Answer in brief
Check the provenance chain, not the “fingers”: first publication, independent material from the event, the source's reaction and, where available, C2PA/SynthID. A missing credential proves nothing.
Five Documented Cases, Not Finger-Counting Advice
Five cases documented from primary sources (two AI images with publicly confirmed origins, one series in which contemporaries mistakenly attributed visual “AI clues” to the wrong image, one manual Photoshop edit of a genuine photograph, and — separately — a genuine photograph that was publicly and mistakenly declared an AI generation), data current as of September 2026 on what Google and Apple can verify, and a repeatable provenance-verification chain that applies to any disputed photo — instead of advice like “look at the fingers.”
Trace the Publication Path, Not the Pixels
When a disputed photo of a well-known person appears in a feed, one of the obvious pieces of advice is to “look at the hands” or “count the teeth.” That method is becoming less reliable: generative models change from version to version, and a harmless inconsistency in an ordinary, unedited shot — uneven lighting, an odd angle, social-media compression — can look just as suspicious as a sign of generation. The five cases documented from 2023–2024 below point to a more durable approach: who first published the image and when, what the source itself, independent witnesses and specialized institutions — photo agencies, award organizers, digital-forensics experts — say about it, and which technical provenance marks, such as C2PA and SynthID, or institutional reactions from agencies and organizers, confirm or disprove the claim of a fake. Among the cases is one episode where two different images from the same day were mixed up, and another where suspicion fell on a genuine photograph rather than a generated one.
Pope Francis in a Puffer Coat: Exposure Came Before the Creator's Admission
In March 2023, a photo of Pope Francis in a large white puffer coat — stylish, modern, and nothing like traditional papal vestments — spread across social media. The image went viral: BuzzFeed News later called it the picture that “fooled the world.” A telling detail for verification method: even before the creator's name became known, the theory that the photo was AI-generated had already spread online — suspicion took shape earlier, and without any input from the creator. On March 27, 2023, BuzzFeed News was the first to publish the creator's story: Pablo Xavier, a 31-year-old construction worker from the Chicago area, who generated the first three versions of the image in Midjourney around 2 p.m. on a Friday and posted them to the Facebook group AI Art Universe, and then to Reddit. Asked about his motive, he said: “I just thought it was funny to see the Pope in a funny jacket” (editorial translation; BuzzFeed News, March 27, 2023). TV host Chrissy Teigen publicly wrote that she, too, had mistaken the photo for real. The case has two independent links of confirmation: an earlier collective recognition online, and, separate from it, the creator's own voluntary account of who he was and why he did it.
The Trump “Arrest” Series: How Two Different AI Images Got Mixed Up
In that same March 2023, Bellingcat founder Eliot Higgins used an AI generator to create a series of images from the prompt “Donald Trump falling down while being arrested” (editorial translation; per The Independent, March 24, 2023). When he posted the series, Higgins captioned it: “Making pictures of Trump getting arrested while waiting for Trump's arrest” (editorial translation). The images racked up nearly five million views. During those same days, Trump himself shared a different AI image — one showing him kneeling in prayer. It was in that separate image, not in the “arrest” series, that visual signs of AI generation were found, according to The Independent citing Forbes: a missing finger on the right hand, “mixed-up” thumbs, and a shoe sole with “brush-like” edges. The distinction matters for the method: a visual clue has to be attached to the exact image it came from — in a stream of viral AI pictures from the same day, it is easy to mix up which clue belongs to which picture, and a conclusion about “signs of fakery” in one image does not carry over to another image from the same day.
Eldagsen and the Sony World Photography Awards: The Dispute Was About Trust, Not Concealment
On March 14, 2023, Berlin-based artist Boris Eldagsen was announced the winner of the Creative Open category at the Sony World Photography Awards with a black-and-white image titled “Pseudomnesia: The Electrician,” depicting two fictional women from different generations. According to the organizers, the World Photography Organisation, Eldagsen confirmed in correspondence before the winner was announced that the image was “co-created” using AI. In other words, the organizers already knew about the AI's involvement before the winner was announced — the dispute was not about concealing that fact but about the good faith of the artist's assurances: the organizers later stated that he had made “deliberate” attempts to mislead them. Eldagsen himself describes the situation differently: according to him, he informed the organizers about the AI's involvement in advance. In April 2023, he announced he was declining the award, publishing a statement on his website: “AI is not photography. Therefore I will not accept the award” (editorial translation; per CNN, April 18, 2023). He explained that he had entered “as a cheeky monkey” — to test whether photography competitions were ready for AI images — and got his answer: they were not. The case is instructive not as an example of “an artist deceiving a jury” but as an example that even when AI involvement is known in advance, a dispute over trust can still continue, and the decisive public gesture turns out to be the creator's own statement rather than technical expertise.
