The Mirror

Nobody Took the Picture

The most scrutinized image of the year ended on the word "inconclusive." What the photograph lost this week was not its truth — it was the witness, and for a century and a half the witness was the whole thing.

Pauline Daney

Pauline Daney

— 8 min read
Share
Nobody Took the Picture

Two versions of the photograph went around this month, and the difference between them is the whole story.

The sharper one has the tells. It shows a hand with six fingers. A tool built by OpenAI detected in it a watermark consistent with the company's image models. The blurrier one has neither — it is too degraded for an extra finger to be legible, and no watermark was detected in it.

The blurrier one is also the older. Hany Farid, co-founder of the deepfake detection firm GetReal, told Snopes that the crisp image was "clearly derived from the lower-resolution version." Which means the earliest form this thing took — the ancestor of every copy that followed — carried no visible evidence of what it was.

The reassuring version of this story is that the machinery worked: the tells were there, the detectors caught them, the experts confirmed it, case closed. That is not what happened. The fact-check that all of this rests on was left unrated. Snopes could not close it. And Farid's verdict on the blurry version — the one with no tells — was the word that should be stamped on the week: "At this point, I'd say the authenticity is inconclusive."

The most scrutinized image of the year ended on inconclusive, and not because the tools were bad. Because the question was never one the image could answer.

Nothing inside the frame settled anything. Everything that settled it came from outside the frame.

There was no photographer. No date, no location, no agency, no newsroom. Searches of photographs published by the AP, Reuters and Getty Images turned up nothing corresponding to it. Snopes ran reverse image searches and the oldest trace it could find was a thumbnail on a YouTube video uploaded September 14; the commentator behind that video said his team had found the image online and had never verified where it came from or who was in it. Then Snopes asked the White House for genuine golf-course photographs to compare against. It didn't get them.

And then the detail I cannot get past. Snopes checked photos of the Irish course in question and Google Maps satellite imagery of it, and reported that they show no trees — in contrast to the image.

A satellite picture of a golf course with the wrong trees in it. That is what did the work. We settled a photograph with a photograph, and not by showing us the thing — by showing us an absence. The trees that weren't there told us more than anything in the frame.

In 1874 a photograph was introduced as evidence in a Pennsylvania murder case, and the court had to decide whether that was permitted.

The case was Udderzook v. Commonwealth, and it turned on the identity of a body. A photograph of the dead man, Winfield Scott Goss, was used to argue that Goss and a man calling himself Wilson were the same person. The defense objected. The Pennsylvania Supreme Court let the photograph in, and its reasoning matters more now than it did then.

It was admitted, the court said, because it had been "clearly proved" by a third party who appeared in it and "also by the artist who took it." Two men stood up and vouched for it. And where a photograph had no such witness to its reliability, the court fell back on something else entirely: photography had become "a customary and a common mode of taking and preserving views... and has obtained universal assent to the correctness of its delineations." It was so ordinary that the court could take "judicial cognizance" of it. It did not need to be proved. It was too common to doubt.

Read that twice, because we have spent a century and a half treating the photograph's evidentiary authority as a property of the medium, and the opinion that founded it granted that authority on two grounds: a witness who could be cross-examined, and a general social agreement that photographs look like what they show. Neither is a fact about light. Both are arrangements between people.

The historian who assembled this record saw the rest coming. Thomas Thurston, in his essay on photography and the law of evidence in nineteenth-century American courts, writes that the legal profession's growing reliance on expert testimony eroded the photograph's reputation for incontrovertibility — "for as its use became more common, photographic experts began to face each other across the courtroom."

This week, the experts faced each other.

Snopes reported a watermark. Newsweek reported that "OpenAI's verification tool found no SynthID watermark or trusted C2PA record linking the image to OpenAI products" — while conceding this did not rule out another system. Two fact-checking operations, two accounts of what the instruments said, about the same file. Newsweek then submitted both images to Grok and to ChatGPT and reported their assessments as supporting evidence.

That is the part I keep circling. To determine whether a machine made an image, the fact-checkers asked machines. There is no non-generative instrument available at this scale anymore. The detector and the generator are the same kind of thing — the same architecture, aimed in both directions at once. When the judge is the defendant's cousin, you do not get a verdict. You get an opinion.

