The Seven-Hour President: Should Early Exit-Poll Results Be Embargoed or Broadcast?

This paper examines whether American broadcast networks should keep early exit-poll results secret until polls close, or release them as the data arrive. The question is usually posed as though the only stakes were paternalism versus press freedom. This paper asserts the real difficulty is epistemic and acute given an increasing number of American elections, at all levels of government, no longer resolve on "Election Night," with ever more escalating to the courts to resolve. The early exit poll is at once the most newsworthy and the least reliable artifact a network possesses: in 2004 it made John Kerry, in journalist Mickey Kaus's phrase, "the seven-hour president." Drawing on the 1980 West Coast controversy, the 2000 Florida misfire, the 2004 leak, the constitutional law of exit polling from Burson v. Freeman through CBS v. Cobb, and on the now-routine multi-day counts seen federally in 2020 and locally in the 2026 Los Angeles mayoral primary, this paper defends a qualified position: the networks are right to embargo projections until polls close, but the underlying rationale — and the appropriate object of secrecy — must shift from protecting voters from information toward protecting the public from a specific, predictable category of error in an era when the count itself no longer concludes on election night.

I. The Question and Why It Is Harder Than It Looks

On the afternoon of every recent Election Day in America, a small and growing set of actors — the broadcast and cable networks, the Associated Press, the campaigns, and an expanding penumbra of websites, newsletters, and reporters — possess information the voting public lacks: early exit-poll estimates of who is winning. The convention, honored since the early 1980s, is for networks to not broadcast projections until polls are closed. This paper addresses whether that convention is justified: should early exit-poll results be secret, or broadcast?

The episode that frames the question is 2004. Around midday, the first wave of exit-poll numbers reached the consortium of news organizations and, almost simultaneously, leaked. The early exits showed John Kerry leading in nearly every contested state — even, briefly, in South Carolina, which had not voted Democratic since 1976. Pro-Kerry blogs propagated the figures; the financial markets dropped; the Kerry campaign began, quietly, to celebrate. The journalist Mickey Kaus memorably dubbed Kerry "the seven-hour president." The actual returns told a different story, and a post-election review by Edison Media Research and Mitofsky International found the exit polls overstated Kerry's share in most battleground states, attributing the gap chiefly to differential nonresponse: his supporters were more willing than the opposition's to stop and complete a questionnaire.

The naive framing of the debate pits two values against each other. On one side, press freedom and the public's right to know:

The embargo is paternalistic toward voters and puts journalists in the awkward and unfamiliar position of concealing information from their readers.

— Julia Turner, Slate

On the other, the integrity of the franchise: the fear that early numbers depress or distort turnout among those yet to vote. This paper asserts both framings are incomplete, because both assume early exit polls are roughly accurate information whose disclosure is merely premature. The 2004 episode shows the deeper problem: early exit polls are systematically and predictably unreliable. As a result, the case for their embargo rests less on protecting voters from true-but-early information than on declining to broadcast flawed artifacts as though they are irrefutable newsworthy facts representing election reality.

II. How the System Came to Be: 1980, 2000, 2004

The modern embargo is the sediment of three controversies. Exit polling itself dates to the 1960s — Warren Mitofsky conducted the first major network exit poll during the 1967 Kentucky governor's race — and by the 1970s it was standard network practice. The first crisis came in 1980, when NBC used exit-poll data to project Ronald Reagan's victory over Jimmy Carter at 8:15 p.m. Eastern — roughly three hours before polls closed on the West Coast — and Carter conceded at 9:50 p.m. Eastern, before Western voting had ended. Congress held hearings on whether early projections depressed turnout; the evidence, notably, was mixed rather than uniformly dismissive. A University of Michigan study of that election estimated that roughly a quarter of Westerners who had planned to vote later in the day, and who heard the early call, ultimately did not vote — a real, if contested, effect, and a reminder that "scant evidence" of harm is not the same as no evidence. In response, the networks adopted the self-imposed rule that survives today: no projection of a state until its polls have closed.

