AI-generated images spreading as real photos
AI-generated images being mistaken for or deliberately passed off as authentic photographs have become a documented and growing source of viral misinformation, particularly around breaking news, natural disasters, and political events.
What we know
The rapid improvement of image generation tools such as Midjourney, DALL-E, and Stable Diffusion since 2022 has made it possible to produce photorealistic images in seconds that can be difficult to distinguish from genuine photographs at a glance, particularly when viewed quickly on a phone screen at reduced resolution within a social media feed. Multiple high-profile examples have been documented and traced by journalists and fact-checkers, including fabricated images depicting an explosion near the Pentagon in May 2023, which briefly circulated widely enough to cause a small, temporary dip in stock market indices before being confirmed as fake by the Arlington Fire Department and reported by Reuters and other outlets, and numerous fabricated images purporting to show celebrities, political figures, or disaster scenes that were later identified as AI-generated by reverse image searches and forensic analysis.
Researchers who study visual misinformation, including teams at the Reuters Institute for the Study of Journalism and various university digital forensics labs, have documented recurring technical tells in AI-generated images, though these tells are shrinking in reliability as the technology improves. Historically common indicators have included anatomically implausible hands with incorrect numbers of fingers, inconsistent or nonsensical text within an image, unnatural repeating patterns in backgrounds, and lighting or shadow inconsistencies across different parts of an image, though newer generation models have substantially reduced these specific errors compared to versions from 2022 and 2023, meaning visual inspection alone is becoming a less reliable detection method over time.
Fact-checking organizations and technology companies have responded with a mix of detection tools and labeling policies. Reverse image search tools, forensic metadata analysis, and specialized AI-detection classifiers are used by outlets including AFP Fact Check and Reuters Fact Check to assess suspect images, though these tools carry acknowledged error rates and are not fully reliable on their own, particularly for images that have been recompressed or edited after generation. Major platforms including Meta, YouTube, and TikTok have introduced policies requiring disclosure labels on AI-generated or heavily AI-edited content, and a coalition of technology and media companies developed the C2PA (Coalition for Content Provenance and Authenticity) standard, an embedded metadata system intended to let viewers verify an image's editing history and origin, though adoption remains uneven across platforms and can be stripped through simple screenshotting.
Academic and journalistic research on the spread of these images consistently finds that emotionally striking or politically charged images spread fastest and farthest before corrections catch up, a pattern well documented in misinformation research generally and not unique to AI-generated content, but amplified by AI tools' ability to produce a large volume of plausible-looking images quickly and cheaply compared to earlier photo manipulation techniques that required more manual skill. Fact-checkers consistently recommend checking whether an image has been reported by multiple independent, established news organizations with named reporters before treating a viral image as confirmed, since single-source viral images, particularly those without any identifiable original photographer or wire service credit, carry a substantially higher risk of being fabricated, regardless of how convincing they appear.
Common claims
- A viral image showing a Pentagon explosion in 2023 was confirmed fake.True, officials and journalists confirmed the image was AI-generated and no explosion occurred.
- AI-generated images can always be identified by extra fingers or garbled text.Increasingly false, newer generation models have substantially reduced these specific visual errors.
- Reverse image search reliably identifies all AI-generated images.Partly true, it is a useful tool but carries acknowledged error rates and limits, especially on edited images.
- Major platforms now require labels on AI-generated content.True, though enforcement and adoption remain uneven across different platforms.