Why Americans Hold False Beliefs: New Hampshire Found the Answer and It Didn’t Matter

Brainwashing is the wrong word. Four replicated findings explain what a state-ordered ballot audit could not fix.

New Hampshire ran the experiment. In 2021 the state did something almost no state had done: it opened a disputed election down to the paper and let three outside experts take it apart in public. The town was Windham, the race four state representative seats in Rockingham County District 7. The machine count on election night and the hand recount nine days later disagreed by hundreds of votes, and nobody could say why.

The answer arrived that summer, and it was boring. Windham had used a folder-inserter machine to prepare thousands of pandemic-era absentee ballots, and it put creases through the vote targets. Optical scanners read some creases as marks. The auditors called the fold lines the primary root cause, found no malware on any tabulator, and reported nothing unusual about the ballot paper. The recount never changed who won.

That is a complete answer, ordered by statute and published in full by named experts. It should have ended the argument. Instead the finding was picked up nationally as proof that voting machines invent votes, and the belief it corrected is still in circulation five years later. If you want to understand why Americans hold false beliefs, start there: not with a claim that survived because nobody checked, but with one that survived because somebody did.

Brainwashing is a Cold War marketing term

The word people reach for is “brainwashed.” It is the wrong word, and it has a paper trail. Edward Hunter, a journalist who had worked in wartime intelligence, popularized it in the Miami Daily News in September 1950, reporting that Chinese communists had perfected a technique for rewriting the mind. Washington believed him enough to spend two decades on MKUltra trying to build the machine. It never worked. What Hunter described does not exist.

What replaced it in the research is less cinematic and worse. Ordinary cognition, working exactly as built, produces large and predictable errors inside an information environment like ours. Nobody is overriding anybody. The mind is doing its job. Four findings carry the weight, and all four have survived decades of replication.

You start with the conclusion

The foundation is Ziva Kunda’s 1990 review in Psychological Bulletin, The Case for Motivated Reasoning. Kunda argued that people do not begin neutral and follow evidence downhill. They begin with an answer they would prefer and recruit memory and inference in its service. The single constraint is that the conclusion has to be defensible. People do not believe anything they want. They believe anything they can build an argument for.

That is not a character flaw. It is a system built to keep people inside their group, in a species where exile was lethal.

Smarter partisans are further apart, not closer

The result every serious argument about political belief has to answer is Dan Kahan’s. In Motivated Numeracy and Enlightened Self-Government, published in Behavioural Public Policy with Ellen Peters, Erica Cantrell Dawson, and Paul Slovic, 1,111 American adults got the same hard two-by-two table of numbers. Half were told it showed a skin rash cream. Half were told it showed a city ban on carrying concealed handguns. The numbers were identical.

On the skin cream the pattern was the expected one: the better you were with numbers, the more often you got it right. On the gun ban it broke. Subjects sorted by politics, and the sorting worsened as numeracy rose. Among the most numerate partisans, data that flattered your side made you likelier to read it correctly by about 45 percentage points. Among the least numerate that advantage was roughly 25. High-numeracy partisans handed threatening data did no better than anyone else.

Kahan called it identity-protective cognition. When a factual question becomes a badge of membership, the mind stops asking what is true and starts asking what it costs to say so. Rational for one person, a disaster for 1.4 million trying to agree on anything.

Corrections do not delete anything

Removing a false belief is harder than installing one. The classic demonstration is Hollyn Johnson and Colleen Seifert’s 1994 warehouse-fire study: readers learn a fire probably started with volatile materials in a closet, then learn plainly that the closet was empty. They remember the correction and repeat it back. Then they reach for the volatile materials anyway.

That is the continued influence effect, reviewed by Stephan Lewandowsky and colleagues in Psychological Science in the Public Interest in 2012 and updated by Ullrich Ecker and colleagues in Nature Reviews Psychology in 2022. Corrections usually reduce reliance on bad information. They rarely erase it, and in some studies they do nothing.

