For text-generation models like GPT-3, I think it's very optimistic to think students will content themselves with only generating an essay's outline, rather than most or all of its contents.
I fully expect these kinds of models to completely obsolete essays of the format "write 500 words on X topic", at least without further restrictions such as sourcing, which I worry is not a high bar to clear, since copy and pasting quotes from sources and letting the model write your analysis seems quite doable.
Anti-plagiarism software seems to have mostly evolved around detecting when students directly copy and paste each other's work or work from the internet, but AI generated text will likely be very, very hard to distinguish from standard student work after a single editing pass - besides possibly being of higher quality than average student work. By these models changing the labor of writing from being generative to an act of editing, I also fear the educational value of learning to produce one's own work will suffer.
For image generation models like DALL-E and Stable Diffusion, I think these generative models are likely to have a dampening effect on new artists pursuing a major or career in the field, particularly in subfields of graphic design, concept art, and character design, as the ability to generate dozens of professional-grade human-artist competitive images in these spaces instantly by quickly modifying a text prompt, is very likely to reduce jobs in the space - and in a field where it's already perceived as quite challenging to make a living.