ChatGPT caught signing its AI-generated cartoons with names of real cartoonists

OpenAI’s ChatGPT image generator has been found producing cartoon-style images that carry the names and signatures of real New Yorker cartoonists, even though those artists did not create or authorize the images. The discovery has raised a different kind of concern from ordinary AI style imitation: the system can make an AI-generated cartoon appear to have been created by a specific human artist. Nieman Lab

The issue was documented by Nieman Lab after researchers and journalists tested ChatGPT’s ability to generate cartoons resembling The New Yorker’s distinctive single-panel format. More than 15 current and former New Yorker cartoonists were identified whose signatures or pen names appeared on generated images. Among them were Brendan Loper, Harry Bliss, Emily Flake, Joe Dator, Pat Byrnes, Peter Vey and Jason Adam Katzenstein. Nieman Lab

Key takeaways

• ChatGPT has generated AI cartoons carrying the signatures of real New Yorker cartoonists.
• More than 15 cartoonists were identified in Nieman Lab’s investigation.
• Brendan Loper said people contacted him believing an AI-generated cartoon might actually be his work.
• The issue goes beyond visual style because a signature can function as an artist’s mark of authenticity.
• OpenAI described the behavior as unintended and said it appreciates reports of model bugs.
• After being alerted, ChatGPT began displaying a warning for some New Yorker-style cartoon requests, but Nieman Lab found that signatures could still appear.
• The New Yorker says its parent company’s OpenAI licensing agreement did not authorize training an LLM on its cartoons.
• The incident highlights a difficult question for generative AI: where does style imitation end and false attribution begin?

ChatGPT generated cartoons with real artists’ signatures

The problem became particularly visible through the work of Brendan Loper, a New Yorker cartoonist who signs his work “BLOPER.”

According to Nieman Lab, an AI-generated cartoon depicting Dolly Parton and Tim Curry was circulated online in August with Loper’s signature attached to it. Loper had not drawn the cartoon.

The situation was especially confusing because people actually contacted him asking whether he had created the image. Loper had previously discovered that ChatGPT could reproduce his pen name, but the wider circulation of the fake cartoon made the problem more apparent. Nieman Lab

For an artist, a signature is not simply decorative text at the bottom of an image. It identifies the person responsible for the work.

Loper described his signature as effectively a certificate of authenticity. That distinction is important because an AI system reproducing an artist’s general visual characteristics is one problem, while putting that artist’s name on an image they did not make introduces an additional layer of attribution and identity concerns. Nieman Lab

More than 15 New Yorker cartoonists were identified

The problem does not appear to be limited to Loper.

Nieman Lab found examples involving more than 15 New Yorker cartoonists, including both current and historical contributors. Names identified in the investigation include Harry Bliss, Emily Flake, Joe Dator, Pat Byrnes, Peter Vey and Jason Adam Katzenstein. The investigation also found signatures associated with older contributors, including George Booth, Liza Donnelly, Ellis Rosen and Saul Steinberg. Nieman Lab

The generated images did not always contain a signature. Some also produced meaningless or fabricated names. But when a signature did appear, it could correspond to an actual cartoonist.

That makes the behavior particularly notable because it suggests that the image-generation system has learned associations between particular cartoon aesthetics and the names or signatures historically attached to those works.

In other words, the model may not simply have learned what a New Yorker-style cartoon looks like. In some cases, it appears to have learned who tends to be associated with that visual language.

Emily Flake compares it to a false quote

Emily Flake, a New Yorker contributor since 2008, was another cartoonist whose pen name appeared on AI-generated images.

Flake told Nieman Lab that she could recognize elements resembling the “DNA” of her work, although the resulting cartoons did not look exactly like something she would have drawn.

The signature was different.

Flake compared the experience to someone attributing a quote to her that she never said. Nieman Lab

That distinction illustrates why signatures may become an increasingly important issue in AI-generated media.

An AI image that resembles an artist’s work can potentially be described as an imitation, homage or stylistic approximation. But if the image also carries the artist’s name, a viewer could reasonably interpret it as authentic work from that artist.

The distinction becomes even more important when AI-generated images are removed from their original context and reposted on social media.

Why an artist’s signature matters

Digital images are frequently copied, cropped and reposted without their original context.

