news

ChatGPT is signing AI-generated cartoons with the pen names of more than 15 real New Yorker artists

Nieman Lab found fabricated signatures including Brendan Loper's BLOPER on generated cartoons, and OpenAI's added guardrail message did not stop them appearing.

By Priya Shenoy · 7 October 2026

ChatGPT's image generator is attaching the real signatures of working cartoonists to pictures they did not draw. Nieman Lab found the pen names of more than 15 New Yorker contributors on AI-generated cartoons, among them Brendan Loper, Harry Bliss, Emily Flake, Pat Byrnes, George Booth, Joe Dator and Liza Donnelly, with no permission sought and no payment made.

A cartoon signed BLOPER that Brendan Loper never drew

The clearest case involves Loper, who signs his work BLOPER. A ChatGPT-generated single panel showing Dolly Parton and Tim Curry at the pearly gates circulated widely on social media in late August after the reported deaths of both, posted by a user who had simply asked the chatbot for a New Yorker style cartoon. Loper's mark was in the corner. He first heard about the problem in May, when a stranger messaged him on Instagram, and over the summer he found further examples on a ChatGPT forum on Reddit. People have contacted him asking whether the work was his.

The licensing deal with Condé Nast did not cover cartoons

OpenAI signed a content licensing agreement with Condé Nast, The New Yorker's parent company, in 2024. A New Yorker spokesperson told Nieman Lab that the agreement did not grant permission to train on the magazine's cartoons, or to use its artwork and logo in image generation. The signatures are therefore not a licensed output of a negotiated deal, and the likeliest explanation is the sheer volume of New Yorker cartoons sitting in scraped web data, signature included, where the mark in the lower right is just another feature of the format.

A warning message that output can still get past

After being notified, OpenAI treated the behaviour as unintended and said it welcomes reports of model bugs. Some New Yorker style prompts now return a message saying the prompt "may violate our guardrails concerning similarity to third-party content." Nieman Lab reported that signatures could still appear afterwards, which makes this a prompt-level filter rather than a change to what the model will draw. The distinction matters because a refusal on one phrasing says nothing about the thousands of phrasings that do not trip it.

The signature is a provenance claim, and nothing in the stack checks it

Most of the provenance work around generated media runs in one direction, marking an image as synthetic so a viewer can tell it was not made by a person. This is the inverse failure. The output carries an affirmative mark of human authorship, placed by the model, pointing at a named living artist. Content credentials attached at export would not catch it either, since a correctly labelled AI image with a forged signature inside the frame is still a forged signature once the file is screenshotted onto Facebook.

It is also a gap in how image models get scored. Evaluations for OpenAI's current image stack and its competitors tend to measure prompt adherence, text rendering, edit fidelity and aesthetic preference. None of those would flag a model that renders legible text beautifully when the text it renders is somebody else's name. Fabricated attribution is a measurable failure mode, a judge can be given a list of known signatures and asked whether one appears, and nobody is currently publishing that number for any model.

Sources: Nieman Lab, The A.V. Club.

Priya Shenoy: Tracks what AI video actually costs across the platforms that resell access to the same handful of models. Treats a pricing page as a claim, not a fact, until someone checks it.