That rule has held for more than four decades across the entire industry. ABC, CBS, NBC, CNN, Fox News, and the Associated Press all draw on the same shared exit-polling consortium — first the Voter News Service, now the National Election Pool — and no documented instance of a member network unilaterally breaking the embargo to release early projections has surfaced since the 2004 leak, including from Fox. The embargo is also broader than the presidential race alone: the same consortium fields exit polls for top-of-ticket gubernatorial and congressional contests, so the "no projection" rule is not a presidential-only courtesy. It is, however, purely a voluntary industry norm with no counterpart in FCC regulation — nothing stops an independent blogger, newsletter, or prediction market from publishing whatever it can obtain, and increasingly they do. Notably, the empirical case that the 1980 call for Reagan changed the outcome of any race is thin now and was considered thin then, even if the turnout-depression question itself is less settled than the "scant evidence" framing suggests. The broader unreliability of pre-election polling as a predictor of electoral intent — as distinct from the narrower exit-poll question this paper takes up — is examined at greater length in my companion essay, Counting the People.

The second crisis, in 2000, was about premature projection rather than premature disclosure — a distinction worth making explicit given how similar the two episodes can look. The networks called Florida for Al Gore at roughly 7:50 p.m. Eastern — while polls in the state's heavily Republican Central-time panhandle were still open — based partly on exit data that overstated his lead, then retracted the call less than two hours later as more panhandle returns arrived; Florida itself would not be settled for another thirty-six days. The difference from 1980 is instructive: NBC's 1980 call was, in the end, correct — Reagan did win — so whatever harm it caused was purely a matter of timing. The 2000 Florida call was not simply early; it was wrong. Two acts that look identical on the surface — announcing a winner before polls close — produced two different species of harm: a disclosure problem in 1980, an accuracy problem in 2000. The 2000 failure, compounded two years later by a computer-system meltdown that again crippled the consortium's ability to deliver results in the 2002 midterms, led to the dissolution of the Voter News Service in January 2003 and its replacement the following year by the National Election Pool, with exit polling conducted by Edison Research.

The third crisis, 2004, was about leakage: the embargo on broadcast held, but the data escaped to blogs and trading desks during the afternoon. The institutional response was procedural and remains in force. Since that election, the handful of analysts who first see the raw numbers have been physically sequestered — placed in a "quarantine" room at an undisclosed location, their phones and BlackBerrys confiscated, wireless devices disabled, minders present even for bathroom trips, with results released only in the late afternoon — precisely so that nationwide exit results cannot circulate while voting continues. The lesson the industry drew was not that the public could not handle the numbers, but that the numbers themselves required "a fine comb" before release — an implicit concession that early exit data are not yet fit to be reported as findings.

III. The Constitutional Floor: What the Networks May Be Compelled to Do

Any normative debate over embargo operates against a constitutional backdrop that sharply limits coercive solutions. The collection and publication of exit-poll results is core protected speech, and government efforts to suppress it have fared poorly. In Daily Herald Co. v. Munro (1988), the Ninth Circuit struck down a Washington statute restricting exit polling within three hundred feet of polling places, holding that exit polling constitutes protected speech and that the restriction was content-based, overbroad, and not the least restrictive means of preserving order at the polls. Two decades later, federal courts enjoined enforcement of similar statutes in Florida and Nevada in CBS v. Cobb (S.D. Fla. 2006) and ABC v. Heller (D. Nev. 2006), with Nevada's secretary of state ultimately agreeing to a permanent injunction against the law's enforcement.

The leading contrary precedent is narrower than it first appears. In Burson v. Freeman (1992), the Supreme Court upheld a Tennessee statute barring solicitation of votes and display of campaign materials within one hundred feet of a polling place, finding that even under strict scrutiny the state's compelling interests in preventing voter intimidation and election fraud justified the "minor geographic limitation." Critics note an important limit on Burson, however: the Tennessee law it upheld targeted campaigning and solicitation before a ballot is cast, not the reporting of a vote already cast. Exit polling is the latter. Commentators have used this distinction — pre-vote persuasion versus post-vote reporting — to explain why Burson does not control cases about exit polling, even though both involve speech near a polling place.