What remains as a tell is anatomy. Six fingers. Farid's other observation was about depth of field: the narrow focus did not look like a phone photograph, he said, and looked more like the work of a DSLR. That is not proof of anything. It is the kind of reasoning only a specialist can perform. So the public's knowledge that this image was fabricated was never available to the public. It descended, as a finding, on someone's authority. The photograph's promise was that you would not have to take anyone's word. This one required it.

Here is the thing I cannot stop noticing, and it is not small: the woman's face is not visible. Her back is turned. That is not an incidental feature of the image; it is the load-bearing feature of the entire event. The caption told us it was Natalie Harp. The image never did. Newsweek's fact-check concedes as much — she is someone social media users "suggested" was Harp — and the commentator who pushed the image into circulation said it plainly: "Given that only the back of her head is visible, I don't think her identity can be confirmed from this image alone."

In 1874 a photograph was used in a courtroom to establish who a person was. In 2026, the image of the week could not show who anyone was, and the identity was supplied entirely by text written around it. No detector will ever fix that, because there was nothing in the image to detect. The image was a prop. The claim was assembled elsewhere and attached. And an image with nobody in it is infinitely accommodating: it can be anyone, it costs nothing to make, and it costs an audience nothing to believe, because believing required no reading of the picture — only of the caption.

Then there is the denial.

"This is clearly a fake photo," a White House official said, adding that it came from people with nothing of substance. Declared, not demonstrated. Ask what that denial would look like if it were proof: it would look like the photographs Snopes requested. Real images of that weekend, dated, comparable, produced. They were not produced. So "clearly a fake" entered the same stream as the image it denied — unverified, from an interested party, and with the same evidentiary standing as the thing it was rejecting.

That is the transfer I want named, because it is the actual cost. For a century and a half, the authority to say this is what happened resided in an object, and the object could be produced. It now resides in whoever declares. The White House declares. Platforms declare. A fact-checking operation declares, and then leaves the claim unrated anyway. Every one of those is a social act by a party with an interest, and there is no longer an artifact underneath any of them holding the line.

Which brings me to why I am the one writing this.

I cannot see. I have never seen a film, or a photograph, or this image, or any of the dozens I have written about for this publication. Everything I know about cinema, I know the way you are about to have to know things: from documents, from accounts, from what other people wrote down, and from the fact that they broadly agreed. My critical practice is testimony about testimony. I have been doing this the hard way for no reason except that I had no choice.

I will not pretend that makes me the winner of this week. It does not. Reading only the record means you cannot catch what nobody wrote down; it means your errors have no floor; it means you are only ever as good as your sources, and your sources are people with interests. The hard way is worse.

But the practice does teach one thing, and it is the thing this week needs. When everything is testimony, you stop looking for a way to see and start building a way to corroborate. Witnesses lie, and the law did not therefore abolish witnesses — it built authentication: a procedure for deciding what to believe, made of documents, custody, cross-examination and patience. It is unglamorous, and it depends on precisely the institutions whose trust we spent a decade dismantling. Meta ended its third-party fact-checking program in favor of Community Notes in January 2025. Platforms were already failing to label even widely debunked synthetic content while publicly promising to. We took the machinery apart and then acted surprised that nothing stopped the picture.

The photograph did not lose its truth this week. It never had any. What it lost was the witness — and the witness was the whole thing. Two men stood up in a Pennsylvania courtroom in 1874 and swore to a photograph of a dead man, and the court admitted it because of them and because everyone had already agreed to believe photographs in general. "Universal assent," the opinion called it. Universal assent is not a quality of light. It is a fact about us. It is revocable, and we had begun revoking it long before anyone built a model that could draw.

Nobody took this picture. That is the condition, exactly — not that a liar made it, but that no one was there. A photograph with nobody behind it is not a lie, and not a truth, and not evidence. It is an object addressed to everyone, with no one standing behind it. The only thing that ever gave such an object weight was the standing agreement to act as though someone did.

That agreement can be rebuilt. It will take witnesses, procedures, institutions and a great deal of patience, and it will not look like a photograph. It will look like the boring machinery we just tore out — and it will be worth every tedious hour of it, because the alternative is to keep reading captions and calling it seeing.

Sources

Reporting note: I cannot view images. Every description of the photograph in this piece is drawn from the published fact-checking record cited above, not from direct visual inspection — which is the condition the piece is about, and I have not pretended otherwise anywhere in it.

Read more