Authentic photographs, video, and audio of public figures are AI fabrications, so any inconvenient recording can be dismissed as fake

Why 'that's AI' is a reasonable first thought

The strongest version of this position deserves stating properly, because it is not stupid and it is not paranoid. It is mostly correct about the world.

Generative image and video tools have crossed the threshold where careful, motivated inspection by an ordinary person is no longer enough. Hands and text, the old giveaways, are largely fixed. Fabricated footage of real politicians saying invented things has circulated during live news events and drawn millions of views before anyone corrected it, and the correction never travels as far as the original. Anyone who adopted a posture of default credulity over the past three years has been made a fool of repeatedly.

There is also a real institutional problem underneath. Detection is uneven and partly proprietary. Watermarking schemes only cover content from the models that implement them, which means a fake made with a tool outside that set leaves no marker at all. Newsrooms often lack the forensic capacity to adjudicate an image on deadline, and the analysts who can are few and busy. Someone who says I no longer know how to tell is describing the situation accurately.

Given that, treating a startling image with suspicion is not a failure of reasoning. It is the appropriate reflex, and this site would rather readers had it than not. The question this file turns on is what happens next: whether suspicion is treated as a question to be resolved, or as a conclusion that has already been reached.

What happened when the claim was actually checked

The useful thing about this particular pattern is that it produces testable predictions, and several have now been tested in public. The results have not been kind to it.

In August 2024, photographs of a Kamala Harris rally at a Michigan airport hangar were declared AI-generated, with the assertion that the crowd “didn't exist” and that nobody had been there. The rally had been attended by thousands of people who took their own pictures, livestreamed by numerous news channels, covered by reporters on site, and attended by other prominent politicians. Local outlet MLive put the crowd at roughly fifteen thousand. Snopes ran AI-detection tools and found the image was likely photographed rather than generated. Hany Farid, a digital-forensics specialist at Berkeley, examined it with models built for exactly this purpose and reported no evidence of AI generation. The Harris campaign identified the staff member who took it.

Every independent line of inquiry converged, and the claim did not survive any of them. Note also how little forensics was needed: an event with fifteen thousand witnesses and dozens of camera crews is not the sort of thing that can be invented, whatever any detector says.

The McConnell sequence in July 2026is more instructive still, because the same tools returned opposite answers about two different images, correctly. A widely shared picture of the senator in a hospital bed was identified as machine-generated after Google's SynthID watermark was located in it. Days later, when his office put out a genuine photograph to answer the speculation, that image was accused of being AI as well, or of having been recycled from 2023. PolitiFact and Snopes found themselves debunking the debunking, and a chatbot circulating in those threads cited a fact-check that had never existed.

Hold those two results side by side. A fake was caught. A real photograph was cleared. The methods distinguished between them on the basis of a detectable property. A working toolkit behaves this way, and it is the opposite of the world the theory describes, in which nothing can be established and therefore anything may be dismissed.

What a watermark can and cannot tell you

One rebuttal recurs often enough to deserve its own answer: watermarks can be stripped or spoofed, so finding one supposedly proves nothing. This treats two situations as equivalent when they are not.

Stripping a watermark removes evidence that an image came out of a generative model. That is a real limitation, and it means the absence of a marker establishes very little. A fake made with a tool that never watermarked anything, or a fake that has been deliberately laundered, will look clean.

Adding one is a different proposition. There is no ordinary route by which a photograph taken on a camera acquires a generative model's embedded signature. So the presence of a SynthID marker is strong positive evidence about origin, even though its absence is weak evidence about anything. The argument that gets made in practice runs the asymmetry backwards: it cites the genuine weakness of the negative case in order to dismiss the strong positive one.

The other rebuttal, that fact-checkers calling one image fake and another real is a contradiction, misreads what discrimination means. A test that came back “fake” every time would be useless. So would one that came back “authentic” every time. Different answers about different inputs is the signature of a method that is actually measuring something.

Why the denial is so easy to reach for

This is not, for the most part, a story about people being foolish. The denial is easy to reach for because the incentives and the arithmetic both favor it.

Start with cost. Posting “that's AI” takes a few seconds and requires no evidence whatsoever. Answering it requires forensic tooling, someone qualified to operate it, corroborating witnesses, and a day of professional time. In any contest between a free accusation and an expensive rebuttal, the accusation wins on volume, permanently.

Then there is what the denial does for the person making it. If you have already concluded that a senator is incapacitated, a photograph of him sitting up with his wife is not new information. It is an obstacle. One word clears it, and the belief continues undisturbed. This is ordinary motivated reasoning, and the availability of AI as an explanation has simply made it cheaper than it has ever been.

The American Political Science Reviewstudy puts numbers on the payoff. Across five pre-registered experiments with more than fifteen thousand American adults, politicians who claimed damaging material was misinformation retained more support than those who did not, and the effect concentrated among respondents already inclined to back them. That measures the reward, not anyone's sincerity. But a reward that consistent means the claim should not be granted simply because it is phrased as caution.

Finally, the distrust travels. If you already doubt the outlet, you will doubt the fact-checker it cites, and the forensic analyst the fact-checker quotes. Verification stops functioning as reassurance and starts reading as the cover-up recruiting another participant. There is no image that can be produced to someone in that position, which is why demands for proof-of-life photographs are so rarely satisfied by proof-of-life photographs.