> The chilling effects of this will be all too real.

Maybe that's the point.

It could also be announced because they plan to use it publicly soon, to "prove" that some person they want to get for political reasons is the same as some evil terrorist/pedophile/serial killer.

> plus the risk of randomly being flagged

This is 100% going to happen, so my guess is it can’t be authoritative. It’s likely to be used to whittle down a list so that humans can review the results.

Feels like the beginning of a hybrid Minority Report + Enemy of the State movie. I’d watch that. I don’t want to live in that world, though.

> > plus the risk of randomly being flagged

> This is 100% going to happen, so my guess is it can’t be authoritative.

Why do you think that it won’t be authoritative?

How many people write similar to you? I’d imagine it is more than one. Someone even did a show HN to find other accounts that write similar to you, which was entertaining.

I didn’t claim it would be correct or accurate. I doubted your allegation that it would not be taken as authoritative.

Won't there be a countermeasure? A grammarly for privacy that takes your intent and outputs an un-anonymizable text?

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.

> If this works

Seems like an impossible task. There's just not enough signal to noise ratio to make that distinction. Not if the goal is to distinguish among thousands of people.

Among tens or hundreds, might work if they don't take any countermeasures.

Even if it doesn't exactly work, like with facial recognition fucking up for non-white skin colors, they will use it anyway and real people will be saddled with real adverse consequences.

Im not smart enough to know if this is the same thing as what you're saying, but it just doesnt seem like there are millions of writing styles. There might just be too many literary doppelgangers.

In the name of national security, the agencies do whatever they like, and what's worse, people buy the narrative