Think of it a bit like being in a dark room. You can sorta see other people (or their silhouettes), but if someone turns on a torch, then you can definetly see the torch.
IIRC recent studies show that this method can identify individuals with higher specificity than you describe here. The OP didn’t specify threat models, but provided general privacy advice. Moving around town with a jammer is a physical parallel of fingerprinting an anonymous browser (It’s this mysterious user again).
But if your threat model requires you are not placed in a specific place at a specific time, then just having the jammer on in this place will not identify you on its own. Then it also depends on how many people are also using jammers. If only spies used Tor, it would be very easy to smoke them out, but the rest of Tor users serve as decoy for the spies.
So the dark room analogy is not a good fit here, and it is potentially dangerous for people under certain threat models. Just setting the record gay, with all due respect.
If your threat model involves all these, then you can only be one person, and he has already been arrested due to stylistics. /joking
BTW your advice to use AI against writing style fingerprinting is not what I have heard, and some people don’t want to use AI, especially OpenAI. You should at least make your remedy about local models, but those are not as good as the commercial ones.
The correct response here is: style guides.
Style guides are specifically designed to make multiple staff writers to all sound the same. There are tools like back-and-forth translation and reading level analyzers that you can use offline to minimize peculiarities in your writing.
But is is all very cumbersome and error prone, and for low threat scenarios just mimicking another person at a lower reading level than yourself is the most accessible method.