The Kate Middleton Photo: Ordinary Editing, Not a Neural Network
On March 10, 2024, Mother's Day in the United Kingdom, Kensington Palace published a photograph of the Princess of Wales with her children, taken by Prince William — the first picture of Catherine since her operation in January. Late Sunday evening, the Associated Press withdrew the photo from its archive, citing a technical reason: “inconsistency in alignment of Princess Charlotte's left hand” (editorial translation; AP, per BBC, March 11, 2024). Following AP, Reuters and AFP withdrew the photo too, then Getty Images; the next day, Monday, the PA news agency followed the same path, citing the palace's lack of explanation. All five agencies used the term “manipulated.” A BBC analysis pointed to specific inconsistencies: in a sleeve, in the blurring of a hand, in a zipper, in the background and on one of the children's knees. The next day, Catherine personally apologized on social media through the official account, signing off with the initial “C”: “...express my apologies for any confusion the family photograph we shared yesterday caused” (editorial translation; March 11, 2024), adding that, like many amateur photographers, she sometimes experiments with editing. BBC states plainly that, as of March 11, 2024, basic details about the photo — when it was taken, what exactly was changed and whether it was a composite — remained undisclosed even after the admission of editing; the editors did not verify any more recent information. The case shows that an official admission of editing does not settle every question about a photo, and that a “neural network” is not the only way to alter a documentary image.
The Reverse Case: A Genuine Harris Photo in Detroit That Was Called an AI Generation
The opposite situation happens too — when a genuine photograph is mistakenly declared to be AI-generated. On August 7, 2024, at a rally near Detroit Metropolitan Airport, U.S. Vice President Kamala Harris was greeted by a crowd of supporters; local outlets Detroit News and Detroit Free Press independently reported thousands in attendance. On August 11, Donald Trump wrote on Truth Social: “There was nobody at the plane, and she 'A.I.'d' it” (editorial translation), and later added that he had “caught” Harris with a “fake crowd.” Supporters of this theory pointed to reflections on the fuselage of Air Force Two as proof that there had been no crowd. Fact-checkers' verification relied on dozens of photos and videos from the event, including images from Getty Images and the Associated Press, which independently confirmed the supporters' presence. Hany Farid, a professor at the University of California, Berkeley's School of Information and a digital-forensics specialist, examined the photo with two different computer models trained to recognize generative-AI patterns — both found no evidence of either generation or compositing. Comparing several versions of the photo, Farid concluded that the only edit was a simple adjustment of brightness and contrast and possibly sharpness — the photo was neither AI-generated nor entirely untouched. He separately commented on a glowing outline around people's heads and silhouettes: according to him, the effect was caused by ordinary hangar lighting, and the same scene is visible in many other photos and videos from the same event. The Harris campaign confirmed that the photo was taken by a campaign staffer and that it was neither created nor processed with AI. This case is worth remembering whenever a “strange halo” in a genuine photo from a poorly lit room looks like a sign of generation — and arguments against a photo's authenticity are worth checking against what was actually said rather than against what merely looks convincing at first glance.
What Watermarks and the C2PA Standard Can Do — and Where They Fall Short
Some content-provenance technologies are technical in nature rather than journalistic. The open C2PA standard (Coalition for Content Provenance and Authenticity) records the creation and editing history of a digital file — “a kind of nutrition label for content.” The Content Authenticity Initiative, founded by Adobe in 2019 and still led by the company, develops open tools for building this standard into websites, apps and services. Google DeepMind applies a similar principle in its SynthID technology: an imperceptible watermark is embedded in the images, audio, video and text produced by Google's generative products and is designed to withstand cropping and compression. By September 2026, the limitation that “SynthID only works with Google content” was no longer entirely true: according to a Google blog post from May 19, 2026, OpenAI, South Korea's Kakao and ElevenLabs were adopting the technology for their own AI content, while NVIDIA was using it for video generated by its Cosmos models. On September 9, 2026, Apple announced that it would add SynthID support in a software update later in 2026 — the marks will appear on most edited images. At the same time, Apple introduced “Apple Reference Image” for the iPhone 18 Pro: the camera sensor signs pixel-level data at the moment of capture, and that signed image can be compared against the photo in the Photos app. The principle behind all these technologies is the same: they can confirm provenance only where a credential was embedded in the first place. If content was made with a tool that lacks such a credential, or passed through a service that strips it out, a C2PA or SynthID check will show nothing — and that does not prove the content is fake.
