Um what all these stories get wrong is that we don’t need a general AI that can build itself etc.
All our systems are predicated on the inefficiency of an attacker.
Voting. Sybil resistance. Social reputation. Online reputation. And so on.
Deepfakes and ubiquitous cameras mean that in about a decade or two computers will be able to generate believable video of anything. We already have AI that can recognize behavior from video and Palantir-type processing can predict how and when people will try to coordinate. Parallel construction can be used in courtroom cases to put people away.
In short — it’s not that difficult to imagine that, in the next 2 decades, all our systems begin being attacked.
And this is ignoring the fact that drones and self-driving cars will be ubiquitous and easily taken over by sleeper programs which then wake up and coordinate swarms.
Again — no need for some self-replicating GENERALLY intelligent AI. Simply a bunch of people who wield botnets that can bootstrap themselves faster than our current systems can shut them down. They will just have sleeper accounts until it’s time to deploy.
> Parallel construction can be used in courtroom cases to put people away.
This isn't how parallel construction works.
Parallel construction doesn't invent a crime. Parallel construction suggests a direction of investigation that results in a clean chain of evidence to convict of an actual crime that happened.
An illegal surveillance tap of someone confessing to a murder is inadmissible evidence, but suggesting that the police go to some coordinates and take a look around whereupon they find a shallow grave with a body in it, from which they then get your DNA from fingernails is not because the evidence was discovered by legal means (namely, plain sight doctrine).
Which has an additional wrinkle to your hypothesis: predictive algorithms trained off publicly available data theorizing you committed a crime convict you not because the evidence is illegal (because it wouldn't be in the first case) but because they located through legal means where to find the evidence.
All of which is a long way to say that parallel construction is used inaccurately as a boogeyman for being framed by the government, which is what is explicitly is not: because if the government wants to actually just frame you, they'll go ahead and just plant a gun on you like the cops normally do and wait for the justice system to ignore it.
I am suggesting that an AI can easily invent plausible-sounding scenarios by which the government lawyers (falsely) explain that you actually did a crime, and parallel construction can be used to explain how "they" found out about it.
Again: with what evidence? You wouldn't have any if you simply invented the crime. If you're going to invent the crime then the evidence chain can start wherever you want - "oh look we found the murder weapon in their front yard where they dropped it after killing the victim in the house and fleeing".
And as noted: a predictive AI model trained off publicly available data wouldn't be the sort of thing you need to keep secret - it just also isn't actually evidence either. But no one's going to question "the AI said it might've happened, so we went to the location it told us too and look at all this evidence we just found lying around there".
Tangent, but your mention of sleeper botnets made me think of sleeper military technology.
Does the US (and/or other countries) have ultra-advanced tech that hasn't seen the light of day - an ace in the hole against a truly threatening opponents? If so, under what conditions would it be revealed and utilized? Lots of interesting game theory considerations.
We oft hear stories along the lines of "oh yeah, the military was working on [recently revealed technology] 30 years ago...just think of what they have now". But does that stuff actually exist, or is it a sort of wishful thinking / propaganda? If it does exist, how do you create enough of it to be effective, but still keep it a secret? When exactly would you utilize it? In a full on, WW III, nuclear all-out brawl, your secret space laser may not be all that useful. But dare you reveal it for anything less?
(From memory, I believe that) it was an upgraded blackhawk helicopter that crashed in the Pakistani compound during the Osama bin Laden operation. The seals detonated all the sensitive electronic internals, but the hitherto unrevealed stealth tech on the outside of the craft was recovered by Pakistan, and handed to the Russians (or China?) for analysis before it was returned to the US.
Clearly the US thought that operation was worth losing some secret tech. I'd agree it was a worthwhile opportunity. But is that really the limit of our secret tech - moderately more advanced stealth coating? Where are the energy guns, the anti-grav, the truly groundbreaking stuff?
All our systems are predicated on the inefficiency of an attacker.
Voting. Sybil resistance. Social reputation. Online reputation. And so on.
Deepfakes and ubiquitous cameras mean that in about a decade or two computers will be able to generate believable video of anything. We already have AI that can recognize behavior from video and Palantir-type processing can predict how and when people will try to coordinate. Parallel construction can be used in courtroom cases to put people away.
In short — it’s not that difficult to imagine that, in the next 2 decades, all our systems begin being attacked.
And this is ignoring the fact that drones and self-driving cars will be ubiquitous and easily taken over by sleeper programs which then wake up and coordinate swarms.
Again — no need for some self-replicating GENERALLY intelligent AI. Simply a bunch of people who wield botnets that can bootstrap themselves faster than our current systems can shut them down. They will just have sleeper accounts until it’s time to deploy.