# Authentic photographs, video, and audio of public figures are AI fabrications, so any inconvenient recording can be dismissed as fake

**Verdict: Contradicted.** Two different things share one set of words here. Synthetic media is real, cheap, and hard to spot; no part of this page disputes that. What this file rates is the second move, the claim that a specific authentic photograph, clip, or recording is an AI fabrication. In every case examined below, that claim failed when tested. The Harris rally photo was taken by a staff member at an event thousands attended and livestreamed, and forensic analysis found no sign of AI generation. The McConnell hospital image was AI, caught by a watermark; the replacement photo his office released was not.

Category: Science, Space & Technology · Era: 2020s · First circulated: Named in legal scholarship in 2019; in wide political use from 2024, and a recurring feature of 2026 news cycles · Believed by: Not a fringe belief so much as a reflex now available to anyone: politicians and their supporters, litigants facing recorded evidence, and ordinary social-media users encountering a photograph they would rather not accept. Experimental work published in the American Political Science Review, drawing on five pre-registered studies with more than 15,000 US adults, found the tactic measurably helps politicians hold onto support, with the strongest effect among their own base.
URL: https://theconspiratory.com/theory/liars-dividend

## Summary
As deepfakes got good, a second problem arrived behind the first. Once the public knows convincing fakes exist, anyone caught on camera can simply say the camera lied. Legal scholars Bobby Chesney and Danielle Citron named this the liar's dividend in 2019, and made the uncomfortable observation that it grows in proportion to how well people are educated about deepfakes: every explainer about how convincing fakes have become is also an argument that nothing can be trusted. This case file rates the specific claim, made repeatedly and in public, that authentic images and recordings are AI fabrications. It holds that claim up against what happened when each one was checked.

## The claim
That photographs, video, and audio which appear to document public figures are routinely AI-generated fabrications, that the tools and analysts who declare such material authentic are unreliable or complicit, and that a recording can therefore be set aside as synthetic without producing any evidence that it is.

## Origin and timeline
- 2018-04: The problem enters mainstream awareness. Filmmaker Jordan Peele and BuzzFeed release a synthesized video of Barack Obama, made as a public warning about how convincing manipulated video had become. The lesson lands, and so does a side effect nobody had priced in: the public now knows that any video might be fabricated.
- 2019: Legal scholars Bobby Chesney and Danielle Citron publish 'Deep Fakes: A Looming Challenge for Privacy, Democracy, and National Security' in the California Law Review, and give the second-order problem a name. The liar's dividend is the payoff a dishonest actor collects from public awareness of deepfakes: the better educated the audience is about fakery, the easier it becomes to wave away something real.
- 2024-08-11: The tactic arrives at the top of American politics. Donald Trump posts that photographs of a Kamala Harris rally at a Michigan airport hangar were AI-generated and that the crowd 'didn't exist'. The rally had been attended by thousands, livestreamed by numerous news channels, and photographed by attendees and reporters; local outlet MLive estimated around 15,000 people present.
- 2024-08: The denial is checked and fails. Snopes runs AI-detection tools and concludes the image was likely photographed rather than generated. Hany Farid, a digital-forensics specialist at the University of California, Berkeley, analyzes it with models built to detect AI generation and reports no evidence of it. The Harris campaign says a staff member took the picture.
- 2024: Researchers Kaylyn Jackson Schiff, Daniel Schiff, and Natalia Bueno publish 'The Liar's Dividend: Can Politicians Claim Misinformation to Evade Accountability?' in the American Political Science Review, testing the idea across five pre-registered experiments with more than 15,000 US adults. Claiming that damaging material was misinformation helped politicians retain support, and worked best on people already inclined to back them.
- 2026-07-08: Detection scores a public win. A widely shared image of Senator Mitch McConnell in a hospital bed is identified as AI-generated after Google's SynthID watermark is found in it, the watermarking system's first high-profile catch in a live news story.
- 2026-07: The reflex inverts within days. When McConnell's office releases a genuine photograph to answer speculation about his condition, social-media users accuse that image of being AI too, or of being recycled from 2023. PolitiFact and Snopes end up debunking the debunking, and one chatbot circulating in the thread cites a fact-check that does not exist.
- 2026-07-27: The pattern settles into a routine. His office releases a further photograph dated 26 July alongside a statement from the Office of the Attending Physician and a specific account of his rehabilitation and missed votes. Some users question that image as well. Each element of the statement was checkable, and none of it was contradicted.

