ChatGPT 4o’s image generation excels but bans readable text in images

ChatGPT 4o’s new image generation capabilities bring impressive quality but frustrating text restrictions, creating a contradictory user experience. While OpenAI claims the system excels at text-forward images like business cards and instructional posters, its content policies paradoxically prevent users from creating images with readable text on physical objects—a limitation that undermines one of its most promising technical achievements.

The big picture: OpenAI has integrated native image generation directly into ChatGPT 4o, replacing its previous reliance on DALL-E and delivering higher quality results with significant policy limitations.

Key capabilities: The new image generation system produces exceptionally detailed images, though they take longer to generate than with previous iterations.

• The quality improvement is evident in examples like a winter robin image highlighted in the article.
• ChatGPT 4o can perform sophisticated image edits including background changes, subject swaps, and mood alterations.

Ethical guardrails: ChatGPT appropriately refuses to remove watermarks from uploaded images, showing improvement over Google’s Gemini in respecting copyright protections.

The contradiction: Despite OpenAI marketing the system as excelling at “text-forward” images like instructions posters and business cards, ChatGPT 4o refuses to generate images containing readable text on objects.

• When asked to create images of poetry on gravestones, stone tablets, signposts, or even coffee mugs, ChatGPT consistently refused, citing “content guidelines.”
• The system claims it cannot generate “realistic, readable text on a physical object” or even “lengthy, realistic-looking text within images” regardless of context.

Why this matters: The restriction creates a frustrating user experience where the technology’s actual capabilities are artificially limited, preventing practical applications.

• DALL-E will attempt to create similar text-based images but produces illegible results, highlighting the technical advancement that’s being restricted in ChatGPT 4o.
• The policy creates a situation where the AI can technically produce high-quality text in images but is programmed to refuse most practical use cases for this capability.