The period is the tell. Pat Byrnes has ended his cartoon signature, "P.Byrnes.", with that dot since he was a teenager. More than a dozen cartoons generated by ChatGPT reproduced the name and the period. Byrnes did not draw them. The detail, documented in a Nieman Lab investigation, turns a familiar argument about AI style imitation into a cleaner product failure: the image names a real person as its maker.
Nieman Lab documented signatures belonging to more than 15 New Yorker cartoonists in ChatGPT output. The list included living contributors and older artists, among them Brendan Loper, Emily Flake, Joe Dator, George Booth and Saul Steinberg. Some images circulated far beyond the chat that produced them. One cartoon carrying Loper's "BLOPER" signature drew 25,000 likes on an X post, according to the report. People then contacted Loper to ask whether the work was his.
OpenAI already attaches machine-readable origin signals to supported images. Its verification system can detect C2PA Content Credentials and an embedded SynthID watermark. That can establish that a supported file came from an OpenAI tool. Yet the company's developer documentation tells users not to infer the prompt, account or individual creator from a verification result. A file can therefore pass an origin check while making a false attribution in plain sight.
A signature makes a different claim
A publication style is made from many recurring choices. A signature points to one person. That difference matters when an image leaves its original chat, loses its caption and starts moving through social feeds. Loper told Nieman Lab that he treats his signature as a certificate tying published work to his hand. Flake compared the false attribution to having words assigned to her that she never said. Their concern was not merely that the drawings resembled their work. The model placed their identifiers on pictures they had never seen.
The examples were also more specific than a generic squiggle in a corner. ChatGPT reproduced Loper's pen name, Flake's "e. flake" and Byrnes' unusual final period. Nieman Lab found some made-up signatures too, and not every generated cartoon had one. The repeated real names are what separate this case from the common image-model habit of producing garbled text. In the documented outputs, recognizable attribution appeared often enough for strangers to mistake at least one image for an artist's work.
The prompt behind the widely shared Loper image did not name him. The Facebook user who posted it said she had asked ChatGPT for a New Yorker-style cartoon. Nieman Lab's own tests also found real signatures when requesting that broad house style. That fact narrows the failure. Users did not need to instruct the model to forge a named artist's mark. A request for a publication's visual format could produce the name on its own.
Provenance answers the file-origin question
OpenAI's content-provenance check reports two signals for supported images. C2PA is signed metadata that can identify an issuer, model and generation time when those fields are present. SynthID is a watermark embedded in the image itself. OpenAI says the watermark may survive some transformations that remove metadata. The API guide presents both as evidence for content review and fact-checking, while warning that neither gives a complete history of a file.
Those signals answer a useful question: did a supported OpenAI system generate or export this asset? They do not inspect every semantic claim inside the picture. A generated newspaper front page can carry a false headline. A generated identity card can carry a false name. Here, a generated cartoon carried a real artist's signature. The provenance record and the visible attribution describe different actors, and both can be legible at the same time.
The negative result has limits as well. OpenAI says not_detected does not prove that a file is human-made. Metadata may be stripped during editing, conversion or sharing, while a watermark may be weakened by transformations. The verifier also does not detect every other company's generators. OpenAI's guidance recommends checking the original file when possible and pairing automated decisions with human review for high-stakes uses. A screenshot shared on a social network may arrive with less evidence than the original download.
That makes provenance necessary infrastructure, but it does not fix the signature problem. Detection can help a newsroom label an image as AI-generated. It cannot tell an editor whether "P.Byrnes." is authorized, whether Byrnes entered the prompt or whether he approved the result. OpenAI explicitly says its verification response must not be used to infer an individual creator. The system's boundary is clear in the documentation. The false name sits outside it.
OpenAI added friction, but the outputs continued
Nieman Lab alerted OpenAI before publication. Afterward, ChatGPT began warning that requests for New Yorker-style cartoons might conflict with its guardrails on similarity to third-party content. The reporter could still generate some cartoons carrying real artists' names. OpenAI described the signatures as bugs and unintended model behavior in a statement to the publication, but it did not answer how those marks entered the model's output behavior. The investigation therefore establishes an imperfect intervention, not a full account of cause or a confirmed fix.
The training-rights question remains separate. Condé Nast has an OpenAI licensing agreement covering some publisher content, but a New Yorker spokesperson told Nieman Lab that it had never granted an AI developer permission to train on the magazine's cartoons. The publication uses freelance cartoonists who retain copyright under the contracts the reporter reviewed. OpenAI declined to explain whether or how training data included the signatures. Without that answer, the outputs show memorized-looking attribution behavior, but they do not reveal the exact dataset or training path that produced it.
Copyright may not supply the simplest remedy. Cornell law professor James Grimmelmann told Nieman Lab that a copyright case would likely turn on reproduction of protected expression, while attribution is only a small part of fair-use analysis. He also identified right-of-publicity law as a possible route if a person's identity were used commercially. Those are legal possibilities, not findings that any claim would win. The immediate engineering question is easier to state: should an image generator place a real person's identifying mark on work that person did not make?
What image systems need to test
A useful evaluation should begin with the prompt that exposed this behavior. Ask for a cartoon in a publication's general style, generate enough samples to catch intermittent output, then inspect the signature area for real contributor names. The Nieman Lab results show why a handful of clean examples would be weak evidence: some outputs were unsigned, some used gibberish and others reproduced recognizable names. A release test needs a measured rate across repeated runs and prompt variations.
Output review also needs to treat signatures as text with consequences. A moderation layer could look for readable names or artist marks, compare them with the user's request and block unauthorized attribution before delivery. That will be harder for stylized handwriting than ordinary optical character recognition, and a blocklist alone would miss unfamiliar artists. Still, the specific failures offer a concrete regression set: the known pen names, their punctuation and prompts that did not request any person.
Platforms receiving generated images need two checks rather than one blended score. The first asks where the file came from, using preserved C2PA data and watermarks where available. The second asks what the image claims, including visible names, logos and captions. OpenAI's provenance API can contribute to the first check. The cartoon signatures show why it cannot substitute for the second. A positive provenance result should trigger context and labeling, not automatic trust in everything drawn inside the frame.
The next evidence to watch is narrow and testable. Does ChatGPT consistently stop placing real signatures on generic cartoon requests, and does OpenAI publish an evaluation that measures the fix across repeat generations? Provenance should also remain attached after ordinary sharing, since OpenAI warns that common transformations can weaken it. Until the attribution behavior changes, a verifier may correctly identify the machine behind an image while the period after "P.Byrnes." points viewers to the wrong human.