I wondered something almost the opposite the other day. With consistent interaction with chat bots and AI output, their linguistic ticks are surely bleeding into every day speech ("smoking gun" and other turns of phrase). What if our linguistic output starts to conform?<p>> Do they, incorrectly, position their adverbial clauses?
Underrated line.
There's been some interesting threads about stylometry over the years [1]. The top link was quite decent at unmasking HN alt accounts with basic ngram analysis whipped up in one day [2].<p>1. <a href="https://hn.algolia.com/?q=stylometry" rel="nofollow">https://hn.algolia.com/?q=stylometry</a><p>2. <a href="https://news.ycombinator.com/item?id=33756141">https://news.ycombinator.com/item?id=33756141</a>
> A stronger conjecture is that we’re heading towards a sort of generalized pseudpocalypse. Perhaps, in the future, if you interact with the world through essentially any high-bandwidth channel, then you identify yourself. Say you wear a mask in public and only speak by sub-vocalizing into a voice changer. That’s fine, you’ll still be identified using your body shape, gait, or chemical signature. Or say you don’t like your car being tracked everywhere, so you stop carrying a phone and you somehow convince lawmakers to ban license plates. No problem, your car will still be tracked using tiny scratches or unique pinging sounds from the engine. Or say you don’t like being tracked on the internet, so you lock down your browser profile, buy stuff only with Monero, and connect through a chain of three VPNs. That’s OK. You’ll still be identified through how you wiggle your finger as you scroll down the page. We’re all just too unique, and the information theoretic limit is coming for us.<p>Forensic research, NSA, Palantir…<p>Btw 42. Sleep, eat, have sex, have fun, be useful.
One would think that it could be possible to make a tool that takes some text and "anonymizes" it by making it a little more standard and boring (uniforming punctuation and sentence structure, changing words with some synonyms, etc). Maybe wouldn't make it particularly compelling, but would be valuable for political dissidents and other people with a high threat model.<p>Does anyone have some tools to share?
Entropy is unfortunately a very bad metric to estimate if these identification techniques will scale. Plugging my work here as example: <a href="https://www.nature.com/articles/s41467-024-55296-6" rel="nofollow">https://www.nature.com/articles/s41467-024-55296-6</a>
So anonymity of written speech is toast. We should, however, strive to preserve other forms of anonymity. For example, donations given to political causes should be kept confidential. Let protesters wear masks up to the point where they break the law.
> So anonymity of written speech is toast.<p>No just use a text-mixer: in goes your text, set parameters to have output match <scapegoat> (or just pick "standard_Neanderthal_3"), out comes text conveying the message you wrote, in the style of your choosing.<p>Of course that would also strip the attributes that made it your creation.
Hmm, the Kill Puppies Movement received $500,000,000 from an anonymous donor.
yes but no, can't we just ask AI to sufficiently shuffle our words or for algos to do so?<p>"boom", pseudoanonymity (spell?) restored?
One potential solution here I suppose is to make deanonymization or contributing to it a serious crime. I doubt that will happen in most places, though, since it’s often the government that wants to do this to its own citizens.
As long as you're fine with losing your distinctive voice (which should be taken as table stakes for people who value anonymity to the extent of worrying about this), it's perfectly reasonable to use tools to stymie stylometry. I'm not even suggesting using an LLM (which may or may not be sufficient, it would be difficult to verify either way); it would suffice to have a tool that rewrites your prose (or analyzes it and flags it for manual correction, etc.) if it isn't written in, say, a form where every sentence is primitive subject-verb-object, limited to the 1,000 most common English words, with no contractions, idioms, or exotic punctuation. Yes, this doesn't completely eliminate all possibly identifying bits of entropy, but it would more than suffice for hiding in a crowd ("hiding" in the sense of obviously standing out as someone trying not to be noticed, and as long as you're also careful about your opsec in other ways, like time of posting, etc).
This always seems theoretical. Has it happened?
[flagged]