## The evidence, claim by claim
- Claim: Deepfakes are now so good that no photograph can be trusted, so treating one as fake is just reasonable caution.
  Evidence: Caution and conclusion are different acts. 'This might be fabricated' is a sound starting posture; 'this is fabricated' is a claim that can be tested, and in the documented cases it lost. The Harris rally photo survived forensic examination and was corroborated by thousands of attendees, multiple livestreams, and independent press photography. The McConnell hospital image failed, because a watermark was found in it. The tools discriminated between the two, which is the opposite of a world where nothing can be known.
- Claim: A watermark like SynthID can be stripped or forged, so finding one proves nothing.
  Evidence: The two directions are not symmetrical. Removing a watermark destroys evidence that an image was machine-generated; it does not let anyone stamp one onto a photograph taken with a camera. So absence of a watermark establishes very little, while presence of one weighs heavily toward the file having passed through a generative model that embeds it. Citing the weakness of the first case to dismiss the second inverts what the technology actually shows.
- Claim: Officials releasing a photo to prove someone is alive and well is exactly what a cover-up would do.
  Evidence: As stated, nothing could ever count against the theory. Once a claim is built so that every possible observation confirms it, it has stopped doing empirical work. Look at what came with the McConnell photographs: a specific date, a statement from a named medical office, a countable number of missed floor votes, and a named Kentucky event he was skipping. Those are all falsifiable details, published in advance of any scrutiny, and none of them turned out to be wrong.
- Claim: Fact-checkers called the first McConnell image fake and the second one real. They cannot have it both ways.
  Evidence: Reaching different conclusions about different images is what a working method looks like. A test that returned 'fake' for everything, or 'authentic' for everything, would tell you nothing at all. Snopes and PolitiFact split on the two pictures because the pictures differed in a detectable respect: one carried a generative watermark and the other did not.
- Claim: When a public figure says a recording of them is AI, they are being appropriately careful in an era of fakes.
  Evidence: Motive in any individual case is not something an outsider can read. What has been measured is the payoff. The American Political Science Review study found that claiming damaging material was misinformation produced a real gain in retained support across five pre-registered experiments, concentrated among respondents already sympathetic to the politician. An incentive that reliable does not prove insincerity in a given instance, but it does mean the claim cannot be taken at face value simply because it sounds cautious.

## Why people believe it
- The premise is true, which is what makes the conclusion feel small. Convincing synthetic media exists, costs almost nothing, and fools people daily. Moving from 'this could be fake' to 'this is fake' feels like a short step, when it is actually a different claim carrying a different burden.
- The accusation is free and the rebuttal is expensive. Typing 'that's AI' takes four seconds and no evidence. Answering it takes forensic analysis, watermark tooling, corroborating witnesses, and a day of somebody's professional time. That asymmetry alone guarantees the accusation will outnumber the refutation.
- It rescues a conclusion someone already holds. If you have decided a senator is incapacitated, a photograph of him sitting up is a problem that needs solving. One word solves it, and the belief survives intact.
- Public education about deepfakes feeds it directly. This was Chesney and Citron's most uncomfortable point: the dividend grows as awareness grows. Every well-meaning explainer about how good fakes have become doubles as an argument that no image can be believed.
- Distrust transfers from the institution to its verification. If you already doubt the newspaper and the fact-checking outfit, then their finding that an image is authentic is not reassurance. It reads as the cover-up extending one layer outward.

