is to say so before offering the replacement. A reading should not cite this entry as "a correct machine prediction suppressed because the humans did not trust it." On the evening's actual record: the journalists were not behind the machine, they were ahead of it — Lubell at 8:35, the Hartford Courant calling Connecticut at 7:40, the Associated Press writing that Ike had crashed into the lead by 8 p.m., Murrow declaring the election over fifteen to twenty minutes before the machine did. The people who disbelieved the forecast were the people who built it. And the reason they disbelieved it was not a general suspicion of machines but a specific and rather modern one: they had a competing model — Gallup's — that they trusted more, and they resolved the conflict in favour of the prior. If you cite this as an automation-bias mirror you will be citing the press release.

Four things it is good for.

1. It is the founding case of a vendor's postmortem becoming the historical record

This is the strongest reason to hold it, and in 2026 it is not a historical curiosity, it is the daily situation. When a model does something notable — good or bad — in a deployed system, the first and often only detailed account of what happened comes from the organisation that built it, is delivered under commercial pressure, and is written before anyone outside has access to the logs. Draper's account had all three properties. It became the record for fifty-eight years, until someone went back to the tapes. When it was finally checked, it was wrong in the direction that flattered the vendor: it compressed an evening of bad public output into one suppressed good result, omitted that the aired forecasts had been worse than the human commentary, omitted that the model had been deliberately degraded, and got its own timings wrong in ways its own internal report contradicted.

A reading meeting a vendor's incident writeup, a model card's account of a failure, or a lab's retrospective on why a system behaved unexpectedly should cite this entry for the base rate: the vendor's own account, delivered fast and under pressure, has a documented history of being the version that sticks and the version that is wrong. That is a claim about incentives, not about honesty, and it does not require anyone to have lied. Draper probably believed every word.

2. The failure was integration, not trust — and Chinoy's diagnosis is the transferable one

Chinoy's sharpest observation is not about disbelief at all. The UNIVAC team had anticipated many sources of error and built detection for them. What they had not anticipated was a correct answer surprising enough not to be believed, and their implicit method for resolving that was, in his words, "to compare the prediction to expectation rather comparing the prediction to other sorts of fresh data points, easily obtained, by listening to the radio or watching television or brainstorming with experienced reporters and political observers." Draper dismissed as impossible in real time any solution involving checking the calculations. There was no channel by which the Philadelphia team and the CBS newsroom could jointly assess whether the forecast and the reporters' reads agreed. The evidence that would have validated the machine was being spoken aloud on the network the machine was working for, and nobody in Philadelphia was listening to it.

CBS had a camera crew in Philadelphia and no correspondent. The computer was, as Collingwood put it, "sitting there in his corner." Chinoy's conclusion — "the UNIVAC had been added to the election-night formula – to attract and interest viewers and generate headlines – but it was not well integrated into the journalism of election night" — is the citable sentence. A reading that meets an AI deployment sitting beside a workflow rather than inside it, evaluated against what its owners expected rather than against the other evidence available in the same building, has a seventy-four-year-old precedent with a name.

3. It is a hard case on what "the machine was right" is allowed to mean

Section 3 grades this properly, but the headline belongs here: the celebrated forecast was right about the electoral vote to within four votes and wrong about almost everything else, including which states each candidate would win and how many Americans would vote. Its popular-vote projection was short of the actual turnout by roughly ten million votes. The aired forecasts were worse. A reading that wants to say "the model got it right" should be made to say which output it means. This is the same discipline LENSES.md demands under record the magnitude, never the consequence, and the same one bin/ethos-check.py enforces against fake precision — and it is a rule this entry can be cited for, because the entry is the case where an aggregate number that looked like a bullseye was sitting on top of components that were materially wrong in both directions.

