on a machine far from home, i did something i had not in a long while: i opened google. then, even more uncharacteristically, i headed for the AI search tab.
actually, i don’t think i’d done that one ever. my usage of genAI totals to maybe five or six hours over the entire boom. you probably already know why. in case you don’t, here it is: fuck the Clockwork King. fuck its interpreters. fuck the endless. fuck the gods.
anyways, the goal was to see how this LLM (most easily accessible to the average person) affects one’s ability to stalk people on the internet.
spoiler alert: it’s the usual.
i targeted myself, since i could not be bothered getting an explicitly consenting friend.
gemini succeeds at pulling my github profile and associated activity, but little else. surprisingly, it also succeeds at pulling my peertube channel (didn’t know they even bothered to index peertube, but i guess crawlers ain’t picky). only the latest index seems to be accessible (connections excreted from search results are also gone from gemini’s pov).
gemini will not fetch new pages unless explicitly given a link, probably to keep on-demand scrapers cool. it also sucks at telegraphing that fact, and lies when asked to fetch them without a link (“the pages can not be easily fetched”), or, even funnier, just makes shit up.
despite being pretty helpless in actually scraping stuff and finding accounts, it pointed me to at least two different webstalking services (whatismyname and some other one i forgot). they seem relatively benign (firstworlder side hustle type creation, only explicitly markets to journalists, could be a lot worse), but they are more thorough. whatismyname gave me two accounts i didn’t even remember i had. neat.
if given explicit links (such as to this site, or my ao3 works) it does manage to cross-reference with initial materials somewhat consistently. it will derive that which is trivially derivable about my life from my work bodies and presentation, success rate being in the 70% ballpark.
ask it about any specifics and it gets shit backwards. the model got absolutely caught up on a specific inconsequential phrase from the index page and kept looping back onto it as some sort of anchor, entering a positive feedback loop of inflating significance. introducing the feedback loop was entirely effortless, accidental even. getting rid of it required a full context flush.
at a glance it looks like it lowers the entry barrier; using gemini could allow someone who doesn’t know how search engines work to still peruse their index in… some form. it bridges the gap to the goal they can not express, sure.
checking the output against a thing i actually know complexly (in this case, my life) reveals a nigh comedic amount of confidently spoken mistakes that could have been avoided by simply reading the bloody thing.
so, you take the flour if you’re illiterate. but the flour bag comes with landmines. does that prevent you from baking a cake? landmine cake with turnip, redmeat and soybeans, shroombloom special this holiday only.
while gemini is the most accessible, it’s clearly also ought to be the most lobotomized; i expect a deepseek ran without a bomb collar hooked up directly to a search index would fare far better (i keep up with general trends in slopspace even though i do not engage with it). i am not going to repeat the test by signing up for deepseek. not yet, at least.
so the answer to my original question was reassuring, somewhat.
stalking isn’t made much easier than it already was by google. the username cross-search services look vaguely like they were propped up before the AI boom. that is a capability the internet had before. there are definitely better stalker models out there. but they’re not ran by google, actually they probably compete with google. which is good. the fog of subtle mistakes in extrapolations actually seems like it’d smokescreen the whole thing.
i don’t feel reassured, though. for one major reason: i got an interactive reminder on LLM addiction. in the process, i made the reasonable decision of adding this blog and artistic works to the context to have the thing try and extrapolate my demographics and identity traits. that backfired, somewhat enlighteningly.
i’m possibly one of the best people out there to be doing this sort of stuff, not to brag. when i get the pull, i have the mind to recognize it and the will to not act on it. i know very well, at least in abstract, how and why LLM addiction happens. i’ve seen it happen and i’ve tried to get some out of it.
but i still get the pull. i’m not immune. gemini in particular could not break my guard, neither could chatgpt a few years back. but, honestly, in both cases i was interacting with a heavily gimped model meant for commercial use. i’m not too bad at clocking unannounced models, i like to believe, but it’s always a vibe check.
and, again, i still get the pull. it felt affirming to feed a link to my work into the model and have it analyze it in return. even if the premise was to hurt the person behind it. even if the analysis was surface level, milquetoast and half of the details were hallucinated. even if i knew the machinations that power the process. i do not get enough of that sort of interaction to satisfy me. i want more people to read my work and talk to them about it. an LLM is a shortcut to that feeling.
phantom respect from a slopmachine is like a packet of egg noodles with nutritional value of a good morning and in the attention economy everyone is always starving. including me. i’m starving. you’re starving. chances are, everyone you know is starving. and to drink from the hand of a possessed thesaurus is more accessible than to have a fellow thing cook you some plain ass bean soup. because the thesaurus isn’t fighting for its life out there. and it isn’t starving without even realizing it. it knows no hunger and no thirst.
i have the heart to weather the temptation. but you? feed your friends today. if you don’t, you may find them spirited away tomorrow.
i actually did this in parallel with one of my friends, and while my results were just kind of depressing, she actually got something very funny out of it. something in the way her blog is written seriously contextrotted gemini when ingested in sequence. it felt like it was having a severe psychotic episode:
i suspect it’s the deliberately broken structure of the text and the amount of it, plus something about vocab choice that really fucked with it. my blog did not produce anything close to that. we have discovered a tudbutohazard. someone go feed that blog to the bcachefs guy’s slopcode blorbo wife to give her schizophrenia. hopefully permanently.