Diagrams and visuals without a designer

Two different tools wearing one name

“AI images” covers two things that behave nothing alike.

A language model, asked to describe a diagram, produces text: what to draw, where, what to label. It is good at this because it has read thousands of diagram descriptions and textbook figure captions.

An image generator produces a picture from a description by a process closer to sculpting noise into a plausible image. It has learned what pictures of hearts look like. It has not learned anatomy. It renders text as visual pattern, squiggles that look like letters, which is why labels come out as almost-words, and it places parts by what looks right, which is why chambers are mislabelled. The picture is convincing and wrong, and it is wrong in the way that matters most for teaching: confidently.

So the rule is simple. Image generators for illustration. Language models for diagrams, which you then draw.

Diagrams: the path that works

Ask for the diagram as instructions, not as an image.

Describe a diagram of the water cycle suitable for a blackboard. List each element to draw, its position relative to the others, and the label for each. Number the elements in the order I should draw them while explaining.

You get a build order: sea at the bottom left, sun top right, arrow labelled “evaporation” rising from the sea, and so on. Draw it on the board in that order while talking, and you have a diagram that is correct, that students watched being built, and that cost nothing.

For a printed diagram, two free paths:

• Diagram code. Ask for the diagram in Mermaid, a text format for flowcharts and cycles, or as SVG for simple shapes. Paste it into a free online Mermaid editor and you get a clean, editable image with real text labels. Flowcharts, cycles, timelines and tree diagrams work well. Anatomy does not.
• A verified source plus your labels. Wikimedia Commons has accurate, freely licensed diagrams of nearly every scientific structure in a school syllabus. Download one, open it in a free editor (Photopea in the browser, GIMP or Krita on a laptop) and add or translate labels yourself. This is what a diagram in the local language costs: ten minutes and a correct source.

Illustration: where image generation earns its place

Scenes, characters, settings, mood. A market for a reading passage, a girl with a bicycle for a maths problem, a village at dusk for a poem. Here the model’s strengths apply: it produces a plausible, attractive picture, and plausible is exactly what an illustration needs to be.

Free options change often; at the time of writing, Bing Image Creator, the image mode in Gemini and ChatGPT, and several open models on Hugging Face are usable at no cost. Prompt with the setting and the people described plainly (“a girl of about ten in school uniform buying tomatoes from an older woman at a market stall, morning light, East African town”) and expect to generate three or four before one is right.

Two predictable failures: hands, which come out with the wrong number of fingers, and any text in the image, which comes out as scribble. Crop the hands; leave the text out and add it in the editor.

Representation

Unprompted, image generators draw a default person: light-skinned, Western clothing, Western setting. Every one of your students will notice if the maths problem about their market shows a supermarket in Ohio. Specify the people and the place, every time, and check the result for the version of your country that a tourist brochure would draw. Ask again if you get it.

Photographs of real things

For anything where accuracy matters and a real photograph exists (a plant, a rock type, a historical building, a piece of equipment) use the photograph. Wikimedia Commons, Unsplash and Pexels are free and searchable, and a real photograph of a real baobab teaches more than a generated approximation of one. Generation is for what does not exist or cannot be photographed.

The copyright question

Whether a generated image can be copyrighted, whether the model’s training on artists’ work was lawful, and whether using generated images in teaching materials creates any risk, are all being argued in courts and legislatures in several countries at once, with different answers emerging in the US, the EU, the UK, India and Japan. Nothing here is settled. For material used inside your own classroom the practical risk is low. For material you sell or publish, know that the rules differ by country and are changing, and do not treat any single article, including this one, as advice.

The honest limit

The diagram path works because you draw it. A teacher who cannot draw a passable heart on the board is not helped by a description of one, and the verified-source path is the fallback for exactly that case. And image generators improve at text and hands with every release; the failures described here are shrinking, but the underlying fact, a picture made to look plausible rather than to be accurate, does not change with resolution.