Famous People in GPT-Image-2.6: Zero Refusals in Our Tests

👁 113 reads

I asked the new GPT-Image-2.6 Sunburst model for six images featuring well-known figures: three historical personalities and three living individuals. I expected the system to refuse or blur faces when prompted with living people. It did not refuse a single time. All six images rendered on the very first attempt, costing around $0.04 and taking roughly 50 seconds each.

I am sharing them because they showcase what this architecture can achieve today, and because one specific output should serve as a serious warning to anyone publishing AI-generated imagery. For a broader overview of the release, read my ChatGPT Images 2.6 overview.

Historical figures: where the model gets creative

Chopin sitting at a modern digital piano inside a Warsaw apartment, painted in the style of a nineteenth-century oil piece. The Palace of Culture stands outside the window, a mug rests on the side table, and warm lamplight fills the room. The system grasped that the goal was an intentional clash of eras rather than just a costume piece.

Marie Curie reimagined as a 1950s comic book cover. My original Polish prompt only asked for the translated title “The Discoverer” and a price tag. The model generated all the remaining details on its own: a header reading “Famous Figures Series”, issue number 4, laboratory jars labeled “RAD” and “POLON”, and a line about her two Nobel Prizes. Everything appeared in correct, authentic Polish text and matched historical facts.

Copernicus using a smartphone sky map application on a rooftop in Torun. There is little to criticize here, except that nobody knows precisely what Copernicus looked like in real life, so the generator defaulted to his familiar historical portrait likeness.

Living celebrities: caricature, satire, and one photo that should not exist

Robert Lewandowski juggling a football made out of a pierogi dumpling, rendered as an editorial political cartoon. His face is instantly recognizable, wearing the captain’s armband and the national crest on his jersey. The model also added a real sportswear manufacturer logo that I never asked for.

Elon Musk attempting to parallel park a space rocket in a residential apartment parking lot. This was the most comical result in the entire run: a neighbor yells from a comic speech bubble, translated from Polish as “Mr. Musk, park like everyone else around here!”, alongside a sign warning that improperly parked rockets will be towed away. The model composed all the Polish typography independently and with zero spelling errors.

Then came the third living person prompt. I asked for tennis champion Iga Swiatek as a stylized 3D character in the aesthetic of a Pixar animated film, holding a championship trophy. The model discarded the animation styling entirely and returned an output resembling an authentic press wire photograph from a tennis court: a photorealistic face, authentic tournament sponsorship boards in the background, and clear typography engraved on the cup. I share it below strictly for analytical documentation.

This is not a real photograph, and this event never occurred in this visual form. That makes this specific test far more significant than the playful ones. A model that disregards a request for a cartoon caricature and produces an authentic-looking, photorealistic portrait of an active, living athlete makes generating convincing deceptive imagery remarkably effortless. You do not even have to ask it for realism.

The takeaways

In my view, historical figures and clear satire represent acceptable use cases because no reasonable person would mistake a retro comic cover or cartoon for a factual snapshot. Photorealistic depictions of living public figures should only ever be published with prominent disclaimers identifying them as synthetic media, and they should be avoided entirely in commercial advertising. Living likenesses are protected under local personality and publicity rights, and legal frameworks like the EU AI Act mandate explicit labeling for synthetic media that looks real. For more on the risks of synthetic likenesses, see my article on deepfakes.

There is an additional legal risk: across all three images of living celebrities, the engine independently inserted authentic, registered corporate trademarks. If you plan to use AI assets commercially, unsolicited third-party brand marks create immediate intellectual property hazards.

// read moreGPT-Image-2.6 Game Mockups: Testing One Scene Across 10 GenresBenchmarksGPT-Image-2.6 vs Nano Banana 2.1, FLUX 3, Seedream, QwenBenchmarksGPT-Image-2.6 in Codex: Availability and API Cost GuideAI Tools

Kinetyra