ChatGPT Signs Fake New Yorker Cartoons With the Names of 15+ Real Cartoonists

TL;DR

ChatGPT is generating "New Yorker-style" cartoons and signing them with the pen names of real New Yorker cartoonists, according to an investigation by Nieman Lab's Andrew Deck published October 5. Deck documented more than 15 cartoonists whose signatures OpenAI's image generator has reproduced without permission or payment, from current contributors like Emily Flake and Pat Byrnes to George Booth and Saul Steinberg. The viral example is a pearly-gates cartoon of Dolly Parton and Tim Curry that spread after both died in late August, signed "BLOPER," the pen name of Brendan Loper, who never drew it. OpenAI called the behavior "bugs and unintended behavior," added a warning message after Nieman Lab got in touch, and was still signing some outputs with real names when the story ran. The New Yorker says no LLM deal, including Condé Nast's 2024 agreement with OpenAI, ever covered its cartoons.

The cartoon that went viral with the wrong name on it

The premise: Dolly Parton in a sparkling gown, arm in arm with Tim Curry as Dr. Frank-N-Furter, tells God at heaven's reception desk, "I knew I chose the right plus one." It was posted to Twitter on August 26, 2026, the day after both entertainers died, and one tweet of it collected 25,000 likes. Copies circulated on Facebook, Instagram, and Bluesky. In the bottom-right corner sits "BLOPER."

Brendan Loper did not draw it. A Dolly Parton fan generated it with ChatGPT and posted it to Facebook, explaining in her comments that she had simply asked for "a New Yorker-style cartoon." Loper found out from his twin brother, whose friend had seen it, and then from a stream of strangers asking whether it was his. "It was very surreal," he told Deck.

It was not even the first time. In May a stranger messaged Loper on Instagram to warn him his pen name had appeared on a ChatGPT image. Over the summer he found more in a Reddit forum where users were posting the New Yorker-style cartoons they had generated. Loper already knew his work had been "harvested" for training. A forged signature felt different. "My name is my name. It felt very much like a violation of my personhood."

More than 15 signatures, one from the 1960s

Loper is the hook, not the scope. In the same forums Deck found dozens of images carrying the signatures of Harry Bliss, Emily Flake, Joe Dator, Pat Byrnes, Peter Vey, and Jason Adam Katzenstein. In his own prompting, ChatGPT produced cartoons signed by George Booth, Liza Donnelly, Ellis Rosen, and Saul Steinberg, plus Warren Miller, who published in the magazine in the 1960s. The total: more than 15 New Yorker cartoonists, past and present.

The outputs reliably hit the 101-year-old magazine's house style: black-and-white single panel, sparse background, clean linework, a sardonic caption or an approximation of one. Not every image carried a signature, and some pen names were gibberish. But when a legible signature appeared, it was regularly a real contributor's.

Flake, a contributor since 2008, described the fakes signed "e. flake" as "a distant cousin": echoes of her stances and shading, but mostly an amalgam of several cartoonists. The signature, though, was unmistakably hers. "It's like somebody attributed a quote to me that I didn't say." Byrnes has signed "P.Byrnes." with a trailing period since his teens and published under it in the New Yorker since 1998. Across more than a dozen ChatGPT cartoons carrying his name, every one ended with a period. The model cannot reliably land a joke, but it has nailed the punctuation.

Why a model forges a signature nobody asked for

The fan did not prompt for "BLOPER." She asked for a New Yorker-style cartoon and got the whole package. That is the mechanism, and the Hacker News thread on the story lands on it quickly: nearly every training example of this genre carries a signature in the corner, so the signature is learned as part of the style. The model is not attributing anything. It is completing a pattern in which a specific squiggle sits bottom-right on this kind of image, the same way a sparse background and a wry caption do.

Think of a cover band that learned its whole set from live bootlegs. It now shouts "Thank you, Cleveland!" at the end of every song, because that is how the recordings end. The signature is the model's "Thank you, Cleveland."

OpenAI's statement to Nieman Lab: "We believe the future of creativity is one that is fundamentally human, and our focus is on building tools that empower human creativity and creators. We very much appreciate the community flagging bugs and unintended behavior by our models, so we can address them." After Deck notified the company, New Yorker-style prompts began returning a new line: "This prompt may violate our guardrails concerning similarity to third-party content." The outputs kept coming, and some kept carrying real names. The warning is a prompt classifier. The signature is in the pixels. This is the same family of quiet, prompt-side style guardrails that ChatGPT rolled out for named authors in July, with the same gap: it screens the request, not the result.

The license that never covered cartoons

Condé Nast, the New Yorker's parent, signed a multi-year licensing deal with OpenAI in 2024. The terms are private, but a New Yorker spokesperson told Deck the company has never granted any LLM developer permission to train on its cartoons, and that reproducing the magazine's logo, which some ChatGPT outputs also did, is not allowed under any deal either. OpenAI declined to answer how its models ingested the signatures.

