AI-Generated Images in Journalism: A Global Crisis of Trust

The Numbers Underneath the Problem

The 2025 Reuters Institute Digital News Report found that 58% of survey respondents were concerned about their ability to distinguish real content from fake in online news.² In 2018 the figure was 42%. That 16-point shift happened over seven years, with the steepest drop in the most recent two.

The 2025 Edelman Trust Barometer measured a record high: 70% of respondents worry that journalists purposely mislead them.³ These numbers compound. Once people doubt the messenger, every image the messenger publishes becomes suspect, regardless of whether any specific image is real or synthetic.

The Anatomy of an Incident

The Southport photograph followed a pattern familiar to anyone who tracks misinformation professionally. A convincing visual circulates in the first hours of a breaking story, before the authentic record catches up. People share it because it confirms something they already believe or fear. By the time verification completes, the image has already done its emotional work.

Ireland's 2025 presidential election produced a starker example. A deepfake video falsely showed the eventual winner withdrawing his candidacy, with fabricated footage of national broadcasters appearing to confirm the news. The Electoral Commission had to issue an emergency statement.⁴

Brazil tracked the same dynamic in data. Agência Lupa found that 16% of claims it fact-checked in 2025 involved AI-generated content, up from 7% the year before. The fastest-spreading single piece of AI-generated content reached 32 million TikTok views before debunking arrived.⁴

The distribution is not random. AI-generated fakes cluster around moments of high emotional salience: elections, riots, disasters. Exactly when verification is hardest and the cost of error is highest.

How Wire Services Are Responding

The organized press moved relatively quickly once the scale of the problem became clear. AFP tested C2PA provenance certification during the 2024 US elections, using a modified Nikon camera to embed a cryptographic signature in each photograph at the moment of capture.⁵ The signature travels with the file and breaks if the image is altered downstream. The Associated Press prohibits AI-generated images as publishable editorial content entirely. World Press Photo, the benchmark for photojournalism integrity, banned AI submissions from its 2024 contest and employs two independent forensic analysts to examine RAW files, EXIF data, and editing histories for every finalist.⁶

These measures work for what they cover. They do not cover screenshots, reposts, or footage from the first people on the scene with a phone, which is increasingly where breaking news visuals originate.

The Liar's Dividend

Legal scholars Danielle Citron and Robert Chesney coined the term "liar's dividend" in 2019 to describe a risk that was then theoretical: once deepfakes become common enough, real footage becomes deniable.⁷ The idea was that the mere existence of convincing fakes would give bad actors a rhetorical escape hatch from genuine evidence. That risk is no longer theoretical.

Research from the 2024 US election cycle showed that telling a person a photograph might be AI-generated reduced their trust in it, even when the photo was real. The knowledge that fakes exist is itself a tool. A politician caught in a genuine photograph can claim AI-generation and expect a portion of the audience to accept the denial. The 2025 Edelman data suggests that audience is substantial.³

The Reuters Institute found that this effect compounds with prior exposure: people who had previously encountered AI-generated misinformation showed greater skepticism toward real images afterward.² Familiarity breeds doubt, not immunity.

What Actually Helps

No single technical fix closes this gap. C2PA covers cameras used by organized press; it does not reach the broader information ecosystem where most people encounter news imagery. Detection tools catch some synthetic images but operate at real-world accuracy rates well below their benchmark numbers, as we documented in the arms-race article. Platform content labels are inconsistently applied and easily circumvented.

The data also points to a useful reframe. AI-generated imagery may be somewhat overstated as a direct threat relative to simpler manipulation. A 2025 study by the News Literacy Project found that cheap fakes — real photographs with misleading captions, out-of-context footage, selectively edited clips — appeared seven times more often than AI-generated content in election-related misinformation.⁴ The problem is real; it just operates inside a larger ecosystem of manipulation that predates generative AI entirely.

The more actionable takeaway is behavioral. The habit of checking where an image came from before sharing it addresses both problems at once. Reverse image search is free and takes under ten seconds. The 2025 Reuters Institute data shows a growing share of news consumers doing it by default.²

Journalism's credibility problem predates generative AI. AI makes it harder to solve and raises the cost of getting it wrong. The Southport image spread because people wanted it to be true. No detection algorithm fixes that.