Why Missing Metadata Proves Nothing
A common argument in disputes over a photo's authenticity is: “the file has no EXIF data, so it must be fake.” That logic is unreliable: many social networks and messaging apps strip out some file metadata — camera model, shooting date, location, exposure — during upload and recompression, regardless of whether the picture was taken with a camera or generated by a neural network. So missing metadata in a file that has passed through a popular platform proves nothing by itself: it can simply be a result of ordinary transmission rather than a sign of the file's origin. The situation is changing: according to Google, starting in 2026 Meta will begin labeling Instagram photos taken with a camera that supports Content Credentials — for some files, provenance data will instead be preserved and made visible. Judging a photo's authenticity solely by the presence or absence of EXIF data, without considering exactly where the file came from, is a common methodological mistake.
Reverse Image Search: A Reader's First Practical Step
Before analyzing pixels or waiting for an official comment, readers have a free tool available — reverse image search. Google Lens (“Search by image”) lets you upload a file, drag it into the search box or paste a link to the picture, and get back a list of sites where the same or a similar image appears. The tool does not promise a strict sort by date of first publication, but it does give a starting point — a list of sources where the image has already appeared. Since May 2026, readers have had a more direct option as well: according to the Google blog, Lens, AI Mode and Circle to Search in Search, along with Gemini in Chrome, can answer the question “Is this made with AI?” by checking for a SynthID watermark and cross-referencing C2PA Content Credentials — whether the frame was captured by a camera unedited or had been altered. The limitation is the same as for the credentials themselves: a meaningful answer is available only for content where the mark was embedded from the start; for every other image, the tool will not say anything definitive.
How to Build a Verification Chain for a Specific Photo
The five cases above come down to a single route that can be repeated for any disputed image — it is laid out in the checklist below. Three things are worth keeping in mind separately from the checklist. First, no single clue — an extra finger, missing metadata, the verdict of one AI detector — is final proof on its own: in the Harris photo case, the expert checked the picture with two different models, not one. Second, an official admission does not settle every question: the Kate Middleton example shows that some details can remain undisclosed even after an admission of editing. Third, a visual clue has to be checked against the specific image in question: in the Trump episode, the AI signs journalists found in 2023 belonged to a different picture of him, not to the “arrest” series — confusion between similar images from the same day happens more often than it seems, and virality says nothing about a photo's origin.
Practical checklist
- Find earlier publications of the image through reverse image search (Google Lens/“Search by image”) — this does not guarantee the very first version, but it gives you a list of sources for comparison.
- Check what was written about the image in its earliest publications: who the creator was, the circumstances in which the picture appeared, and whether an AI-origin label was present from the start.
- Do not treat missing EXIF metadata as proof of a fake — many platforms strip it out during ordinary uploads regardless of the file's origin.
- Look for independent confirmation from the scene — other photographers, video from different cameras, witnesses, local media.
- If the file carries a provenance credential (C2PA Content Credentials, a watermark such as SynthID), check it with available tools (Gemini, Lens, AI Mode) — but remember that a missing credential proves nothing if it was never embedded in the first place.
- Do not rely on a single clue — a detail in the image or the verdict of one third-party detector — as final proof, and make sure the clue actually belongs to this image and not to a similar picture from the same day.
- Wait for an official reaction from the source — a press office, the creator, or a relevant institution — and keep in mind that even this may not close every technical question about the photo.
Questions and answers
Is It True That a File Missing Camera and Date Data Signals a Fake?
Not by itself. A photo's metadata is often stripped during an ordinary upload to a social network, messenger or compression service — this happens to genuine photographs and AI images alike. At the same time, some platforms are moving in the opposite direction: for example, according to Google, Meta will begin labeling Instagram photos from cameras that support Content Credentials. What matters is not simply whether metadata is present or absent, but whether there is independent confirmation of the image from the moment of its first publication.
Can You Rely on Free Online AI-Image Detectors?
This piece did not run a documented test of specific commercial detector services, so the editors cannot assess their accuracy. A clear principle emerges from the cases examined: even professional expertise, as in the Harris photo case, relied not on a single tool but on two different detector models together with independent outside confirmation — the same logic is worth following when using public detectors.
What Should You Do If You Cannot Verify the Image's Original Source?
Treat the image as unconfirmed and do not spread it further as documentary fact until at least one independent confirmation turns up — a first-hand publication, a reaction from a relevant institution, or agreement with other independent footage of the same event.
Does the Same Method Work for Video, Not Just Photos?
Partly: searching for the first publication and checking witness reactions work the same way, but a full verification of video requires additional steps — comparing frames, timecodes and footage from different sources — that go beyond this piece.