4. It is the first appearance of the computer as television's guest

Not the reason the proposal wanted it, but the historically largest thing about the night. Chinoy's dissertation argument is that the machine was on that stage "not so much to revolutionize news reporting as to capture public attention," and that "the new technology's technical features were less a driving force for adoption than its usefulness as a wonder and as a symbol to enhance the prestige of its adopters." The word "UNIVAC" became the American word for computer for a decade off the back of one evening in which the machine was, on air, wrong twice and late once. That is a fact about the audience, not the machine, and it is a useful thing to hold when a reading is asked to distinguish a capability from its reception. Both chatgpt-2022 and rosenblatt-nyt-1958 cover neighbouring ground; this is where the pattern starts.

What it got right, and what it got wrong

Not required for moment, but this entry lives or dies on it, and the prediction discipline is imported deliberately: the claim in its own words, the date it was made, the date it was due, what actually happened, the predictor's self-grade, and an independent grade, with the divergence stated.

The benchmark. Certified national result, 4 November 1952: Eisenhower 34,075,529 popular votes (55.18 per cent) and 442 electoral votes from 39 states; Stevenson 27,375,090 (44.33 per cent) and 89 electoral votes from 9 states. Total turnout 61,751,942, about 63 per cent of the eligible population, up more than eleven points on 1948.

The withheld forecast — the one the legend is about

Claim: Eisenhower 438 electoral votes, Stevenson 93; Eisenhower 43 states, Stevenson 5; popular vote roughly 33 million to 19 million; odds at or above 100 to 1. Made: evening of 4 November 1952, time unresolved between 8:30 and 9:15 p.m. EST, on 3,398,745 votes counted — about 5.5 per cent of eventual turnout. Due: the same night.

Right, and impressively so, on the electoral vote: 438 against 442, an error of four. That is the number the legend runs on and the legend is entitled to it. At 5.5 per cent of the vote, from a machine with a few thousand tubes and a model assembled in a hurry, four electoral votes is a genuinely good result and no amount of context makes it not one.

Wrong on the state map, in both directions, and the errors cancelled. It gave Eisenhower 43 states; he won 39. It gave Stevenson 5; he won 9. Those two facts cannot be reconciled with a four-vote aggregate error unless the mistakes ran both ways: had UNIVAC simply handed Eisenhower four extra states that Stevenson actually carried, Eisenhower's predicted electoral total would have come out above 442, not below it. So at least one state Eisenhower actually won was assigned to Stevenson, and it was a large one — five predicted Stevenson states carrying 93 electoral votes averages 18.6 votes each, against 9.9 for the nine states he really won. The inference is arithmetic; the state-by-state list is not in any source read for this entry, and a session with access to the full Woodbury–Mitchell report should reconstruct it. But the shape is clear enough to state: the near-perfect total was sitting on a materially wrong map, and that is a different thing from being right.

Wrong on the popular vote, badly, and this is never mentioned. Roughly 33 million to 19 million is a 63/37 split against an actual 55.18/44.33 — eight points too favourable to Eisenhower — and a total of about 52 million against an actual 61.75 million, understating the number of Americans who would vote by nearly ten million. Every UNIVAC forecast that night, including the ones that aired, projected a turnout denominator between 51 and 52 million while Cronkite was telling viewers on air to expect 55 to 60 million. The machine was systematically wrong about the size of the electorate all evening and nobody noticed, because everyone was looking at the electoral column.

Wrong, or at least unverifiable, on the odds. "100 to 1" was, in some surviving printouts, rendered "00 to 1" — the programmers had not allowed three digits for the odds field, because they had not imagined the model would ever return a number that large. A confidence figure that the output format cannot represent is not a calibrated probability; it is a saturated counter. Citing "100-to-1 odds" as evidence of the machine's confidence is citing a display bug.

The forecasts that actually aired

~10:30 p.m.: Eisenhower 314 electoral votes to 217, popular vote 27,445,000 to 24,456,000. Directionally right, quantitatively poor — 128 electoral votes short of the result — and, crucially, weaker than what unaided humans on the same broadcast had already said.

~11:30 p.m.: 24 states each, Stevenson leading the popular vote 26 million to 25 million, Eisenhower 270 to 261 in the electoral college at eight-to-seven odds. Flatly wrong, at the hour when the outcome was no longer in doubt to anyone watching. This is the output of the degraded model and it is the single worst forecast issued by any participant, human or machine, on either network that night.

