Forum
Images That Can Read: Typography Arrives in Generative Design

For most of their short history, image generators have been bad at words. Signs came out as gibberish, posters needed their text added by hand, and garbled lettering was one of the easiest ways to spot a synthetic image. Two launches this month suggest that era is ending.
On 22 April, OpenAI released ChatGPT Images 2.0, powered by a model available to developers as gpt-image-2. OpenAI describes a model that reasons before it draws, planning composition and checking constraints, and that renders text accurately in many languages, including scripts such as Japanese, Korean, Chinese, Hindi and Bengali.
A week earlier, on 16 April, Canva introduced Canva AI 2.0 at its Create event in Los Angeles. It runs on what the company calls the Canva Design Model, a foundation model it says was built to understand the structure and hierarchy of real-world design. Instead of producing a flat picture, it is meant to generate layered, editable layouts from a prompt. Canva calls it its biggest launch since 2013 and is releasing it as a research preview.
Type was the last line
Typography has quietly been the boundary between generated images and designed communication. A picture can be evocative without words; a poster, a package or an interface cannot. As long as models could not handle text, designers kept a clear role at the end of the process: placing the words, choosing the type, making the hierarchy legible.
Models that write accurately inside images and tools that output editable layouts move that boundary. A first draft of a poster, a social post or a slide can now arrive with its headline set and its structure in place.
The two launches come from opposite directions. OpenAI is teaching an image model to handle words and layout; Canva is teaching a layout tool to generate. They meet at the point where a picture becomes a piece of communication, which is exactly where most professional design work happens.
What changes for designers
This is not the end of typographic craft, but it changes where the craft is applied.
- Drafts start further along. The designer’s first task becomes critique: is this hierarchy right, is this typeface appropriate, does the layout serve the message?
- Multilingual work gets easier and riskier. Accurate rendering in many scripts is a real gain for global work. It also means errors may be subtle rather than obvious, and need native readers to catch them.
- Editable output matters more than pretty output. A layered layout can be corrected and built into a system; a flat image cannot.
- Brand consistency becomes the main job. When anyone can generate a decent layout, the value lies in rules that make every layout unmistakably yours.
The trust problem
There is a less comfortable side. Garbled text was a useful tell. When generated images can include convincing documents, labels and screens, viewers lose one of their easiest checks. Photography competitions, newsrooms and platforms will have to rely more on provenance records and less on spotting mistakes.
Designers who work with generated text should hold themselves to a simple rule: never produce something that imitates a real document, interface or brand in a way that could mislead.
Implications for award entries
Juries have often found AI-assisted entries easy to read because the seams showed. That will change. Clear process documentation becomes the most reliable way to understand an entry. The International AI Award Rules page asks entrants to explain which models and versions they used and how they prompted them, along with the techniques and parameters they applied. Entries that show how typography and layout decisions were made, and which ones a person made, will be far easier to judge fairly.
The takeaway
Generative tools are learning the grammar of design: text, hierarchy, layers and structure. That makes them more useful and more consequential. The designer’s role shifts from producing the first layout to judging every layout, setting the rules that keep work coherent and making sure that what reads clearly is also true.



