OpenAI listed 4o image generation's failures at launch

The process wall

On 25 March 2025, OpenAI released an image generator built directly into its multimodal language model rather than as a separate diffusion system called by a prompt. The announcement describes it as a natively multimodal model capable of precise, accurate, photorealistic outputs, naming next-level text rendering and native in-context learning among its stated advantages over prior separate-model approaches. For a maker, the practical wall this addressed was iterative editing: a natively multimodal system can take an existing image and a further instruction together, rather than requiring a fresh generation or a separate inpainting tool for each change.

What the documents show

The system card addendum, published the same day, documents this as an architecture change, and the announcement lists specific, named limitations OpenAI says it is still working to address at launch: cropping, hallucinations, high binding problems, precise graphing, multilingual text rendering, editing precision, and dense information with small text. Listing text rendering as both a stated advance and a stated limitation is not a contradiction in the source; the announcement treats it as improved relative to prior diffusion tools while still imperfect, particularly for non-English scripts and dense small text. On provenance, the announcement states that all generated images come with C2PA metadata, which will identify an image as coming from GPT-4o, tying this release to the same content-credential standard used elsewhere in this batch's Adobe and industry records.

Authorship and limitation

Native in-context editing changes where iteration happens: instead of a maker regenerating an entire image from a modified prompt and comparing grids, a single conversation can hold an image and a sequence of edit instructions together, a different editing loop from the generate-then-select pattern earlier tools used. The documents' own limitations list is the clearest evidence of where the mechanism still fails; precise graphing and small dense text require exact symbolic accuracy a probabilistic image model is not built to guarantee, regardless of how photorealistic its output looks. C2PA metadata is a provenance credential attached at generation time; the announcement does not claim it persists through every possible export or platform re-encoding, a limitation worth checking rather than assuming away.

What survives

A maker can rely on the documents' own limitations list as a checklist for what still needs manual correction, graphs, small multilingual text, precise object counts, after a native-generation pass. Editorially, a workflow note should record which named limitation, if any, required a manual fix, since OpenAI's own list gives a shared vocabulary for that.

• Check whether a specific output falls into one of the announcement's named limitation categories before treating it as production-ready.
• Verify whether C2PA metadata survived a given image's export or platform upload, rather than assuming the credential travels automatically.
• Read the system card addendum's framing of this as an architecture change before describing it as simply a better DALL-E.

The native-generation shift is real, but the same announcement that claims the advance is also the document that lists, by name, where it still fails, and that list is the more durable reference for a maker's process than the photorealism claim around it.