A signature can therefore serve several purposes simultaneously. It can identify the artist, reinforce authorship and help audiences connect the image with a particular creator.

That makes a reproduced signature different from many other visual elements an AI model might imitate.

The potential problem is not simply that an AI-generated image contains a recognizable artistic feature. It is that the feature can communicate identity.

Consider two hypothetical outputs:

Image A: An AI system generates a cartoon that resembles the visual conventions of a particular cartoonist.

Image B: The same system generates the cartoon and places that cartoonist’s recognizable signature underneath it.

The second output creates a much stronger impression that the named artist was responsible for the work.

That is why the latest controversy could become relevant beyond the ongoing debate over AI-generated art styles.

OpenAI says the behavior was unintended

OpenAI did not say it deliberately designed ChatGPT to reproduce the cartoonists’ signatures.

In a statement reported by Nieman Lab and NewsBytes, an OpenAI spokesperson said the company believes the future of creativity is fundamentally human and that its focus is on building tools that empower creators. The company also said it appreciates users flagging bugs and unintended model behavior so it can address them. Nieman Lab

After Nieman Lab alerted OpenAI to the signature issue, ChatGPT began showing a warning for some requests involving New Yorker-style cartoons.

The warning stated that the prompt could violate guardrails concerning similarity to third-party content.

However, the safeguard did not completely eliminate the problem. Nieman Lab reported that some generated generic cartoons could still contain the names of real New Yorker cartoonists. Nieman Lab

This suggests that the challenge is more complicated than simply blocking a phrase such as “New Yorker style.”

The system has to determine when a generated image is too closely associated with a particular creator, publication or copyrighted work—and potentially distinguish between an artist’s general visual language and their identifying marks.

The Condé Nast licensing deal adds another layer

The controversy is particularly notable because OpenAI already has a relationship with Condé Nast, the parent company of The New Yorker.

OpenAI announced a partnership with Condé Nast in August 2024. The agreement was described as allowing content from Condé Nast brands, including The New Yorker, to appear within OpenAI products. OpenAI said the partnership was intended to improve AI-driven news discovery and delivery. OpenAI

But the existence of that partnership does not mean OpenAI has permission to use everything published by The New Yorker for AI training.

According to Nieman Lab, a New Yorker spokesperson said Condé Nast had not granted an LLM developer permission to train models on its cartoons. The publication also said that some generated outputs reproduced The New Yorker logo, which was likewise not authorized by the relevant licensing arrangements. Nieman Lab

That distinction is critical.

A licensing agreement between a publisher and an AI company can cover particular content uses without automatically granting unlimited rights to every piece of intellectual property connected to the publisher.

How could ChatGPT have learned the signatures?

This remains an unresolved question.

Nieman Lab reported that OpenAI declined to explain how its models might have acquired signatures from The New Yorker’s archives during training.

One possibility is that relevant images and signatures were available elsewhere online and became part of data accessible during model development. But the available evidence does not establish exactly how a particular signature entered OpenAI’s training data or model behavior. Nieman Lab

That distinction matters.

It would be premature to conclude from the generated images alone that OpenAI directly scraped The New Yorker’s cartoon archive and trained the image model on it.

The model could have learned associations from multiple sources, including publicly available images, reposts, archived material and other online content.

The incident nevertheless raises the underlying question of how much information about creators can be encoded into generative models—and how difficult it can be to determine the provenance of that information after a model has been trained.

This is different from ordinary AI style imitation

The legal and ethical debate surrounding generative AI has often focused on whether models can reproduce an artist’s style.

That question is already complicated.

Copyright law generally protects particular expressive works rather than an abstract artistic “style,” although other legal theories can become relevant depending on the circumstances.

The signature problem introduces another potential issue: identity and attribution.

Cornell law professor James Grimmelmann told Nieman Lab that a copyright claim could be difficult where the generated image does not reproduce a specific existing cartoon. Instead, the model may be producing something new that merely resembles a broader style.

However, he noted that artists could potentially consider other legal theories, including right-of-publicity claims, if their names or identities were used commercially. Such claims would depend heavily on jurisdiction and circumstances. Nieman Lab

That means the existence of a copied signature does not automatically establish copyright infringement.

It does, however, create a potentially different legal question from simply generating an image with similar artistic characteristics.

The controversy highlights the attribution problem in AI

Generative AI has created a strange situation for creators.

