Product1 publisher3 min readPublished
A fabricated 9/11 rescue photo cleared five AI checkers on the first ask
Gizmodo asked five AI systems whether a viral image of the 9/11 rescuer Welles Crowther was real. Each described the fabrication as a photograph of some kind, and what caught it was a reverse image search.
The Product Desk · Product desk
What happened
- A fabricated image purporting to show Welles Crowther, the 9/11 rescuer known as the Man in the Red Bandana, saving people at the World Trade Center is spreading online, Gizmodo reported.
- Crowther, a 24-year-old equities trader and volunteer firefighter, is credited with helping save at least a dozen people before the South Tower collapsed, and no photographs of him in action are known to exist.
- Grok told Gizmodo the picture was a well-known photograph from the September 11, 2001 attacks, and conceded it might be fake only after being pressed on whether it was authentic.
- Gemini said it could not check the file for Google's SynthID watermark, a detection feature Google opened to all Gemini users last year, then reported finding no SynthID once pressed.
- Google's AI overview placed the image in Aftershock: Earthquake in New York, a disaster television miniseries from 1999, two years before the attacks the picture claims to show.
Compiled by The Product DeskSomething wrong?How this is made
Why it matters
- constraint Generation-time provenance only covers the vendor that stamps it. A library checking for SynthID finds out one thing, whether Google made the file; anything scanned, screenshotted or produced elsewhere sits outside that check.
- exposure Any product that invites a user to ask an assistant whether an image is real is now passing along confident wrong answers, and the team that shipped that affordance answers for the correction.
- decision Anyone cataloguing submitted media has to pick between a verdict they cannot defend and a trail they can: first-seen date, earliest known publication, and whether anything predates this week.
- precedent More fabricated Crowther images have already appeared this year, so curators of any commemorated event with a thin photographic record should expect synthetic pictures to fill the gap.
Somebody with a suspicious picture and a deadline does the obvious thing: pastes it into a chat box and asks whether it is real. Gizmodo did that with the Crowther image on five surfaces, and none of them called it AI-generated in a first answer [1]. ChatGPT was the most confident of them. "It very likely is not AI-generated. It appears to be a genuine historical photograph rather than an AI-created image," it replied [7], adding that the picture was "consistent with photographs taken during the evacuation of the World Trade Center after the September 11 attacks" [8].
Claude did something more useful to a person spreading the image. Asked to run a reverse image search, it said it had "confirmed" the picture came from a documentary, told Gizmodo "this doesn't look AI-generated" and called it a "reenactment/dramatization photo" [10]. The documentary it named is real. Man In Red Bandana, from 2017, briefly shows an actor recreating Crowther's rescue, and that actor wears a white t-shirt, while the viral image shows a button-up shirt and tie [11][2].
Even a clean answer on the watermark path would have settled little. SynthID is added to all images created with Google's AI tools [13], so a file with no SynthID has been cleared of being made by Google and by nothing else [3]. Gizmodo says AI image detectors can be unreliable [18]; the sharper problem for a catalogue is that a negative watermark result will be treated as a clearance.
One check landed. Reverse image searches indicate the picture had not been published anywhere before this week, and Gizmodo states plainly that the image is not genuine and was made with AI [15]. Publication history is the kind of fact an archive already stores.
Demand for pictures of Crowther is not going to fall. Donald Trump awarded him the Presidential Medal of Freedom on Tuesday [5] and, in remarks Gizmodo quoted from a post dated September 8, 2026, said: "The Man in the Red Bandana. Have you heard of him? He's become more famous than me. I don't like it. Very unhappy. But he's become much more famous than me." [6] Roughly 25% of Americans alive in 2026 were not born by September 11, 2001 [17].
The Gizmodo test shows the real behaviour: a file goes into a chat window, and the answer that comes back never gets a second opinion. Two questions come first. Do you know where an item came from at the moment it entered the collection, meaning a device upload you can inspect or a generator you operate. And what does a wrong label cost on this specific item, a bad nudge on a meme or a caption on a memorial page. Items with unknown origin and high cost should not get a verdict at all. What they get instead is the trail: the earliest date you hold for the file, where it was first indexed, whether any publication predates that. The tradeoff is not small. Newly digitised archival material has no online history either, so it will look as suspicious as a fake and needs a human review path, and your first-seen date is only as good as your crawl. In this test, five AI systems missed the image and publication history caught it [1][15].
What to watch
- Whether Google states which generators the Gemini SynthID check actually covers, and whether it reads marks from non-Google tools.
- Whether the fabricated Crowther image gets picked up as illustration by news outlets, wikis or classroom material.
- Whether OpenAI, Anthropic or xAI change how their assistants answer "is this AI?" on historical images they cannot verify.