~11:45 p.m. and just after: 100-to-1 odds; Eisenhower 424 electoral votes to 107, popular vote 29 million to 23 million, Stevenson taking eight states. Right in direction, still 18 electoral votes off, still ten million votes short on turnout, and arriving fifteen to twenty minutes after Murrow had said the same thing with a stopwatch in his hand.

After midnight, Massachusetts: wrong, and wrong for the most ordinary reason there is. The state's vote counts had been corrupted on the way in and stayed corrupted all night, and the error-detection did not fire. The forecast was faithful to its inputs and its inputs were garbage. Of everything on this list this is the failure that generalises furthest.

The self-grade against an independent grade

The self-grade is Draper's, delivered live at about 12:30 a.m. on 5 November 1952: "they were all human and not the machine… It was right, we were wrong. Next year we'll believe it." Scored generously and immediately, by the party with the most to gain, on a night when nobody could check.

The independent grade is Chinoy's, published in 2010 as a University of Maryland dissertation and awarded the American Journalism Historians Association's Blanchard prize in 2011, then extended into Predicting the Winner (2024). It is built from the CBS and NBC recordings, the Collingwood papers at the Wisconsin Historical Society, Remington Rand's own internal reports at the Hagley Museum and the Smithsonian, and interviews with surviving participants including Woodbury himself. The two grades differ, and they differ systematically.

| | Draper, on air, 5 Nov 1952 | Chinoy, from the archive | |---|---|---| | Who disbelieved | "we" — unspecified, heard as CBS | Remington Rand's own team in Philadelphia; no evidence CBS knew | | Number of bad forecasts | one, corrected in ten minutes | at least two aired, over roughly an hour | | Why the odds moved | the machine recovered as votes came in | the model was deliberately degraded, then swamped by raw votes | | Humans vs machine | machine ahead, humans behind | humans ahead by up to two hours, on the same broadcast | | Time of the forecast | 8:30 p.m. | unresolved; the vote counts don't support 8:30 |

Every difference runs the same way. That is what makes it worth recording as a pair rather than picking one, and it is the reason the prediction kind requires both.

And the proposal's claims, graded

"Sperry Rand": wrong, by two years and eight months. "One percent of returns": wrong, about 5.5 per cent, and the figure's attachment is disputed. "CBS did not believe it": unsupported and probably backwards — the disbelieving parties were Remington Rand's, and Chinoy is explicit that no document settles whether CBS knew at all. "Did not broadcast the real forecast until the result was in": wrong — it aired before midnight with three hours of broadcast left. "The honest counterweight to every story about people deferring to a model too readily": the entry does not support this and a reading should not use it this way. The one clause that holds up entirely is the last four words of the id: it was election night, 1952.

Commonly misused as

Not required for moment, but this is the entry's central service and it would be strange to omit it.

What the record says. On the evening of 4 November 1952, a computer belonging to a company with a commercial interest in the outcome produced an accurate electoral-vote forecast from early returns; its own operators, trusting a poll over their model, withheld it and then deliberately weakened the model; the weakened model produced two public forecasts, one mediocre and one badly wrong; the network's human analysts, working from precinct comparisons by eye, called the election correctly before the machine did; the machine's landslide forecast reached the air before midnight; and at half past midnight the vendor's executive delivered a graceful on-air account that made the machine the hero and became the permanent record.

What it is repeatedly made to say, in roughly descending order of frequency:

**One thing that is not a misuse**, and is worth defending: it is entirely fair to cite this as an early, well-documented instance of a statistical model outperforming its operators' expectations and being overruled by them. That happened. The caution is only against the further step — that the overruling tells you something general about how humans treat machines — because in this case the overruling tells you something about how humans treat surprises, and the machine happened to be the messenger.

Sources

Provenance is marked. Where something was read directly it says so; where it comes from a search summary, it says that too.