A Congressional Research Service assessment captured the resulting asymmetry: the First Amendment would generally forbid Congress from prohibiting the media from interviewing departing voters or from reporting the results, though Congress could ban voter solicitation near the polls and might be able to deny the media access to official ballot counts before polls close. This constitutional floor is also office-agnostic: the reasoning in Munro, Cobb, and Heller turns on the nature of the speech and its location, not on whether the race atop the ballot is presidential, gubernatorial, or municipal — a point the 2026 Los Angeles mayoral primary, discussed below, illustrates directly. The constitutional upshot is decisive for the present question: the embargo cannot be a legal mandate imposed on the networks; it can only be a self-imposed norm. The debate is therefore not "should the law require secrecy" but "should responsible broadcasters choose it" — a question of journalistic ethics and democratic prudence, not of permissible regulation.

IV. The Affirmative Case for Broadcast

Three arguments support releasing early exit-poll results. The first is the anti-paternalism principle. In a free press system, the default is disclosure; withholding newsworthy information from the public on the theory citizens cannot be trusted to handle it is, as Sasha Issenberg argued in defending the 2016 Votecastr experiment — in which Slate published live, unofficial turnout-based projections throughout Election Day — a "self-imposed gag order" resting on the paternalistic premise that "voters cannot be trusted with live information." The second is consistency. The networks already report a flood of real, consequential election-day information — turnout figures, anecdotal reports of lines and machine failures, early-vote totals where state law permits — so a blanket embargo on one category of data is arbitrary. The third is the leveling argument, and it has only grown stronger with time: prediction markets, poll aggregators, and anonymous accounts on social media now routinely circulate real-time turnout modeling and unofficial tallies well before any network projection, in 2024 as in earlier cycles. Since campaigns, traders, and insiders obtain such numbers regardless of any network embargo, the embargo does not produce secrecy so much as asymmetry — restraining only the platforms bound by the National Election Pool's contract while leaving the wider information ecosystem free to publish. The 2004 leak is the paradigm case: the information moved, but only to those with access.

The empirical record behind the turnout fear that motivates embargo is genuinely mixed rather than one-sided. Some studies of the so-called West Coast effect have struggled to find robust evidence that early projections meaningfully depress or alter turnout — but as noted above, the University of Michigan's contemporaneous study of the 1980 election found a real, non-trivial effect among Westerners who heard Carter's early concession. The honest summary is that the turnout question remains empirically contested, not settled in either direction; anyone citing "scant evidence" for the anti-paternalist position should also reckon with the studies that found otherwise.

V. The Case for Embargo, Reframed

The strongest case for embargo does not depend on the turnout-suppression hypothesis, which is empirically contested, nor on paternalism, which is normatively unattractive. It depends instead on a fact about the data themselves: early exit-poll results are not merely premature true information but are instead predictably biased information, and broadcasting them is closer to publishing a known error than to disclosing a withheld fact.

A. The Bias Is Directional, Not Random

If early exit numbers were merely noisy — wrong in random directions — the disclosure debate would reduce to a question of communicating uncertainty. But the 2004 evidence indicates a directional bias: differential nonresponse, in which one candidate's supporters are systematically more willing to be interviewed, pushes early estimates toward the candidate whose supporters are most willing to communicate their voting decision. Edison and Mitofsky's evaluation attributed the 2004 overstatement of Kerry to exactly this, compounded by a programming error that overstated the share of women among early respondents and an operational bias resulting from younger interviewers tending to attract younger, more Democratic respondents. A network that broadcasts the raw afternoon numbers is therefore not informing the public of the state of the race; it is broadcasting an estimate with strong prior reason to believe the estimate is skewed, before the corrective discipline of actual returns can be applied. Edison's own analysts in 2004 telephoned the networks to warn them and share their doubts about some of the reported results.

B. The Object of Secrecy Should Be the Projection, Not the Datum

This reframing dissolves part of the dispute. The defensible embargo is not on the existence of exit data — networks already air exit-derived demographic color ("the electorate is younger than in 2020") during the afternoon — but on horse-race projections of who is winning before polls close and before returns can anchor the estimate. The 2000 Florida debacle was a projection error; the 2004 episode was a disclosure-of-unanchored-estimate error. In both, the harm flowed from presenting a provisional, bias-prone number as a finding. The embargo, properly understood, is a discipline against premature conclusions, not against information as such.

