Ideogram detector

When the words are the giveaway — or aren’t

For years the easiest way to spot an AI image was to read it: shop signs, book covers and T-shirts were full of letter-like squiggles. Ideogram, launched in 2023 by former Google Brain researchers, was among the first generators to render words reliably, and typography has been its focus since. Today several models write well, but Ideogram remains a go-to for designs where the text is the point.

That shift means readable text is no longer evidence that an image is real. A crisp “CLOSED UNTIL FURTHER NOTICE” sign, a menu board or a protest placard can all be generated in seconds, and they are often shared as photos of real places.

What still gives text-heavy AI images away

• Too-perfect lettering in imperfect settings. A hand-painted sign with flawless kerning, or text that stays razor sharp on a crumpled surface.
• Small print breaks down. Headlines are usually right; fine print, URLs, prices and secondary text often turn into near-words.
• Fonts that drift. A letter shape that changes slightly between two occurrences of the same word.
• Styled scenes. Designs placed on a mock-up wall, shirt or storefront with lighting that is a little too even.

Alongside these cues, GPTTrace runs its neural classifier and forensic checks. Graphic designs with large flat areas are harder for classifiers than photographic scenes, so expect more “Inconclusive” results for plain posters and logos and more confident ones for text placed in photographic settings.

Checking a meme or screenshot

Memes and screenshots are often edited by hand in ordinary apps, not generated, and a detector can’t tell whether a human typed false words into a real template. For screenshots of conversations or posts, the best check is the source: search for the original post, or ask the person shown in it. Use the AI check here when the image itself — not just its words — might be invented.

Common uses that make text images matter

Text-rendering generators are behind a growing share of fake evidence: images of shop notices announcing closures or policies that never existed, “leaked” memos photographed on a desk, fake product labels and ingredient lists, invented warning signs at tourist sites, and posters for events that aren’t happening. When the text in an image makes a claim, verify the claim at its source — the shop, the organisation, the venue — as well as checking the image.

Logos are a special case: a business that shows a polished logo, website banner and product mock-ups but no real photos of its premises or staff may be a generated storefront rather than a real company.