The leading explanation is mental models. A retraction punches a hole in the story you were carrying. Given a coherent account containing a discredited claim or an incoherent one with a gap, people keep the coherent version. The correction loses on points to the cost of the hole.

Windham cuts both ways here. The audit supplied a replacement cause, the thing the research says works. Not the machines are fine, but here is the folder-inserter, here is the crease, here is the scanner reading the crease as a mark. That beats what most newsrooms manage. It still lost, because by mid-2021 the claim had stopped being a question about scanners and become a marker of political identity, where Kahan’s mechanism takes over. The Union Leader reported that the folded-ballot finding was seized on nationally to renew claims of 2020 fraud, the opposite of what the auditors found.

Repetition beats knowing better

The fourth mechanism is the cheapest to exploit. Lynn Hasher, David Goldstein, and Thomas Toppino showed in 1977 that plausible statements were judged likelier to be true simply for having been seen before, true or not. A 2010 meta-analysis of 51 studies put the effect at d = 0.39 within items and d = 0.50 between them; a 2021 systematic map found it sprawling from trivia to news headlines.

The hardest result for anyone who thinks better information fixes people came in 2015. Lisa Fazio and colleagues, in the Journal of Experimental Psychology: General, showed that knowledge does not protect against illusory truth. Subjects who could name the Pacific as the largest ocean still drifted toward believing the Atlantic was. The authors called it knowledge neglect. What you know sits in one place. Fluency works somewhere else.

The honest counterpoint

The literature is not unanimous, and a piece pretending otherwise is selling something. Gordon Pennycook and David Rand, in a 2019 Cognition paper titled Lazy, not biased, tested 3,446 subjects and found that higher scorers on the Cognitive Reflection Test spot fake headlines better whether or not the headline flatters their politics. Their reading: the problem is not that Americans weaponize reasoning, but that they never engage it.

The two do not fully reconcile. The best synthesis available is that they work at different layers: identity-protective cognition governs evidence already coded as tribal, analytic laziness governs everything not yet coded that way. The first explains Windham. The second explains why a stranger’s screenshot circles a town Facebook group before anyone checks the date on it.

Mechanism	Finding	Source
Motivated reasoning	Cognition serves a preferred conclusion that must stay defensible	Kunda, 1990
Identity-protective cognition	Reasoning ability widens the partisan gap	Kahan et al., 2017
Continued influence	Corrections shrink reliance on bad information, seldom erase it	Johnson & Seifert, 1994
Illusory truth	Repetition raises perceived truth even when you know better	Fazio et al., 2015

The environment is the other half

A brain that works this way stays calibrated when its inputs are. The damage comes from the collision between these mechanisms and an information supply that has worsened at both ends: fewer sources putting friction on a false claim, more channels repeating it.

Start with the referee. Gallup’s September 2025 survey found 28 percent of Americans expressing a great deal or fair amount of trust in the mass media to report news fully and accurately. That is the first sub-30 reading in roughly half a century of asking, down from 31 a year earlier and 72 in 1976. Republican trust registered 8 percent. Independents, 27. A bare majority of Democrats, 51.

Then the supply. Northwestern’s Medill Local News Initiative counted 136 newspapers lost in the year before its 2025 report and close to 3,500 since 2005, leaving 213 counties with no local news source, another 1,524 with exactly one, and about 50 million Americans with little or none. Public media took its own hit: after Congress clawed back $1.1 billion, the Corporation for Public Broadcasting announced its wind-down on August 1, 2025 and ended most staff positions that September. NHPR’s president said the closure did not change the station’s course but named the shared services now unfunded, including training and the satellite plumbing connecting stations to each other.

Now the other end of the pipe. Two days before New Hampshire’s January 2024 primary, thousands of Granite State households answered the phone and heard a synthetic Joe Biden telling them to save their vote for November. NBC News, which broke the story, called it the first known use of a deepfake in national American politics. It was aimed here. Attorney General John Formella indicted the consultant behind it, Steven Kramer, on 26 counts; the FCC proposed a $6 million fine.

