Users have discovered that GPT Image 2, OpenAI's newest image generation model released April 21, occasionally produces images containing visible Gemini branding, an artifact that is not evidence of a technical integration between the two companies but rather a symptom of a much larger and more consequential problem: the web is now saturated with AI-generated content, and the models training on that web are learning its artifacts along with everything else.
The Reddit thread that broke the story on April 24 is worth reading carefully. A parent described prompting GPT Image 2 with random nonsense to entertain their children. One of the resulting images contained visible Gemini branding embedded in the texture, despite neither Gemini nor any Google product appearing in the prompt. The comment thread that followed is, unusually for Reddit, more technically precise than most news coverage: "The watermark leak isn't evidence of distillation," wrote one user. "It's evidence of the training set. Web image data in 2025 is saturated with Gemini-generated images." That observation cuts to the actual issue.
GPT Image 2, which OpenAI shipped April 21 as a significant upgrade over its predecessor, was trained on image data that includes the open web. The open web in 2025 and 2026 contains an enormous and growing volume of AI-generated images, many of them produced by Gemini models. Google's Gemini image models apply SynthID invisible watermarks to every output, a cryptographic signal imperceptible to humans that survives cropping, compression, and format conversion. But the text-based Gemini branding that appeared in viral screenshots is different: it is a visible artifact, likely from images generated during early Gemini access periods when some outputs carried more prominent labeling, or from images that were processed and re-processed in ways that surfaced watermark artifacts as visible texture. When a model trains on enough of these images, it learns that Gemini-style text appearing in images is a normal visual pattern. In the right generation conditions, that learned pattern resurfaces in outputs.
OpenAI's situation here is not unique. Every major image generation model training on web data in 2025 and 2026 is training on a corpus that contains an unknown but substantial percentage of AI-generated images. Midjourney outputs, Stable Diffusion generations, DALL-E images, Imagen outputs, and Gemini generations have been uploaded to Pinterest, Instagram, Reddit, Flickr, stock photo sites, and personal websites at a rate that makes filtering them out of web crawl data an unsolved engineering problem. A December 2025 research paper on loss landscape degeneracy in neural networks noted that training on iteratively generated synthetic data introduces subtle degradation patterns that compound over successive training generations. The visible Gemini watermark in GPT Image 2 outputs is the most legible version of a problem that usually manifests as less visible artifacts.
The announcement from X user Anish Moonka, which spread widely on April 6 before the Reddit thread amplified it further, put it memorably: "OpenAI trained on so many Gemini-generated images that the model thinks a watermark is just what pictures look like." That framing is reductive but directionally accurate. The model has not confused itself into thinking it is Gemini. It has absorbed a visual pattern that appears frequently enough in its training distribution that it occasionally surfaces in generation. The mechanism is the same one that causes early diffusion models to generate extra fingers: the training data contains hands with six fingers often enough that the model's internal representation of hands includes that artifact as a low-probability but non-zero output.
What This Means for AI Provenance
Google's SynthID system represents the most technically sophisticated public approach to AI content provenance currently deployed at scale. The invisible watermark embeds a signal that survives substantial image manipulation and can be detected by Google's verification tools even after format conversion and re-compression. The Coalition for Content Provenance and Authenticity standard, backed by Adobe, Microsoft, and a growing number of publishers, takes a different approach: attaching cryptographic metadata to images at creation time that records origin and edit history. Both systems face the same structural challenge, which is that neither works retroactively on the billions of AI-generated images already distributed across the web without provenance metadata.
For the AI industry, the GPT Image 2 watermark incident is an early warning of a compounding problem. As AI-generated content becomes an increasing share of the training data for the next generation of models, the artifacts, biases, and stylistic conventions of current models get baked into their successors. A model trained on web data in 2026 is partially training on the outputs of models from 2024. A model trained in 2028 will partially train on the outputs of models from 2026. The synthetic data contamination is not a bug that gets patched. It is a structural feature of how web-scraped training data works in an era where AI image generation is ubiquitous. OpenAI has not commented on the specific cause of the Gemini watermark artifacts. The likely explanation, a training corpus saturated with Gemini-branded images absorbed without filtering, is both more mundane and more consequential than the distillation conspiracy theory the viral posts suggested. The mundane version does not make for as compelling a headline. It does, however, point to a problem the industry will have to address before the next generation of training runs begins.
Also read: DeepSeek V4 Pro costs 15x more to run than V3.2 and the efficiency narrative just got complicated • Researchers say deep learning can have a rigorous scientific theory and the math to prove it is already emerging • A microwave clock and a Reddit thread just taught millions of people how AI image editing works
