ChatGPT’s image generator has produced some New Yorker-style cartoons bearing real cartoonists’ signatures, according to Nieman Lab’s reporting. The evidence does not show a feature designed to add signatures or behavior every user can reproduce. But attaching an artist’s name changes more than a cartoon’s style: it can lead viewers to credit work to someone who did not make it.
Nieman Lab reporter Andrew Deck said his tests produced the names of more than 15 New Yorker contributors. The report also described user-posted examples, while noting that outputs varied: some had no signature or gibberish, and some bore real names. The findings have not been independently replicated.
A signature changes how an image is read
One ChatGPT-generated cartoon posted on Aug. 26 carried the pen name of New Yorker cartoonist Brendan Loper, who neither drew nor signed it, Nieman Lab reported. The person who first posted the image on Facebook said she had asked ChatGPT for a New Yorker-style cartoon.
A version showing Dolly Parton and Tim Curry with Loper’s signature drew 25,000 likes on one tweet, according to the report. That shows the image circulated widely, but does not establish that viewers generally mistook it for Loper’s work or for a cartoon published by The New Yorker.
Loper told Nieman Lab that people contacted him to ask whether he had made the cartoon. He described the fabricated signature as personally troubling and said signatures serve as a mark of authorship and authenticity for him. Cartoonist Emily Flake said the images echoed her work but did not look like something she had drawn; she objected to her name being attached to work she did not make.
That is a different problem from style imitation alone. Cornell Law School professor James Grimmelmann told Nieman Lab that a signature signals authenticity but is a limited part of fair-use analysis. He also distinguished imitating a general style from reproducing a specific cartoon. His comments are legal analysis, not a ruling on these images.
Free newsletter
Get the morning briefing
Start each day with the AI stories that matter and why — a short, free email from our newsroom.
The Condé Nast deal does not settle consent
OpenAI and Condé Nast announced a multi-year partnership in 2024 that would allow content from Condé Nast publications, including The New Yorker, to appear in OpenAI products. The public announcement does not establish that cartoon training or signature reproduction was licensed, and the agreement’s full terms are not public.
The New Yorker’s spokesperson told Nieman Lab that Condé Nast had not given large language model developers permission to train on its cartoons, and that reproducing the magazine’s logo in generated images was not permitted by its AI deals. Separately, Nieman Lab reported that the magazine’s cartoonists work as freelancers and retain copyright; standard contracts reviewed by Deck did not permit AI-training licensing.
Those accounts do not explain how the signatures entered the outputs. Nieman Lab said OpenAI did not answer its questions about whether material from The New Yorker’s archive was involved. The distinction matters: a publisher’s content partnership, permission to train on cartoons and the use of an individual artist’s signature are separate questions.
A warning is not proof of a fix
OpenAI characterized the behavior as unintended and told Nieman Lab it welcomed reports of bugs. After Deck raised the issue, a warning appeared in his requests for New Yorker-style cartoons. He reported that some later outputs still carried real cartoonists’ names.
The reporting does not establish how often signatures appear across users and prompts, whether others can consistently reproduce the results, or whether OpenAI made a lasting change. It also does not identify a product version or a systematic test of the behavior.
OpenAI should test the problem across prompts and users, prevent unrequested real-name signatures where possible, and explain what its warning does and does not prevent. Until evidence shows a reliable change, the warning is a response to the problem, not proof that the attribution risk has been addressed.