A test with three models suggests that drawing in SVG avoids errors in science diagrams

In the group, members compared the results of asking Claude Fable 5.2, an OpenAI model and a Google model for a teaching diagram of an animal cell in SVG format, the vector drawing format. According to the person who ran the test, all three delivered a diagram with the correct labels; Claude's was done in a few minutes, ChatGPT's shortly afterwards, and Gemini took close to half an hour.

The explanation offered is that, in a vector drawing, the language model itself writes the drawing and draws on everything it knows about the subject. Bitmap image generators, by contrast, receive a description and pass it to another model with less knowledge, and that is where the mistakes appear. This is the author's hypothesis; they say they are convinced of it but have yet to test it with other models, so it is not an established result.

They preferred Claude's sober, clear style, while the other two seemed worse as illustrations: in one, elements overlapped and it was impossible to tell where the arrows went. In the discussion that followed, someone pointed out that the hard part today is knowing which tool to choose for each task, and another member said they already ask AI to make the images in their presentations SVG. It is the line the diary already reported on creating infographics in HTML and SVG.

For the classroom, the practical rule that emerges is to ask for SVG where accuracy matters, and to check the diagrams anyway before using them.