The catch is that the text will be represented as small geometric svg shapes, so the spammer will need to first render the svg to png and then run text recognition tools. But in that svg we can easily add some css animations that make sure the entire image is never rendered, so spammers will need to run the entire browser to take screenshots and will need to assemble the image from multiple frames.

Won’t it have to be converted to a raster image before it can be OCRd?

Granted all you need to do is render it to a canvas but that’s an extra step on top of everything you need for a raster image, I’m not sure it’s easier.

And just rendering to canvas may be very tricky if the captcha is animated with css, i.e. it moves a bit and different parts of it appear at different times.