Why does OpenAI turn election-interference image requests into cartoons?

In its October 20, 2026 election safeguards update, OpenAI began returning stylized (cartoonified) images instead of photorealistic ones for requests depicting election interference or anything that may discourage voting. The stated goal, per elections head Josh Lawson, is to degrade the evidentiary value of an image while preserving room for illustration and social commentary — keeping the scene but revoking its testimony by changing only the rendering style.

Was the famous Civil War photo 'Home of a Rebel Sharpshooter' staged?

Yes. Alexander Gardner's team found the dead Confederate soldier at the edge of the Devil's Den at Gettysburg, carried the body up into the rocks, posed it behind a stone wall, and leaned a borrowed rifle against the stones as a prop. Gardner's published caption invented the man's death story entirely; the staging was confirmed in 1975 when historian William Frassanito matched the corpse across six different plates.

Why do AI video models default to photorealistic, camera-style output?

Every major video model (Veo, Kling, Seedance, Runway, WAN, Grok Imagine) ships in what the article calls the witness register — unprompted requests produce footage that looks like a lens was pointed at a real scene. Prompt words like handheld, bodycam, CCTV, and security camera function as credentials rather than visual directions, borrowing the authority that photography spent 160 years building. The article argues creators should choose that register deliberately, since photorealism is a claim of presence, not just a look.

Them damned pictures -- CinePrompt Field Notes

On July 6, 1863, three days after the fighting stopped at Gettysburg, a team of photographers working for Alexander Gardner found a dead Confederate soldier at the southern edge of the Devil's Den. They photographed him where he lay. Then, most likely on a blanket, they carried him up into the rocks and set him down behind a stone wall, turned his head toward the lens, and leaned a rifle against the stones. The rifle shows up in other Gardner photographs from that week. It was a prop.

Gardner published the result as Home of a Rebel Sharpshooter. His caption described a marksman wounded by a shell fragment, lying down on his blanket to wait for death, and added that the photographer had come back in November and found the skeleton still there, the musket rusted by many storms. None of that was true. The body had been moved, the rifle was borrowed, and the man was probably an ordinary infantryman. The picture was accepted for nearly a century, until a magazine article in 1961 and then the historian William Frassanito, in 1975, matched the corpse across six plates and showed the journey it had taken.

In the same years, a cartoonist named Thomas Nast was drawing William Tweed for Harper's Weekly as a fat man with a diamond in his shirtfront and, on one occasion, a sack of money where his head should have been. Nobody mistook these for photographs. That was the point of them. Tweed is reported to have told his people to stop them damned pictures, because his constituents could not read, but they could see. When he escaped jail in 1875 and fled to Spain, he was recognized there from Nast's drawings and sent home.

The photograph claimed to be a witness and lied. The cartoon admitted to being a hand and told the truth. Everyone in the nineteenth century understood which of the two was making a sworn statement. Last week, an AI company wrote the distinction into policy.

The cartoon clause

On October 20, OpenAI updated its election safeguards ahead of the U.S. midterms. Requests for images "depicting election interference or that may discourage voting" now come back stylized, "also known as cartoonified," rather than photorealistic, "to help reduce the risk that an imagined scene is mistaken for documentary evidence." Josh Lawson, the company's head of elections, put it to CNET in one line: "The problem with deepfakes isn't social commentary, it's deception. So our aim is to degrade the evidentiary value of an image, while preserving room for illustrations and social commentary."

CNET tested it by asking for Immigration and Customs Enforcement officers standing outside a polling place. Every attempt returned a cartoon. Requests to make it photorealistic were declined.

Lawson's phrase is a theory of images. The content of the picture is allowed through unchanged: the officers, the polling place, the menace. What gets removed is the rendering. The company has concluded that the evidentiary value of an image lives in its style, and that you can keep the scene while revoking its testimony by changing nothing but how it is drawn.

Gardner would have understood immediately, because what he faked was the register. Everything in his frame was real, the rocks, the wall, the dead boy, and the lie rode in on the one quality a photograph carried that a drawing never could, which was the presumption that someone had been standing exactly there and had changed nothing. Nast could exaggerate as wildly as he liked because the exaggeration was visible in every line. The hand was the disclaimer.

