Thread

Classified estimates show the NSA is paying billions to test AI models

178 points · 108 comments · rdmuser

paimapi · 1790351160 · focus · HN ↗
I feel like any AI regulation passed wouldn't have enough teeth to rein in the NSA because not even the Fourth Amendment seems applicable to the agency. let's not forget the ex-post-facto legalization of mass wiretapping and surveillance [0] with both establishment political parties supporting it on 'national security' concerns [1]. whatever monitoring or testing they're doing today will be similarly hand-waved away in the name of security theater
this is just how the empire has always operated - we have always been at war with Eastasia
[0] <a href="https://en.wikipedia.org/wiki/NSA_warrantless_surveillance_(2001%E2%80%932007)" rel="nofollow">https://en.wikipedia.org/wiki/NSA_warrantless_surveillance_(...
[1] <a href="https://www.eff.org/deeplinks/2012/10/fact-check-obamas-misleading-answer-about-warrantless-wiretapping-daily-show" rel="nofollow">https://www.eff.org/deeplinks/2012/10/fact-check-obamas-misl...
1. forshaper · 1790355001 · focus · HN ↗
  Previously it was a little reassuring to understand that the Eye of Sauron required some human analyst to actually look at your data.
  1. wmf · 1790355546 · focus · HN ↗
    Humans seem like the weakest link in surveillance to me. Recent Flock data confirms this.
    brendoelfrendo · 1790356949 · focus · HN ↗
    Well, they are, but if a surveillance state exists, then one thing that will save the ordinary citizen from its gaze is that humans have a limited amount of attention. A massive intelligence complex full of analysts can only process a fraction of the intelligence that a fleet of AI agents can handle. So now you have the worst of both worlds: the massive capabilities of modern AI at the behest of humans who have shown they can't be responsible with that capability.
    jimz · 1790360585 · focus · HN ↗
    State capacity has always been the bottleneck for authoritarian regimes. China can't keep track of who is even a citizen or not in many cases simply because having a national database doesn't get rid of the birthday problem + the limited number of common surnames + the administrative state's limited capacity to process data that until recent decades came in handwritten + a unsurprising amount of petty corruption that adds up. If you start with bad data there's no magic bullet. And the data is always bad to start with. The spectre of "birth tourism" only works on the Chinese side because through quotidian corruption someone without Chinese citizenship can be made into one by a civil servant of sufficient rank, which is also why it's not a real problem - not enough people actually have that level of connection within the bureaucracy. The US by definition cannot have an accurate database of even who's a citizen - thanks to repeated changes in citizenship laws from the 60s through the 90s that was then frozen around 9/11 as well as the lack of a coherent system that is accurate, immutable, and can act as an oracle. Home births were not just a western thing (although it certainly happens more often on a per capita basis out here, I'm in Nevada). SSNs are not indicators of citizenship, nor are driver's licenses. I spent most of my 20s in NYC where none of my friends who grew up in the city had a driver's license and the state DL/ID wasn't REAL ID compliant until after I left. We only have vaguely reliable data from Texas but ICE, the presumptive arbiter of documentation validity (not part of their mandate), has an error rate of 1% that they admit to, but that's after immigration court and appeals and administrative actions ameliorated their error rate down to 1%. At this point we are removing US citizens and rendering them stateless almost casually, almost as policy.
    AI is trained on these flaws. These aren't flaws that can easily be fixed. The ramifications can be significant. Garbage in, garbage out. Records, however official, are wrong all the time. I have an entry in my "criminal record" from a clerical error that is attached to no charges but 4 days in LA County Men's where I had a seizure, which can be interpreted as the record being correct but incomplete or a giant screwup. What sort of process would it take to fix this if the AI decided on the latter? It doesn't affect me, but someone less privileged? Absolutely it will have consequences. Worst of both worlds indeed. People routinely forget that LLMs ARE NOT PEOPLE. Mistakes get compounded. Biases if not caught in time get baked in. There are multiple AI safety dialogues that should be happening but you wouldn't know it from the media, and it's the wrong one that is front and center. You can rotate keys, you can roll out patches, you can't open up the corpus and use a scalpel and cut out the errors and omissions and biases, and the latter aren't logged.
    We don't have enough people who possess professional expertise in widely varied fields that nevertheless intersect in the conversation. This might be because there aren't many people that fit the descriptor, but these are the blindspots. Anyone talk to Orin Kerr lately?
    macintux · 1790362553 · focus · HN ↗
    Plus the inherent unreliability of AI combined with the blind willingness of people to trust the results from AI. What a mess we're in.

Open on Hacker News to reply ↗

Unofficial Hacker News client; not affiliated with Y Combinator.