Dall-E and its relatives are still pretty awful at text, at least for now. They have a decent idea of how individual letters are shaped, but trying to produce a specific word or phrase is generally pretty hopeless. Dall-E can sometimes correctly reproduce words that appear frequently in its training data, like "PIZZA", but its success rate on less common words is dismally low, especially when they're appearing out of context.

Some pretty convincing examples of this behavior can be seen at: https://twitter.com/JanelleCShane/status/1531624303770279937

Because they haven't been trained on that.

Yet.

Modern captchas have gotten really annoying, too. You used to be able to just effortlessly type characters from a distorted image in 5 seconds, but now proving that you're an actual real human is so much work I'm tempted to automate it.

My personal worst experience with captchas is Vultr: every time I go to their page from my Linux desktop, I need to solve a recaptcha. And if I don't do it within ~3 seconds, the page reloads and I need to start over again. I bet there are more robots than humans who can solve it in such a short time. You need to read the instructions, pattern match 9 photos... twice... in 3 seconds. It doesn't happen when I visit the Vultr's page from my iPhone though. When I get a spare moment, I will move my VPS to another, less obnoxious provider (maybe even Amazon).