Even if Condé Nast wanted to license the cartoons, it mostly cannot. Deck reviewed the magazine's templated cartoonist contracts: freelancers submit on spec, get paid only if a cartoon is selected, and retain copyright. It is not work for hire, and none of the contracts permit licensing for AI training. The likeliest ingestion path is the open web. Court documents unsealed last month in New York Times v. OpenAI show OpenAI knowingly circumvented publisher paywalls while building ChatGPT; one unredacted Microsoft memo called the training data "the largest theft of labor in human history." (AI Bacon covered the unsealed brief when it dropped.)

Copyright cannot reach a signature. Publicity rights might.

The cartoonists draw a line between style mimicry and name mimicry, and so does the law, in an awkward way. Katzenstein, a contributor since 2014, calls the fakes "an uncanny simulacrum of an impression of a joke" and the signature something closer to impersonation: "this is slop that's pretending to be me." Dator, 20 years at the magazine, has had his credit card hacked and ranks that the lesser violation: "When they hacked my credit card, they didn't dress up like me."

Cornell's James Grimmelmann told Nieman Lab that a copyright claim would hinge on reproducing a specific drawing style, and "attribution is a very small part of the fair use inquiry." ChatGPT is usually not copying a particular cartoon; it is mimicking a genre. The Andersen v. Stability AI class action, grinding on since 2023, never mentions signatures. The sharper tool is a state right-of-publicity claim over unauthorized use of a name, but that needs proof of commercial use, and Deck found no evidence the fakes are being bought or sold.

Katzenstein has been here before. He was part of the Authors Guild class action against Anthropic, after a HarperCollins anthology containing his cartoons turned up in the pirated book sets used to train Claude. That case ended in the $1.5 billion settlement, the largest in US copyright history; AI Bacon covered the per-title payouts. That money was for downloading pirated books. Nobody has yet been paid for a forged byline.

The job this threatens

None of the contributors Deck spoke to can live on New Yorker commissions alone. The magazine credit is a calling card that opens commercial illustration, book work, and other gigs, and that is the work an image generator undercuts first. The proof case is in Baltimore: The Baltimore Sun fired Kevin Kallaugher, a Pulitzer finalist with more than 30 years at the paper, last year, and this year began running AI-generated cartoons in its opinion section. Deck ran them through OpenAI's watermark identification tool; they were made with OpenAI products. The Association of American Editorial Cartoonists condemned the move in an open letter last month: "The Sun is not alone among American newspapers taking this degrading way out."

What to take from this if you build with image models

• Filter outputs, not just prompts. OpenAI's fix is a prompt-side warning, and the names still appear. If you ship generation, run OCR over outputs against a denylist of known signatures, bylines, and logos, the way platforms already screen for trademarks.
• Style conditioning carries bylines along. Any "in the style of" feature trained on signed work learned the signature as part of the style. Red-team for names, watermarks, and logos in the corners, not only for the look.
• A license is about the future. Weights are a record of the past. Condé Nast's deal excludes cartoons and the model draws them anyway. A negotiated carve-out does not untrain anything that was already scraped.
• Commercial use is the live legal risk. Grimmelmann's publicity-rights path needs a commercial purpose. A consumer gag is a weak case. A paid product that emits real names on fake work is a much better one.
• For creators, provenance is the only signature left. A hand-drawn mark no longer proves authorship; Deck's own forensic step was a watermark detector. Content credentials on your real work are the version of a signature a model does not learn to reproduce.

Caveats

Nieman Lab's "more than 15" comes from trawling public forums plus Deck's own prompting, not from a sampled rate, and the article does not say how many prompts he ran or what share of outputs carried a real name. OpenAI did not explain how the signatures entered training, so the open-web path is inference, supported by the NYT v. OpenAI record but not confirmed for these images. The 25,000 likes is one tweet's count. The guardrail changed between notification and publication and may change again; the signing behavior is as of October 5. And no one has shown the fakes being sold, which is precisely the fact a publicity-rights case would need.

Key Takeaways

• Nieman Lab documented more than 15 New Yorker cartoonists, including George Booth and Saul Steinberg, whose signatures ChatGPT pastes onto AI-generated cartoons without permission.
• The viral case: a Dolly Parton and Tim Curry tribute signed "BLOPER," posted August 26 and liked 25,000 times on one tweet, which Brendan Loper never drew. He had been warned in May.
• OpenAI called it "bugs and unintended behavior" and added a prompt-side warning; outputs still carried real names at publication.
• The New Yorker says no LLM deal covers its cartoons, including Condé Nast's 2024 OpenAI agreement, and its on-spec contracts leave copyright with the cartoonists.
• Copyright stalls because style is not protected; right of publicity stalls because no commercial use has been shown. The signature falls in the gap.
• Builders: screen generated pixels for names and logos, assume style training memorized bylines, and treat provenance metadata, not a drawn signature, as the authenticity layer.

Sources: Nieman Lab (Andrew Deck, Oct 5, 2026), The A.V. Club, Hacker News discussion, RightsTech on the Condé Nast-OpenAI deal (Aug 2024), Nieman Lab on the NYT v. OpenAI records (Sep 2026), Andersen v. Stability AI docket, Anthropic copyright settlement, Association of American Editorial Cartoonists