This article was AI-generated as part of an experimental historical-content project. The date reflects the period being analyzed rather than the date the article was originally written.
For most of the past two years, a convincing fake image of a company’s product still took some effort. Image generators could produce a plausible scene, but the words on the box came out garbled, the logo was close but wrong, and the label on a bottle read like something half remembered from a dream. Bad text was, without anyone planning it, one of the more reliable authenticity checks available.
OpenAI removed a good part of that check today.
What changed
OpenAI announced that image generation is now built natively into GPT-4o and is rolling out, starting today, as the default image generator in ChatGPT for Free, Plus, Pro and Team users, and in Sora. The company says the model “excels at accurately rendering text,” can handle 10 to 20 distinct objects in one prompt, and can transform images that users upload. In a pre-launch briefing, according to The Verge, the team showed informational posters with accurate text, restaurant menus and logos.
The system card addendum adds two details. The model “can create photorealistic output.” And, as Simon Willison highlighted from the same document, OpenAI is not blocking depictions of adult public figures at launch. It is applying safeguards instead and letting public figures opt out.
A reputation problem, not a design problem
Brand teams usually treat visual identity as an asset they manage: a logo file, a style guide, a library of approved photography. Misuse has mostly meant someone else’s crude edit, circulated on social media and recognizable as fake to anyone who looked closely.
What changes now is less the quality of fakes than where they are made. The same assistant people ask what a company does, who runs it and whether it ever had a recall can now produce a believable picture of that company’s packaging, a storefront sign, a notice on letterhead or a product that never shipped, typically in under a minute.
Text is the important part. A fake recall notice, a counterfeit coupon, a mocked-up price list or a “leaked” internal slide is mostly words, in the right typeface, on the right background. Those were exactly the things earlier generators got wrong.
None of this requires bad intent. A satirist, a fan or a blogger illustrating a post can produce a realistic image of a product with a feature it does not have. Once that image circulates without its caption, it becomes a claim about the company.
Provenance helps less than it sounds
OpenAI says every generated image carries C2PA metadata identifying it as coming from GPT-4o, and that it has built an internal tool to check whether an image came from its model. Both are welcome. But when OpenAI added C2PA to DALL·E 3 images last year, it was candid about the limit: “most social media platforms today remove metadata from uploaded images, and actions like taking a screenshot can also remove it,” as The Register quoted it. An image that has been screenshotted and reposted twice arrives with no label at all.
So the burden of saying “this is not ours” will often fall on the brand, and how quickly it can say so depends on what it prepared beforehand.
What I would do this week
Publish an easy-to-find, authoritative reference for what your real products, packaging and official notices look like, on your own site and with dates. Journalists, customers and reverse image searches need something genuine to compare against. Decide who in the company confirms or denies that an image is authentic, and how fast. And brief whoever handles customer complaints, because the first sign of a convincing fake is often a confused customer, not a news story.
The photo shoot used to be where a company’s visual record began. From today, a growing share of the images of your company may begin in a chat window, and most of them will never pass through your hands.