In June 2025 a Belknap County jury acquitted him on every count. His defense argued the primary was an unsanctioned straw poll outside the voter-suppression statute, and that a message never naming Biden impersonated no candidate. Formella said the state’s “commitment to enforcing election laws remains steadfast.” Kramer told the Associated Press he will not pay the federal fine. The state where the technology debuted also produced the first proof that the legal backstop has holes in it.

What works, including the disappointing parts

Interventions with real support exist. All are smaller than their press coverage.

Accuracy prompts. Nudging people to consider whether a claim is true before scrolling on reduced misinformation sharing in work published in Nature in 2021. A preregistered replication by Jon Roozenbeek, Alexandra Freeman, and Sander van der Linden failed at first pass and found a real but small effect only after roughly doubling the sample. Useful, not a cure.

Prebunking. A weakened dose of a manipulation technique, delivered before someone meets the real thing, produces measurable resistance. Roozenbeek, van der Linden, and colleagues tested five short videos in Science Advances in 2022, each targeting a tactic rather than a topic: emotional manipulation, incoherence, false dichotomies, scapegoating, personal attacks. Teaching the pattern instead of chasing the claim is what lets it scale.

Replacement, not negation. The correction research is consistent: give people a competing causal story, not a denial. Windham’s auditors did it right. That it still lost tells you the ceiling, not that the method fails.

What does not work is calling your neighbors brainwashed and waiting. Identity-protective cognition strengthens when the price of changing your mind is your standing at the transfer station.

What this obligates a newsroom to do

I run a one-person newsroom in Northfield, so let me be specific about my own shop.

Never put a false claim in a headline, even to knock it down; the headline is the repetition and the knockdown is the fine print. Publish the replacement cause, not the denial. Keep the fact-check inside the article where readers hit it, which is why the table below is not an appendix. Name sources instead of writing “critics say,” and link documents so a skeptic can go around me to the record. Correct fast and in public. A correction I bury is a correction that does not exist.

New Hampshire voters are not more gullible than they used to be. They are running the same equipment their grandparents ran, with fewer local reporters, weaker institutional trust, cheaper fakes, and a legal system that just showed it may not punish the first person who used one on them. The research lets nobody off the hook. It sets the difficulty level, and that level is higher than anyone selling media literacy admits.

