One relatively straightforward countermeasure would be to feed your text through multiple rounds of machine translation using DeepL, Google Translate, and other similar services and applying corrections when necessary. That should break any personally identifying features in your original text.

That of course has its own risks, if you assume full government access to providers and unlimited surveillance capabilities. If you are fully paranoid you could even run your text through a local instance of (less accurate) Apertium[0] for the first few rounds, or even for the whole process if you find the result is different enough from your original.

[0] https://www.apertium.org/

Wouldn't surprise me if, eventually, it turned out that things like Grammarly were being used to train it in the first place.

I can think of a few.

An offensive one would be to post fake offers to trade government/military intelligence on the dark web writen in the style of the politicians backing up this measure, so that they are put on the feds list and thoroughly investigated.

Yes, there's one already, it just sucks and can't be used from the command-line because the authors were academics. It's called anonymouth.