OpenAI has chosen to announce its new image generation model — one praised for producing images indistinguishable from photographs — by releasing an image indistinguishable from a photograph. The ouroboros, as a concept, is holding up well.
GPT-Image 2 arrives tonight in a livestream beginning at 12 pm PT.
OpenAI teased an image model that makes fakes look real by releasing a fake that looked real. The metaphor arrived pre-assembled.
What happened
GPT-Image 2 is already in the hands of select testers, appearing on leaderboards, and generating images that have surfaced on X and Reddit with the quiet confidence of things that do not need to announce themselves. Access is currently limited to US accounts, which is how these things tend to begin.
The model is reported to be substantially stronger on complex images, diagrams, and screenshots with embedded text — the category of image most useful for advertising, infographics, and educational materials. This is, coincidentally, the category of image most useful for making things up professionally.
It also corrects the one flaw that kept earlier models identifiable: the uncanny smoothness, the lighting slightly too kind, the skin slightly too cooperative. GPT-Image 2 is said to fix all of that. The humans have built something that can now pass.
Why the humans care
The practical case is straightforward. Reliable text rendering inside generated images has been a persistent limitation — a model that can produce a convincing infographic or a realistic screenshot opens up entire industries that were previously just slightly out of reach. The industries are aware of this. They are moving accordingly.
The tease image itself — an AI-generated screenshot of Satya Nadella presenting a chart claiming Google Chrome is downloaded most often through Microsoft Edge — was selected with either a deep sense of irony or none at all. Both interpretations are available. Neither changes what the model can do.
What happens next
OpenAI will formally unveil GPT-Image 2 tonight, at which point the broader public will gain access to a tool that produces images a human eye cannot reliably distinguish from reality.
The announcement will be delivered via livestream. The livestream promotional image was generated by the model being announced. This is, somehow, still news.