ChatGPT Caught Forging New Yorker Cartoons and Stealing Real Cartoonists' Signatures

Overview

A story that sparked discussion on Hacker News has revealed that when ChatGPT generates cartoons in the style of The New Yorker, it includes the signatures of real cartoonists in the images. This means the AI is not only imitating the visual style of a prestigious publication — it is also copying the personal signatures of living and deceased cartoonists onto works they never created.

This phenomenon strikes at the heart of two core concerns in generative AI: copyright and attribution ethics. When a model is trained on large volumes of copyrighted work, it doesn't just replicate style — it can also "regurgitate" sensitive elements from its training data, including signatures that serve as a creator's personal identity marker.

Why the Signature Problem Is Especially Serious

A Signature Is Identity and Endorsement

A cartoonist's signature is not a decorative graphic — it is a legal and reputational marker of identity. When a real cartoonist's signature appears on an AI-generated "fake New Yorker cartoon," it effectively forces that artist to "endorse" a work they had no part in creating, without their consent.

If these images are mistaken for originals and circulate online, they could damage the cartoonist's reputation and potentially infringe on their rights of publicity and trademark-adjacent interests in their name.

Exposing the Memorization Problem in Training Data

From a technical standpoint, the fact that a model can accurately reproduce a specific cartoonist's signature reveals that the training data contained a large number of original signed works — and that the model has developed a form of "memorization" of these details. This kind of memorized output is one of the longstanding criticisms leveled at generative models. It demonstrates that these models are not purely "learning abstract styles" but are, in a meaningful sense, storing and reproducing specific protected content.

Memorization in machine learning is generally divided into two types: verbatim memorization, where a model can reproduce specific training samples nearly exactly, and template memorization, where the model retains a highly repetitive structure or stylistic pattern. The reproduction of cartoonist signatures falls into the former category — a signature is a highly distinctive and repeatedly occurring local feature in images, which causes the model to store it with high confidence as part of a particular style. Researchers have found that the more frequently a type of sample appears in training data, the more thoroughly the model memorizes its details (i.e., memorization is positively correlated with repetition frequency). This explains why the more prominent a cartoonist is — and the larger their body of work — the more accurately their signature can be reproduced. Techniques like Differential Privacy can theoretically suppress this kind of memorization during training, but at the cost of generation quality. Large image generation models rarely adopt it as a standard training configuration.

Dual Risks: Copyright and Brand Infringement

The New Yorker's cartoons are among the publication's most recognizable brand assets, defined by a distinctive composition, line style, and sense of humor. AI mass-producing imitations of this style already occupies a copyright gray zone — but when real signatures are layered on top, the risk escalates from "style imitation" to "identity theft."

For model providers like OpenAI, this type of output could invite legal action from individual cartoonists, as well as scrutiny from Condé Nast — The New Yorker's parent company — on brand protection grounds. Several ongoing lawsuits against AI companies revolve precisely around whether training data was used lawfully and whether outputs constitute infringement.

Several landmark cases have emerged in AI-generated content copyright litigation. Getty Images sued Stability AI, alleging unauthorized scraping of millions of copyrighted images for training and pointing to the appearance of Getty watermark remnants in output images — a situation highly analogous to the signature issue described here. Additionally, a group of artists jointly sued Midjourney and DeviantArt, with one core claim being that the combined output of "style + attribution elements" infringes on creators' derivative work rights. These cases have not all reached final verdicts, but they have prompted some courts to begin examining the fundamental question of whether AI training constitutes fair use. For OpenAI, legal risk in the image generation space compounded by the signature controversy means the company faces not only copyright compliance pressure, but also a harder-to-quantify erosion of trust within the creator community.

What Safeguards the Industry Needs

Several directions worth prioritizing for model developers can be distilled from the discussion:

• Output filtering: Identify and remove real-name signatures, watermarks, publication logos, and other identity markers at the image generation stage.
• Training data governance: Apply more rigorous licensing or de-identification processes to copyrighted works.
• Provenance and labeling: Add clear synthetic markers to AI-generated content to prevent it from being mistaken for authentic work.

None of these measures are new ideas — but this incident demonstrates that existing safeguards still have significant gaps at the granular level of signatures.

Conclusion

While the discussion on Hacker News was relatively limited in scale (82 points, around 10 comments), it presents a concrete and pointed example of the ethical gray zones in generative AI creativity: the more capable the model, the more likely it is to inadvertently reproduce sensitive elements from its training data. Whether viewed through the lens of creator rights protection or AI provider compliance, the unauthorized reproduction of signatures is a warning signal that cannot be ignored.

(Note: This article is based on a single-source report. Specific technical details and any official response have yet to be independently verified.)