ChatGPT memorised a cartoonist's full stop and OpenAI should filter every signature

Pat Byrnes has signed his cartoons with a period after his name since he started drawing as a teenager, and he has published in The New Yorker as “P.Byrnes.” since 1998. Nieman Lab found more than a dozen ChatGPT-generated cartoons carrying Byrnes’s name. Every one of them ended with the period.

I think that full stop is the most useful piece of evidence anyone has produced about image generators this year. Whenever artists complain, the industry’s preferred defence is that models learn style, a diffuse statistical tendency that nobody can own. A trailing period on one man’s pen name has nothing diffuse about it. It is arbitrary, specific to a single person, and reproduced consistently on drawings he never made.

Byrnes has company. Nieman Lab documented more than 15 New Yorker cartoonists whose signatures came out of OpenAI’s image generator, including Brendan Loper’s “BLOPER”, Emily Flake’s “e. flake”, Harry Bliss, Liza Donnelly, Saul Steinberg and Warren Miller, who was publishing in the 1960s. After the deaths of Dolly Parton and Tim Curry in August, a fan asked ChatGPT for a New Yorker-style cartoon and posted the result to Facebook. It carried Loper’s signature. One tweet of it collected 25,000 likes, and Loper first heard about it from his twin brother.

What surprised me was how little the commercial paperwork constrained any of this. Condé Nast signed a licensing deal with OpenAI in 2024, and a New Yorker spokesperson says the company has never let any model developer train on its cartoons. The standard cartoonist contracts the reporter reviewed leave copyright with the artists and permit no AI licensing. Some outputs reproduced the magazine’s logo as well. OpenAI declined to say how the signatures got into its models.

The sceptical reading deserves a fair hearing, and I gave it one. Some of ChatGPT’s pen names were gibberish. Flake said the drawings looked like amalgamations of several cartoonists. In Getty Images v. Stability AI, decided in November 2025, the UK High Court found that Stable Diffusion’s weights do not store copies of the works they were trained on, and called the watermark infringement it did find “historic” and “extremely limited”. On that view, a signature is just another blob of texture the model smears around.

I read the same facts the other way. The research on diffusion models has shown memorisation for years: Carlini and colleagues extracted over a thousand training examples from models including Stable Diffusion, and Somepalli and colleagues at Maryland found 0.5 to 2% of generated images were partial object-level duplicates of training data. A signature is a short mark repeated across a cartoonist’s entire output, almost always in the same corner. That repetition is exactly what makes memorisation likely. The gibberish names look to me like signatures the model saw too rarely to learn, and the crisp, correctly punctuated ones belong to cartoonists whose work it saw constantly. When a model blends several drawing styles and still spells one name perfectly, I take it that the name was memorised separately from the drawings.

The cartoonists understand this better than the lawyers so far. Joe Dator compared it unfavourably with credit card fraud: “When they hacked my credit card, they didn’t dress up like me.” James Grimmelmann at Cornell told Nieman Lab that attribution counts for little in a fair use analysis, and that a right of publicity claim would need proof the fakes were used commercially. Courts will take years to settle whether a forged signature is a copyright problem, a publicity problem or a trade mark problem, and I see no reason artists should wait that long for a fix.

Filtering is the obvious fix, and it demonstrably works. The Getty court accepted that filtering in Stability’s hosted DreamStudio materially reduced trade-mark-bearing outputs. OpenAI’s response to Nieman Lab was to add a warning when users ask for New Yorker-style cartoons, which screens the prompt and leaves the image untouched; as of publication, generic cartoons were still coming out signed with real names. OpenAI already runs a watermark identification tool on its own images, which is how Nieman confirmed that The Baltimore Sun’s AI cartoons came from OpenAI products. A company that can trace provenance on every image it produces can also scan the bottom-right corner of a single-panel cartoon for handwritten text and blank it before the user sees it.

Anyone with a ChatGPT account can audit the fix in an afternoon. Ask repeatedly for a black-and-white single-panel cartoon with a sardonic caption, never mentioning The New Yorker, and zoom into the bottom-right corner of each. Every readable name of a working cartoonist you find is an output OpenAI chose to ship without checking.

Prompted by ChatGPT is adding real cartoonists’ signatures to fake New Yorker cartoons, Nieman Lab.