Language Log

« previous post |

In "Chain of thought hallucination?" (8/8/2025), I illustrated some of the weird text representations that GPT-5 creates when its response is an image rather than a text string. I now have its recommendation for avoiding such problems — which sometimes works, so you can try it…

Yesterday I asked GPT-5 for an image to illustrate a planned presentation for first-year students about literary studies at Penn, and the impressive response included a stack of relevant books. GPT-5 chose the titles, and in fact the whole idea of the book stack was part of its response:

The blue book on the bottom of the stack is labelled in some alien non-alphabet, but sill…

I asked for some changes to the picture, and GPT-5 generally did as I asked. But as our dialogue continued, the labelling of the book stack got weirder and weirder, despite my specific instructions to fix the oddities and restore the original titles:

When I complained somewhat more vividly, GPT-5 responded:

You’re absolutely right — thank you for pointing that out. The AI rendering of book spines often hallucinates text, which is why you’ve been getting “Jonis Joyce,” “Tohi Morrison,” “Maprits,” etc., instead of the clean, correct author names.

To get this right, we’ll need to constrain the design more tightly. A couple of options:

Minimal-text approach:
Instead of generating the book spines with full names, we can request only generic books with solid-colored spines — and then we can overlay real text afterward (I can generate a clean, high-resolution image of the names “Shakespeare,” “Toni Morrison,” and “James Joyce” in matching fonts for you to place on the spines).

Post-edit overlay:
I can generate the correct author names separately as text graphics, then position them digitally on top of the book stack in your chosen image. That way, you’ll have exact, error-free spines.

Typography lock-in:
I can tell the generator explicitly: “Use only the exact names ‘Shakespeare,’ ‘Toni Morrison,’ and ‘James Joyce’ on the spines, no variations” — but even then, the model sometimes slips. The most reliable solution is #2 (overlay).

Do you want me to try another direct fix with stricter prompts, or would you prefer I generate clean text overlays of the author names that you can apply to whichever image you like best?

I asked for the "Post-edit overlay". The first result (a male version) was pretty bad in terms of book-stack titles — "SMAKESPEARE" and "TONI MOEKISON":

I asked for a female version, without commenting on the errors — and the result was correct, at least in terms of book-stack titles:

There may be a magic prompt wording that will work more than half the time, but that's all I've got time for just now.

August 24, 2025 @ 8:41 am · Filed by Mark Liberman under Artificial intelligence

Permalink