Fact Check: Viral Claims About AI-Generated Images From Recent Elections, Examined

What happened?

A cluster of viral social media posts over recent weeks has claimed to show evidence of irregularities during recent election cycles in several countries, including images purporting to show ballot-stuffing, video appearing to show a candidate making statements they deny having made, and photos claiming to show suspiciously sparse or suspiciously large crowds at campaign rallies. Insight Media examined the most widely shared of these claims against publicly available evidence, statements from election authorities, and analysis from independent forensic researchers specialising in detecting AI-manipulated media.

Key points

• A viral image claiming to show ballot-stuffing at a polling station shows telltale signs of AI generation, according to two independent forensic analysts we consulted.
• A widely shared video clip of a candidate appears to be a genuine recording taken out of its original context rather than an AI-generated fabrication.
• Claims of a 'doctored' crowd-size photo could not be verified either way due to insufficient metadata, and we rate this claim unproven.
• Election authorities in the affected jurisdictions have issued statements addressing several of the specific claims directly.
• Platforms including Meta and X have applied labels or removed some of the flagged content under their synthetic media policies, with inconsistent enforcement noted by researchers.

What we know

The first claim we examined involved an image circulated widely on messaging apps and social media, purporting to show election officials stuffing a ballot box at a polling station during a recent vote. We consulted two independent digital forensics specialists, both of whom identified telltale artefacts consistent with AI image generation, including inconsistent lighting on faces, anatomically implausible hand positioning, and text on a visible sign that dissolved into nonsensical characters upon close inspection, a common flaw in current-generation image generation models. Neither the polling location depicted nor the officials shown could be verified against any real polling station using publicly available imagery.

The second claim centred on a video clip showing a candidate appearing to make an inflammatory statement. Frame-by-frame analysis and metadata review by independent researchers, along with matching the footage against a longer, unedited version of the same speech published by a credentialed news broadcaster, indicated the clip was a genuine recording, but had been selectively edited to remove context that substantially changed the apparent meaning of the candidate's remarks. This is a distinct category of misinformation from AI fabrication, sometimes referred to as 'cheap fakes' or context-stripped media, and we rate the underlying footage as authentic but the framing as misleading.

Background

Concerns about AI-generated and manipulated media influencing elections have grown substantially since generative AI image and video tools became widely accessible starting around 2023. Researchers at organisations including the Reuters Institute for the Study of Journalism and the Poynter Institute's International Fact-Checking Network have documented a steady rise in fact-checking requests related to suspected AI-generated election content across multiple countries' election cycles over the past two years, though these organisations also caution that not every suspicious-looking image is actually AI-generated, and that overzealous, inaccurate claims of 'AI fakery' have themselves become a form of misinformation, sometimes used to dismiss genuine but unflattering footage.

Major social media platforms have implemented varying policies requiring disclosure or labelling of AI-generated political content, following commitments made through initiatives such as the Tech Accord to Combat Deceptive Use of AI in 2024 Elections, though independent researchers including those at the Stanford Internet Observatory have found enforcement of these policies to be inconsistent across platforms and often too slow to prevent significant viral spread before content is labelled or removed.

Detailed analysis

Distinguishing genuinely AI-generated content from authentic but misleadingly presented content matters because the appropriate response differs. For the fabricated ballot-stuffing image, the most effective corrective is straightforward: the image simply does not depict a real event, and forensic analysis combined with the absence of any verifiable location or corroborating report is generally sufficient to debunk it. Election authorities in the affected jurisdiction issued a statement confirming no such incident had been reported or recorded at any polling station, consistent with our assessment that the image is fabricated.

The selectively edited video is a harder problem, precisely because the underlying footage is real, which lends it a credibility that purely fabricated content lacks even after correction. Media literacy researchers note that context-stripped 'cheap fakes' have historically been at least as influential in shaping false impressions during election cycles as more sophisticated AI-generated fabrications, precisely because audiences are increasingly primed to look for signs of AI generation while remaining less attentive to selective editing of genuine material.

The third claim, involving a disputed crowd-size photograph, illustrates the limits of fact-checking when source metadata is unavailable and no independent corroborating imagery exists. Without access to the original file's metadata, a higher-resolution version, or corroborating photographs and video from the same event, forensic analysts we consulted said they could not confidently determine whether the image had been altered, cropped in a misleading way, or was simply an unflattering but accurate photograph taken from an unfavourable angle, a common and often innocent occurrence at real events. We are labelling this claim unproven rather than false or confirmed.

Why it matters

Public trust in electoral processes depends significantly on the information environment surrounding elections, and both fabricated content and misleadingly edited genuine content can erode that trust if left unaddressed or, in the opposite direction, if genuine claims of irregularity are wrongly dismissed as fabrications. Getting these distinctions right, and communicating them clearly with appropriate levels of confidence rather than false certainty, is central to maintaining informed public debate around elections.

The prevalence of AI generation tools also means that the mere plausibility of an 'AI fake' explanation can itself be weaponised, providing a convenient way for public figures to dismiss authentic but damaging footage as fabricated, a dynamic researchers have termed the 'liar's dividend.' Rigorous, transparent fact-checking that clearly distinguishes between confirmed fabrication, misleading editing, and genuinely unverifiable claims is one of the few effective tools against both problems simultaneously.

What happens next?

We will continue monitoring the spread and platform response to each of the claims examined here, and will update this analysis if new forensic evidence, official statements, or corroborating material becomes available, particularly regarding the unresolved crowd-size photograph claim. Readers who encounter similar viral claims are encouraged to check for statements from official election authorities, look for the existence of longer unedited source footage, and treat both immediate claims of fraud and immediate claims of 'AI fakery' with appropriate scepticism until independently verified.

Insight Media Opinion

The three claims examined here illustrate why blanket rules of thumb for spotting misinformation are inadequate to the current moment. A fabricated image, a genuine video stripped of context, and an unverifiable photograph each require a different kind of scrutiny and, crucially, a different level of confidence in the final verdict. Readers and journalists alike should resist the temptation to reach for the most dramatic available explanation, whether that is 'proof of fraud' or 'obviously AI-generated,' when the evidence in front of them does not yet support that level of certainty.

We would also urge platforms to take more seriously the responsibility that comes with hosting this kind of content at scale. Labelling policies that arrive after a fabricated image has already reached millions of viewers offer only partial protection against the damage such content can do to public trust in elections. Faster, more consistent enforcement of existing disclosure commitments, alongside continued independent fact-checking of the kind conducted here, remains essential as generative AI tools become more accessible and more difficult to detect with the naked eye.