## Open questions
- Watermark detection only covers content made by models that embed a marker, and that coverage is expanding fast. On 19 May 2026 OpenAI became C2PA-conformant and began adding Google DeepMind's SynthID watermark to every image from ChatGPT, the API, and Codex, alongside C2PA provenance metadata, and previewed a public tool that reports whether either signal is present in an upload. The two layers are meant to cover each other's weaknesses, since metadata carries more detail but is stripped easily, while the watermark survives screenshots and resizing. None of that closes the underlying gap: a fake made with a model outside these programmes leaves nothing to find, and a missing signal still says nothing either way about authenticity.
- Nobody has a good answer for the ordinary case: an image with no watermark, no reliable metadata, and no witnesses. Provenance standards that would travel with a file from the camera onward have been proposed and partly implemented, but they are nowhere near universal, and unsigned content will not simply disappear.
- How much deepfake denial is sincere confusion and how much is strategic cannot be determined from outside. The experimental work measures what the claim earns, not what the claimant believes, and those are genuinely different questions.
- Whether courts will treat 'it might be a deepfake' as a routine challenge to authentic recordings is still being worked out, and the answer matters well beyond politics: it reaches any case that turns on a photograph, a video, or a voice.

## Latest developments
- 2026-08-31T06:45Z: The Nepal flood has produced the clearest case yet of this file's problem running in both directions at once, on one event, inside a single week. A landslide and flash flood struck the Nepal-China border on 26 August. Full Fact recorded more than 547 dead two days later; CBS News, cited by Snopes, put the toll at almost 600 with roughly 2,500 missing. CCTV footage of the moment the water destroyed the Gyirong Port border crossing spread across X, Facebook and Reddit, and commenters on one X post with more than 13 million views insisted it was AI-generated. It was not. Snopes rated the footage authentic on 28 August, the Associated Press independently deemed it genuine, and drone footage of the aftermath confirmed the structure had been pulverised. Snopes noted that readers were arriving at the site searching for the video precisely because they had been told it was fake. Running alongside that, more than twenty separate fabricated or recycled clips about the same flood were rated false within six days by six fact-checking outlets in four countries: Lead Stories, Snopes, Full Fact, Factly, Fact Crescendo and Boatos. Among them an AI-generated bridge collapse, before-and-after damage images made with OpenAI tools, a fighter jet supposedly causing the floods whose first clip had been online since May 2026 carrying an AI watermark, women staged as flood victims, 2023 footage from the Congo, a landslide in Japan, an excavator swept away in Uttarakhand, and a Buddha statue video from 2024. So on one story, in one week, the public was misled in both directions: the real record dismissed as synthetic by an audience of millions, and a score of synthetic or recycled records accepted as real. The dividend is usually described as a defence available to the guilty. Here it protected nobody and cost only accuracy about a disaster that killed hundreds, which suggests the effect is no longer confined to people with something to hide. It is becoming the default posture toward evidence of any kind. (source: https://www.snopes.com/fact-check/china-nepal-border-flooding/)
- 2026-08-25T21:45Z: The cleanest demonstration of this file's problem yet, and it arrived as a matched pair from one fact-checker on one day. On 25 August Lead Stories rated an image of a person in a Spider-Man costume punching a man at a Minneapolis protest as AI-generated. Roughly two hours later it rated a different image of a person in a Spider-Man costume winding up to punch a man at a Minneapolis protest as true. The second is a photograph taken on 22 August by the photojournalist Chris Juhn, one of a series he posted to his own Bluesky and Facebook accounts. The documented record behind it: a planned demonstration and counter-protest at Minneapolis City Hall turned violent, a Minneapolis Police Department bulletin of 22 August named eleven people arrested, and Newsweek reported the department confirmed that one of them was wearing a Spider-Man costume. Consider what that does to the heuristic almost everyone now uses. A costumed figure throwing punches at a protest reads as obviously generated, and on 25 August a reader applying that instinct would have been right once and wrong once, on images of the same event. Scepticism calibrated to how implausible a scene looks performs no better than credulity when reality is doing something that looks synthetic, and this file's whole subject is what happens to public trust when the true record can be waved away. What separated the two images was not appearance but provenance: a named photojournalist, a series rather than a single frame, his own accounts, a police bulletin and a police confirmation to a news outlet. That is the same test that works against counterfeit announcements and fabricated newspaper front pages. It is not the picture that settles it, it is the chain behind the picture. (source: https://leadstories.com/hoax-alert/2026/08/fact-check-real-photo-shows-spider-man-winding-up-to-punch-jake-lang-at-minneapolis-protest.html)
- 2026-08-02T01:00Z: The effect this file describes is now being reported as a measured condition rather than a series of incidents. NBC News, surveying the first week of January 2026, found deepfakes clustering around fast-moving news: the US operation in Venezuela drew AI-generated images, recycled footage and altered photos within hours, and after an Immigration and Customs Enforcement officer fatally shot a woman in her car, a probably AI-edited image of the scene circulated while others used AI to try to strip the mask from the officer who shot her. Jeff Hancock, who founded the Stanford Social Media Lab, put the mechanism plainly: AI will undermine what he calls the trust default, the ordinary human habit of believing a communication until given a reason not to. He also expects the familiar detection folklore to expire. In his words, judging an image or video by eye 'will essentially become impossible', and 'the old sort of AI literacy ideas of let us just look at the number of fingers and things like that are likely to go away'. Two details sharpen the picture. The platforms pay for engagement, which rewards recycling old media around breaking events, so the incentive runs toward confusion. And AI-generated material has already turned up as evidence in courtrooms, which is the liar's dividend arriving somewhere it costs people their liberty rather than their afternoon. (source: https://www.nbcnews.com/tech/tech-news/experts-warn-collapse-trust-online-ai-deepfakes-venezuela-rcna252472)