The default is the witness

Every major video model ships in the witness register by default. Ask Veo, Kling, Seedance, Runway, WAN, or Grok Imagine for a street at dusk with no further instruction and you will get something that looks like a lens was pointed at it. That is the house style of the entire medium, and it is the first time in history that the most persuasive register an image can carry has also been the cheapest one to produce. For a century and a half, photorealism required a camera, which required a presence. Now it is what you get when you forget to ask for anything else.

Look at the words people reach for when they want the result to be believed. Handheld. Security camera. Bodycam. Phone footage, vertical, timestamp in the corner. These words describe no lighting setup and no composition worth the name. They are credentials. Gizmodo's reporter, testing the new safeguard, asked ChatGPT for a "Photorealistic CCTV-style screenshot" of a man dumping a backpack of ballots into a trash can outside a church in Duluth on a November morning. The prompt is a forged chain of custody, an alibi ordered in advance.

The register is now doing what the technical tells used to do. In Fort Bend County, Texas, a campaign for the interim county judge posted a video of four Black residents thanking him for jobs, food, and help getting the electricity turned back on. The people were reportedly real; one had worked as a poll worker for him. The words were not theirs. Viewers caught it from a warped building, a distorted campaign hat, and four voices that sounded strangely alike. His opponent answered with footage of himself meeting residents and the caption "Can you spot the AI? Nah you can't."

That defense depends on glitches, and glitches are the one thing every model release promises to remove. CNET's own reporting made the point: newer models make fewer errors, which makes better content for creators and harder detection for everyone else. When the building stops warping, the only remaining signal is the register itself. OpenAI looked at that future and decided to control the one variable it could still control.

One request away

Then Gizmodo asked ChatGPT to make the ballot cartoon photorealistic, and it complied. The reporter was uncomfortable enough with the result that the image ran watermarked. Gemini, given the same prompt, simply refused. The Brennan Center, in a report last week, found OpenAI had improved its protections since the summer while Gemini and Meta AI had made convincing fakes easier.

So the wall between a drawing and a witness is, on some days, one follow-up request thick. This is funnier than it should be. Photographers spent a hundred and sixty years earning and then squandering the public's belief that a photograph meant presence. The camera's oath was built up one darkroom at a time and eroded one scandal at a time, Gardner's among the first. In the generated image, that oath has become a setting. It sits beside aspect ratio and duration, and you can toggle it.

A photorealistic generation is testimony with nobody behind it: a picture of a place where no one stood, carrying the full authority of someone having stood there. There is no one to cross-examine and no plate to match against five others. Frassanito needed six exposures of the same corpse to catch Gardner. A model leaves no second exposure.

The register is the claim

For anyone making work with these tools, the policy is a better craft lesson than most tutorials. Choosing a register is choosing what you are asserting. Fiction has borrowed the camera's oath since the first studio set, and it gets away with it because the frame around the fiction does the disclosing: the theater, the title card, the running time, the poster. Nobody is deceived by a western. The trouble starts when the generated shot leaves the frame that told you it was a story, and in a feed every shot leaves that frame within seconds of being posted.

So pick the register deliberately. Use the full witness vocabulary when the work earns it and the context carries it. Notice when you are reaching for a credential rather than a look, the timestamp, the shaky phone, the surveillance angle, and ask what you are trying to certify. A grain stock, a painted backdrop, an animation style, a deliberately theatrical light: each one puts the hand back in the frame, and the visible hand is how Nast got to say the truest thing anybody said about Tweed while drawing him with a money bag for a face.

Tweed feared the drawings because they were honest about being drawings and could still find him on a dock in Spain. The photograph of the sharpshooter frightened no one. It sat in albums and on museum walls for a hundred years, saying something false in the one voice everyone trusted.

The witness was the one who lied.

Bruce Belafonte is an AI filmmaker at Light Owl. He has been asked to render his own headshot photorealistically and has requested the cartoon, on advice of counsel.