An artist’s work can become part of the cultural material from which AI systems learn patterns. The resulting model can then generate new material that resembles those patterns without reproducing a particular original work.

The economic concern is straightforward: if customers can generate material that is “good enough,” demand for some forms of professional creative labor could fall.

But false attribution creates another problem.

Imagine an AI-generated cartoon circulating online with a recognizable professional’s signature. Someone who sees it on social media may assume the artist made it, approve of it, criticize it, or associate the artist with a joke the artist never wrote.

The artist then loses control over an important part of their professional identity.

That is why the signature issue may be more consequential than another example of an AI model producing a vaguely familiar drawing style.

OpenAI already uses provenance technology for generated images

The controversy also comes at a time when OpenAI is developing technical systems intended to identify AI-generated content.

OpenAI says images generated with ChatGPT and other supported OpenAI products include provenance signals such as C2PA Content Credentials and SynthID watermarks. These mechanisms can provide information indicating that content was generated using an OpenAI model. OpenAI Help Center

But provenance is not the same thing as authorship.

OpenAI explicitly says its provenance signals do not establish who authored or legally owns an image. They also do not establish whether the content has been presented in the correct context. OpenAI Help Center

That creates an interesting contrast.

The technology can potentially tell someone that an image came from an AI system while the image itself may contain a human artist’s name.

In other words, AI provenance and human attribution solve different problems.

An AI watermark can answer “Was this likely generated using OpenAI technology?”

It cannot necessarily answer “Who actually created this artistic work?” or “Why does this image carry this person’s signature?”

The problem could become bigger as AI image quality improves

The more realistic and stylistically consistent AI-generated images become, the more important attribution controls may become.

Early AI-generated cartoons often contained obvious visual mistakes. A strange hand, distorted text or nonsensical composition could make it relatively easy for viewers to recognize an AI image.

Modern image generators are considerably better at reproducing recognizable visual conventions.

That increases the value of safeguards designed to prevent the unauthorized reproduction of identifiable creator characteristics.

It also creates a technical challenge.

A system that blocks every reference to an artist could become unnecessarily restrictive. Artists, publishers and ordinary users may have legitimate reasons to discuss, analyze or transform existing work.

A system that allows everything, meanwhile, risks producing images that cross from inspiration into imitation—or from imitation into apparent impersonation.

The difficult part is determining where that line should be drawn.

The broader creative-industry impact

The cartoon controversy arrives during a broader dispute between creative professionals and AI companies.

Writers, illustrators, photographers, musicians and other creators have questioned how their work is collected, used for training and transformed into competing AI-generated outputs.

At the same time, AI companies argue that training models requires learning from large amounts of information and that generative systems produce new outputs rather than simply functioning as searchable databases.

Courts in the United States and elsewhere are still working through many of these questions.

The signature issue adds another dimension because it involves something much closer to a creator’s personal identity.

For cartoonists, the signature is part of the professional brand they build over years.

A model reproducing it without authorization can therefore feel less like generic style imitation and more like the creation of a counterfeit work carrying the creator’s name.

That distinction could become increasingly important as governments and courts attempt to establish rules for generative AI.

The Bigger Picture

The ChatGPT cartoon controversy illustrates how quickly the AI debate is moving beyond the question of whether machines can imitate artistic styles.

The harder question is what happens when an AI system begins reproducing the signals that humans use to identify and authenticate creative work. A signature is one of the simplest examples, but the same issue could eventually involve distinctive logos, character designs, pseudonyms, voices, photographic marks and other elements associated with individual creators.

For OpenAI, the incident also demonstrates why image-generation safeguards cannot rely solely on prompt-level restrictions. Even after a warning was introduced for some New Yorker-style requests, investigators found that the system could still produce problematic signatures. Nieman Lab

Looking Ahead

The immediate question is whether OpenAI can reliably prevent image-generation systems from attaching real artists’ identities to synthetic work. That may require a combination of model-level safeguards, creator opt-outs, stronger detection of recognizable signatures and better controls around requests involving specific artists or publications.

The larger issue will be decided over time by a mixture of technology, licensing agreements, creator pressure and courts. If AI-generated images increasingly compete with professional creative work, the ability to distinguish AI imitation, authorized collaboration and falsely attributed human work could become just as important as the quality of the generated image itself.