— Dexter Dow, Granite State Report

Fact check

#	Claim	Status	Source
1	Windham audit examined the Nov. 2020 Rockingham District 7 state rep race; recount did not change the outcome	VERIFIED	Verified Voting summary of the SB 43 forensic audit report (CHECKED)
2	Fold lines through vote targets were the primary root cause; no malware on any tabulator; ballot sample showed nothing unusual	VERIFIED	SB 43 audit report via Verified Voting; NHPR, July 13, 2021 (CHECKED)
3	Folder-inserter machine creased pandemic-era absentee ballots off the score lines	VERIFIED	NHPR interview with auditor Harri Hursti, July 13, 2021 (CHECKED)
4	The folded-ballot finding was used nationally to renew 2020 fraud claims	ATTRIBUTED	Union Leader contemporaneous coverage, May 27, 2021; paywalled (CHECKED)
5	Edward Hunter popularized “brainwashing” in the Miami Daily News, September 1950	VERIFIED	Smithsonian Magazine history of the term (CHECKED)
6	Kunda (1990) established the motivated-reasoning framework in Psychological Bulletin	VERIFIED	PubMed record, Psychol Bull 108(3):480–498 (CHECKED)
7	Kahan et al. study used 1,111 U.S. adults; skin rash cream vs. concealed-handgun ban framings of identical data	VERIFIED	Behavioural Public Policy 1(1):54–86, full text read (CHECKED)
8	Polarization grew with numeracy; ~45-point advantage for high-numeracy partisans on flattering data vs. ~25 points for low-numeracy	VERIFIED	Kahan et al., multivariate analysis section, same paper
9	Johnson & Seifert (1994) warehouse-fire paradigm: readers keep invoking retracted cause	VERIFIED	J Exp Psychol Learn Mem Cogn 20(6):1420; described in Lewandowsky et al. 2012
10	Corrections reduce but rarely eliminate reliance on misinformation; sometimes no effect	VERIFIED	Lewandowsky et al. 2012, PSPI 13(3); Ecker et al. 2022, Nat Rev Psychol 1:13–29 (CHECKED)
11	Illusory truth meta-analysis of 51 studies: d = 0.39 within-items, d = 0.50 between-items	VERIFIED	Dechêne et al. 2010, PSPR 14(2):238–257, figures reported in Psychon Bull Rev 2021 map (CHECKED)
12	Fazio et al. (2015): repetition raised belief even among subjects who knew the correct answer	VERIFIED	APA full-text PDF, J Exp Psychol Gen 144(5):993–1002 (CHECKED)
13	Pennycook & Rand (2019) tested 3,446 subjects; CRT predicted discernment regardless of ideology	VERIFIED	Cognition 188:39–50; PubMed record (CHECKED)
14	Gallup, Sept. 2025: 28% trust in mass media; 8% Republicans, 27% independents, 51% Democrats; 72% in 1976	VERIFIED	Gallup poll release, Sept. 2–16, 2025, n=1,000 (CHECKED)
15	Medill 2025: 136 newspapers lost in a year, ~3,500 since 2005, 213 news-desert counties, 1,524 single-source counties, ~50M people	VERIFIED	Medill Local News Initiative, State of Local News 2025 (CHECKED)
16	CPB announced wind-down Aug. 1, 2025 after $1.1B rescission; most staff ended Sept. 30, 2025; transition team through January 2026	VERIFIED	NPR, Aug. 1, 2025; CPB statement quoted therein (CHECKED)
17	NHPR president said CPB’s closure did not change the station’s course and named the unfunded shared services	VERIFIED	NHPR, Aug. 5, 2025 (CHECKED)
18	AI robocall reached thousands of NH households two days before the Jan. 23, 2024 primary; first known political deepfake of its kind	VERIFIED	NBC News investigation and indictment coverage (CHECKED)
19	Kramer charged with 26 counts by NH AG; FCC proposed a $6 million fine	VERIFIED	NBC News, May 2024 (CHECKED)
20	Belknap County jury acquitted Kramer on all counts in June 2025; Formella statement; fine unpaid per AP	VERIFIED	NHPR, June 13, 2025; Associated Press reporting (CHECKED)
21	Accuracy-prompt replication failed at first stage and produced a small effect after doubling the sample	VERIFIED	Roozenbeek, Freeman & van der Linden 2021, Psychol Sci 32(7):1169–1178 (CHECKED)
22	Five prebunking videos targeted technique, not topic, and improved resistance on social media	VERIFIED	Roozenbeek et al. 2022, Sci Adv 8:eabo6254; PubMed record (CHECKED)

Editor’s note. This piece expands and corrects an earlier Granite State Report essay on the same subject. The earlier version cited a single meta-analytic effect size of d = 0.53 for the illusory truth effect; the published meta-analysis reports d = 0.39 within items and d = 0.50 between them, and the figures here reflect that. The earlier version also placed the Corporation for Public Broadcasting’s closure in early 2026; the wind-down was announced August 1, 2025, with most staff positions ending September 30, 2025 and a transition team retained into January 2026. Every link above was opened and confirmed live on July 18, 2026. Four publishers block automated retrieval — SAGE, Elsevier, Science, and SSRN — so those papers are linked through PubMed or the publisher’s own free full text instead. Johnson and Seifert (1994) and Hasher, Goldstein and Toppino (1977) are cited in plain text because no stable public link resolved. The Union Leader item in row 4 is paywalled and is marked ATTRIBUTED. Corrections: Granite State Report corrects verified errors promptly and appends a note identifying what changed and when.