C. The Cost Is Reputational and Civic, Not Merely Electoral

Even if early disclosure does not change a single vote, it carries a civic cost that the turnout-centric debate ignores: it manufactures a false narrative that must later be publicly reversed, and each such reversal erodes trust in both the press and the count. The "seven-hour president" was not a turnout problem; he was a credibility problem. In an environment where confidence in elections is already strained, broadcasting a number likely to be contradicted hands a ready-made grievance to anyone inclined to allege manipulation — the apparent swing from the afternoon "landslide" to the night's defeat becomes, in bad-faith hands, evidence of theft. This is precisely what happened in 2020, when a genuinely accurate but partial count was seized on as proof of fraud once it "shifted." The dynamic was muted but recognizable again in 2024: national House and California state races that lean heavily on mail ballots processed over the following weeks again saw margins move as later ballots were counted — a reminder that the red-mirage/blue-shift pattern is a structural feature of how different coalitions vote by mail versus in person, not a one-time 2020 artifact. This cost has grown sharply with a development the original 1980 framework never anticipated: the disappearance of election night itself.

D. A Practitioner's Parallel: Why Experimenters Don't Peek

The embargo's real logic — that a provisional number, examined before its collection process has run its course, is not merely incomplete but systematically misleading — has an exact analogue in the discipline I practice professionally. At Google, and now at Meta, where I lead ads-growth work for Media Networks, product decisions are justified by online controlled experiments (A/B tests) run across populations that can rival a national electorate in size. The central operational rule of that discipline is that a live experiment's dashboard should not be treated as a verdict while data are still accumulating — a rule the television networks adopted for exit polls in 1980, arrived at independently and for a structurally identical reason.

Johari, Koomen, Pekelis, and Walsh's study of Optimizely's production experiments showed why the rule is necessary. Standard statistical tests assume the decision to stop looking at data is independent of the data itself, but in practice engineers and product managers check a running experiment continuously and are tempted to call a result the moment it first crosses a significance threshold — "peeking." Because early data are noisier and more prone to transient swings than a fully accumulated sample, peeking inflates the rate of false positives well beyond the nominal five percent; the industry's fix is either to commit in advance to a fixed sample size and refuse to look early, or to adopt "always-valid" statistical methods built to tolerate continuous monitoring without inflating the error rate. Either way, the operative principle is the one this paper has already named: the danger is not the existence of early data but the act of treating early data as a finding before the process that generates it has finished.

The Los Angeles mayoral primary, discussed below, is a peeking problem in miniature. Each daily tally as Los Angeles County processed its mail ballots was a real, unbiased snapshot of votes counted so far — but a network, or a resident, or a reporter who mistook Wednesday's snapshot for Sunday's answer would have committed exactly the error that always-valid methods in online experimentation are built to prevent, and that the exit-poll embargo has enforced by different means since 1980.

VI. The Decisive Modern Complication: When the Count No Longer Ends on Election Night

The entire embargo framework was built for a world in which polls closed, votes were counted, and a winner emerged that night. That world is gone. The rise of mail and provisional voting means in many jurisdictions the count now spans days or weeks, and — critically — the ballots counted first differ systematically from those counted later. This interacts with exit polling in a way that makes careless disclosure far more dangerous than it was in 2004.

A. The Federal Case: 2020 and the "Red Mirage"

Because Democratic voters became markedly more likely to vote by mail and Republicans to vote in person, and because many states count in-person ballots first, early returns in 2020 showed Republican leads that eroded as mail ballots were tallied — the phenomenon variously called the "red mirage" or "blue shift." An MIT Election Data and Science Lab analysis by Curiel, Stewart, and Williams confirmed that states were slower to report where they had large mail volumes, prohibited preprocessing, or accepted late-arriving ballots, and that counties Biden won counted more slowly than counties Trump won. The presidential race was not projected until November 7, four days after Election Day; Alaska was not called until November 11. President Trump seized on the visible swing — true to the data, innocent in its cause — to advance baseless claims that the election was stolen. The 2020 episode demonstrates that even accurate early numbers, if their systematic relationship to later numbers is not understood, can be weaponized into a fraud narrative.

