It's a cat-and-mouse game for sure. At the first level, any dongle that simply types the AI response through a fake HID device will be easy to detect. No real essay writer just types an entire document in one go, with no edits. They move paragraphs around, expand some, delete others, etc.

So this dongle will have to convincingly start with a worse version that's too short (or too long!). It'll have to pipe the GPT output through another process to mangle it, then "un" mangle it like a human would as they revise and update.

If trained on the user's own previous writings, it can convincingly align the AI's response with the voice and tone of the cheater.

Then the spyware will have to do a cryptographic verification of the keyboard ("Students are required to purchase a TI-498 keyboard. $150 at the bookstore") to prevent the dongles. There will be a black market in mod chips for the TI-498 that allow external input into the traces on the keyboard backplane. TI will release a better model that is full of epoxy and a 5G connection that reports tampering...

... Yeah, I also predict more in-person learning :)