Short answer: yes. Ask ChatGPT for an SVG and it writes valid XML you can save as .svg. For simple geometric icons it is genuinely good. For anything organic it fails in a predictable way, and knowing why saves you a lot of wasted prompting.
Why the failures all look the same
ChatGPT has no visual feedback loop. It predicts coordinates the way it predicts words, and nothing ever renders the result back to it. So:
- coordinates are guesses, and elements overlap or drift
- bezier control points, the hardest thing to place blind, produce lumpy or self-intersecting curves
- proportions drift on anything drawn "from memory"
- raw
<text>elements land at the wrong size in the wrong place
The debugging loop is worse than the first attempt: every "move it left a bit" is another blind guess. Ten rounds on one shape is a normal experience.
What it IS good for
- 24x24 line icons and geometric shapes: describable without ambiguity
- correct boilerplate:
xmlns,viewBox, sensible structure - mechanical edits on code you paste in: recolor fills, scale strokes, strip groups
- explaining path syntax better than most docs
A workflow that ends with a usable file
- Constrain the prompt: canvas size, stroke width, flat colors, no gradients.
- Save the code block as
.svg, or paste the markup into a vector editor. - Fix visually, not conversationally. Dragging an anchor point beats another round-trip.
- Export.
I use SVG Lab for the visual step because it runs in the browser with no install: paste ChatGPT's code, see the render, drag the points, export SVG, PNG, or PDF.
If you want to skip the loop
SVG Lab has AI generation built into the editor itself: prompt or image in, editable vector paths on the artboard out, live preview before you commit. It is on the paid plans (free accounts get a preview). The practical difference from ChatGPT is not the model, it is the loop: the AI's output lands where you can see it and fix it with real tools.
Division of labor that holds up: ChatGPT for icons, boilerplate, and code edits. An editor with AI built in for artwork you will actually iterate on.