B. The State Cases: Texas, Georgia, Tennessee

The state level shows how the timing of disclosure depends on idiosyncratic local law. Texas, a Central-time state, releases its large early-vote and mail-ballot totals immediately once polls close at 7:00 p.m., producing an enormous initial tranche that can look like a landslide before many Election-Day precincts have reported — an artifact of reporting order, not of the electorate's actual division. Georgia's 2021 election law (SB 202) allows counties to begin processing — though not tabulating — absentee ballots up to two weeks before Election Day, which has the effect of compressing the post-Election-Day reporting window that helped generate 2020's blue shift, even though the law's authors framed the change primarily as a response to the administrative strain of a record 1.3 million absentee ballots in 2020 rather than as a fix for narrative risk. Tennessee, the home of Burson v. Freeman, combines an early poll-closing time with a constitutional regime that protects exit polling itself; there, the only available lever over premature narratives is the networks' voluntary restraint rather than any law.

C. The Local Case: The 2026 Los Angeles Mayoral Primary

The most vivid recent illustration is local. In the June 2, 2026 Los Angeles mayoral primary, incumbent Karen Bass led comfortably and was projected on election night to advance to a runoff, but the contest for the second runoff slot was a near-deadlock. For several days after election night, early returns showed former reality-television figure Spencer Pratt holding second place over City Councilwoman Nithya Raman. As Los Angeles County — with roughly six million registered voters, more than most states — processed its vote-by-mail and provisional ballots over the following week, Raman closed the gap on Friday, narrowed it further on Saturday, and overtook Pratt on Sunday, June 7, with the county registrar's tally showing Bass at 34.68 percent, Raman at 27.12 percent, and Pratt at 26.69 percent, a margin of roughly 3,100 votes with hundreds of thousands still to count. California law gives counties up to thirty days to canvass, accepts ballots postmarked by Election Day for several days afterward, and requires signature verification on each envelope — so vote totals were expected to keep moving for weeks.

The Los Angeles case is the 2004 problem inverted and intensified. In 2004 the early exit poll was wrong and the night's count was right; in Los Angeles the early count was "wrong" (in the sense of non-final) and only the patient, days-long canvass was right. In both, the danger is identical: a provisional, systematically unrepresentative early number presented as if it settled the matter. A network that had projected Pratt into the runoff on election night would have committed precisely the error the embargo norm exists to prevent — and would have done so using real returns, not exit polls. The lesson is that the disclosure problem has migrated from the exit poll to the count itself.

VII. Synthesis: A Qualified Defense of Embargo

The question — secret or broadcast? — admits no absolute answer, but the balance of considerations supports a qualified embargo, justified on revised grounds. The networks should continue to withhold horse-race projections, whether derived from exit polls or from partial counts, until polls close and until the relationship between counted and uncounted ballots is understood. But they should abandon the original paternalistic rationale — that voters must be shielded from information, which is both empirically contested and normatively unattractive — in favor of a journalistic one: that broadcasting a number known to be systematically unrepresentative is not disclosure but the publication of a likely error.

Three commitments follow. First, the embargo should attach to conclusions, not data: networks may and should report exit-derived context and the mechanics of the count, while declining to declare winners prematurely. Second, transparency should replace concealment wherever possible — rather than hiding early numbers, broadcasters should explain, in advance and in real time, why early exit polls lean one way and why early counts shift, inoculating the audience against the fraud narratives that 2020 showed will rush to fill the vacuum. Third, the norm must be updated for the post-election-night era: the most consequential "early result" is no longer the afternoon exit poll but the election-night partial count, and the discipline of restraint must extend to it.

The deepest point is that secrecy was always a crude proxy for the real goal, which is accuracy responsibly communicated. The 1980 framework treated the problem as one of timing — hold the true number until everyone has voted. The modern problem is one of epistemology — distinguish the provisional from the final, the biased from the representative, and refuse to let the first masquerade as the second. So reconceived, the answer to whether early exit-poll results should be broadcast is: not as projections, not yet, and not without the context that turns a misleading number into an understood one. Karen Bass's challenger was not, after all, the seven-hour runoff candidate; the count simply had not finished speaking.