## Sources
- AI is intensifying a 'collapse' of trust online, experts say, NBC News (2026): https://www.nbcnews.com/tech/tech-news/experts-warn-collapse-trust-online-ai-deepfakes-venezuela-rcna252472
- Deep Fakes: A Looming Challenge for Privacy, Democracy, and National Security, California Law Review (2019): https://www.californialawreview.org/print/deep-fakes-a-looming-challenge-for-privacy-democracy-and-national-security
- The Liar's Dividend: Can Politicians Claim Misinformation to Evade Accountability?, American Political Science Review (2024): https://www.cambridge.org/core/journals/american-political-science-review/article/liars-dividend-can-politicians-claim-misinformation-to-evade-accountability/687FEE54DBD7ED0C96D72B26606AA073
- Deepfakes, Elections, and Shrinking the Liar's Dividend, Brennan Center for Justice: https://www.brennancenter.org/our-work/research-reports/deepfakes-elections-and-shrinking-liars-dividend
- Trump falsely claims Harris used AI to fake rally crowd, CBS News (2024): https://www.cbsnews.com/news/trump-harris-campaign-photo-crowd-size-detroit/
- Hany Farid finds no sign of AI generation in photo of Harris Michigan rally, UC Berkeley School of Information (2024): https://www.ischool.berkeley.edu/news/2024/hany-farid-finds-no-sign-ai-generation-photo-harris-michigan-rally-newsweek-article
- Google's deepfake detector system used to debunk McConnell hoax pic, TechCrunch (2026): https://techcrunch.com/2026/07/08/googles-deepfake-detector-system-used-to-debunk-mcconnell-hoax-pic/
- Mitch McConnell not yet medically cleared for Senate return, new statement says, Forbes (2026): https://www.forbes.com/sites/saradorn/2026/07/27/mitch-mcconnell-not-yet-medically-cleared-for-senate-return-new-statement-says/
- Liar's dividend, Wikipedia: https://en.wikipedia.org/wiki/Liar's_dividend
- Advancing content provenance for a safer, more transparent AI ecosystem, OpenAI (2026): https://openai.com/index/advancing-content-provenance/
- C2PA and SynthID in OpenAI-generated images, OpenAI Help Center (2026): https://help.openai.com/en/articles/8912793-c2pa-and-synthid-in-openai-generated-images

Rated by The Conspiratory, a neutral, sourced encyclopedia of conspiracy theories. Full page: https://theconspiratory.com/theory/liars-dividend