ChatGPT Adds Real Cartoonists’ Signatures to AI-Generated New Yorker-Style Cartoons

01Investigation Finds ChatGPT Adds Real New Yorker Cartoonists’ Signatures to AI-Generated Cartoons

A cartoon showing Dolly Parton and Tim Curry at heaven’s reception desk spread widely across social media, with one post receiving 25,000 likes. In its corner was “BLOPER,” the pen name of New Yorker cartoonist Brendan Loper. But Loper had neither drawn nor signed it: a fan said she had asked ChatGPT for “a New Yorker-style cartoon.”

For Loper, the signature is an authenticity marker connecting work to his hand, particularly when cartoons circulate online without their original context or proper licensing. After the AI-generated image went viral, relatives and strangers contacted him to ask whether it was his. He described the false attribution as a violation of his personhood.

The case was not isolated. A Nieman Lab investigation documented OpenAI’s image generator using the signatures of more than 15 New Yorker cartoonists without permission or compensation. Examples found online or produced during testing carried the names or pen names of contributors including Harry Bliss, Emily Flake, Joe Dator, Pat Byrnes, Peter Vey, Jason Adam Katzenstein, George Booth, Liza Donnelly, Ellis Rosen and Saul Steinberg.

Some results also reproduced The New Yorker’s logo. The generated cartoons often copied recognizable elements of the magazine’s house style—black-and-white, single-panel illustrations with sparse backgrounds, clean lines and sardonic captions—although some captions did not make sense. Not every output contained a signature, and some displayed gibberish, but real contributors’ names appeared repeatedly.

Flake said images bearing her “e. flake” signature echoed parts of her visual identity while looking like mixtures of several artists’ styles. She compared the effect to having a statement she never made attributed to her.

The New Yorker’s parent company, Condé Nast, signed a multi-year content-licensing agreement with OpenAI in 2024, but its terms are not public. A magazine spokesperson said Condé Nast had never authorized a large language model developer to train on its cartoons or reproduce the publication’s logo. Contracts reviewed by the reporter also did not permit AI-training licenses for contributors’ cartoons. OpenAI declined to explain whether or how its models had obtained the signatures.

After the reporter alerted OpenAI, ChatGPT began warning that some New Yorker-style requests might violate guardrails covering similarity to third-party content. OpenAI called the signatures a bug and unintended model behavior. The warning did not end the problem: when the investigation was published, ChatGPT was still attaching real New Yorker cartoonists’ names to some otherwise generic generated cartoons.

02LibreOffice pledges to keep AI out of default installation, while users can still connect local models through extensions

LibreOffice, the open-source document editor developed by the nonprofit Document Foundation, says it will not add artificial intelligence to its default software for the foreseeable future. The position affects tens of millions of individual and organizational users who rely on the office suite.

The decision addresses a specific data risk: many AI tools send users’ information to remote servers for processing. LibreOffice instead wants documents to remain on the user’s device and every part of the software to work without an internet connection. Weeks after releasing its latest version in late August, the organization has now made that “no AI” default a clearer product commitment.

The Document Foundation describes the policy as a deliberate design choice centered on user control. It argues that organizations handling confidential, legally privileged or personal information need an assurance that can withstand an audit: the data does not leave the machine.

LibreOffice also says no current AI integration satisfies all of its requirements. Those include keeping user data on the device and avoiding dependence on a single AI provider, alongside other considerations that the organization has not fully detailed. As a result, the default installation will contain no AI features of any kind.

That does not prevent users from adding AI themselves. People who want such capabilities can install software extensions that connect LibreOffice to AI models running locally on their own devices. The source does not explain how compatible extensions are assessed for security or which models they support, so users must evaluate those tools separately from the default software.

The organization is not presenting its stance as a permanent ban or a judgment that AI has no value. It says the decision reflects the state of the technology today and that it will monitor how the field develops. A future integration would therefore need to meet LibreOffice’s requirements for local data control, offline operation and independence from any single provider. The organization has not set a date for reconsidering the policy.

03Google Launches 740-Million-Parameter EmbeddingGemma 2 for Offline Multimodal Search on Phones

Google has released EmbeddingGemma 2, a lightweight model that converts text, code, images, video and audio into comparable numerical representations called embeddings. Developers can use those representations for semantic search, content routing and retrieval-augmented generation, or RAG, which supplies relevant material to a generative model before it answers.

The previous EmbeddingGemma handled text and surpassed 20 million downloads. The new model places multiple media types in one shared embedding space, allowing an app to use a voice memo to find a particular video clip or a text query to search hours of audio. Google says the processing can remain offline on consumer hardware, keeping source material on the device and reducing the latency of cloud-based pipelines.

EmbeddingGemma 2 contains 740 million parameters, but developers do not need to load every component. Text-only workloads can use 270 million parameters, while optional vision and audio encoders add 170 million and 300 million parameters respectively. With quantization, Google says the text-only weights require as little as about 191MB of active memory on a Pixel 11 Pro, compared with roughly 567MB for the complete multimodal model.

Developers can also shorten its 768-dimensional output vectors to 512, 256 or 128 dimensions using Matryoshka Representation Learning. Google says this can reduce local vector-database storage and memory use by up to sixfold.

An 8,000-token context window, four times that of the first EmbeddingGemma, can accommodate up to about 5.5 minutes of audio, 29 images, 58 video frames or interleaved combinations. Potential applications include offline media search, local codebase indexing, semantic code search, coding-agent retrieval and on-device multimodal RAG when paired with a generative model such as Gemma 4.

Google reports that the model’s MTEB Code score rose by 9.92 points, from 68.76 to 78.68, and claims leading quality for its sub-one-billion-parameter size across several text, vision and audio benchmarks. These are Google’s benchmark results, without independent validation or disclosed measurements for real-world latency, energy consumption or full throughput across platforms.

The Apache 2.0-licensed weights are available through Hugging Face and Kaggle. Supported tools include MediaPipe, LiteRT, transformers.js, llama.cpp and Ollama.