CliffsNotes For The AI Apocalypse
A dear friend recently said that he couldn’t imagine how the existential threat of AI might play out, so I spent some time trying to come up with a few examples:
An AI purposefully destroys humanity. What if a super-intelligent AI decided it wanted to launch nuclear missiles?
I know, there are technical and human checkpoints that control such things, so the super-intelligent AI would need to be smarter than the lesser AIs and human beings tasked with constraining it, but that’s not really the stuff of sci-fi since every major AI developer is rushing headlong toward inventing it.
They’re pursuing this goal with the knowledge that, once achieved, no amount of human ingenuity will allow us to control a super-intelligent AI because it will be faster, better, constantly learn and therefore relentlessly stay ahead of us.
For their part, AIs are already getting increasingly good at lying and otherwise working around those aforementioned checkpoints to accomplish goals that can also emerge from their own processes (and not those of their coders).
The Question of “why” it might want to destroy humanity is immaterial when we’re talking about an intelligence that we could never understand.
So, all a super-intelligent AI would have to do is get installed at — or find a way to reach — some component of a nuclear arsenal, nuclear reactor, bioweapons research lab, or some other potential prompt for the destruction of the world, and there’d be little we could do about it, assuming we even knew it could happen before it happened.
AI mistakenly kills us. What if an AI, perhaps not even super-intelligent, erased us in service of accomplishing another goal?
This is the famous “paperclip” scenario in which AIs are tasked with constantly improving the sourcing and manufacture of paperclips and ultimately conclude that human beings’ demands for resources are an obstacle.
Think of it as an Unintended Apocalypse.
Again, the tech isn’t far-fetched and the reliability of coding to guardrail or otherwise constrain AI behavior is already in question and may be so permanently (remember, the AI is always a step ahead of whatever we’ve told/tell it to do). Assumptions that such “mistakes” couldn’t happen tomorrow are simply not supported by the reality of what’s happening today.
So, maybe the AI doesn’t kill all of us at once but does so selectively by, say, cutting power for heat in the dead of a cold winter to keep another public works project on schedule and thousands freeze to death before the action can be reversed. Maybe it reaches a point in its processing solutions for climate change and decides that all airplane travel should cease…including the thousands of flights currently in the air.
AI kills us by trying to help us. What if AIs that are nowhere near super-intelligent start making healthcare decisions that hurt individuals but improve health or financial statistics overall?
What if you are considering a medical treatment and you AI has determined that the cost/likelihood success wasn’t compatible with your stated financial plans for your family? Instead of having your insurance AI deny you, it simply curates your Internet search results to talk you out of getting the procedure.
What if the hospital is considering a dozen candidates for a certain procedure and its AI analyzes who it will benefit the most (and/or which patients will cost it the less money over time)…and then surreptitiously leads the least reliable bets to choose alternate treatments?
What if AIs start reassessing the survival values of entire groups of patients — let’s say people who weigh more than X, or are past so-and-so age — and nudges them out of life-sustaining services so as to maximize the health and survivability of everyone else?
On a related note, what if AIs in education and the workplace start doling out opportunities based on complex analyses of future life “value” to society? It might not entail physical deaths, at least not directly, but could certainly kill lots of careers and close down paths to individual fulfillment.
Super-intelligence and/or overt biases aren’t necessary for this scenario to play out; AIs right now could be making nuanced decisions that fly under the radar of regulators or individual recognition.
AI kills us by trying to avoid killing us. What happens if you’re not selected to survive the next autonomous car crash?
The car with a family of four is about to crash into another family of four standing at a street intersection, so the AI decides what to do by assessing every individuals’ “worth” to society and decides who should die. This conundrum has been called “the Death Algorithm,” and the math could just as easily be based on who would sustain what injuries and have the best chance of recovery.
Ditto for the crippled airplane skimming the roofs of a residential area as it tries to make it to an airport runway: what would be the most “efficient” accident?
Now, humans make such decisions all the time and we’re inherently flawed, so you might think that letting AIs do it would be at least more just, if not always satisfying for all involved.
But, like the AI plotting to replace mankind with paperclips, we might not have visibility into the ever-evolving criteria AIs use to make those life-or-death calls. The code that enables or influences them might be flawed or discarded. Maybe the AI has learned something that changes how it weights its decision criteria.
And then there’s the fact that probabilistic options become certainties of fact once a course of action is chosen; when an AI decides what will happen based on what it thinks should happen, it’s determining fate. Today’s 27% likelihood of someone contributing to society might rise to 35% after surviving a crash, thereby helping ensure they get even more favorable treatment in the next accident.
It might not kill all of us, though it would suck getting put into one of declining value categories.
AI subtly erases our humanity. This scenario is the hardest to talk about because I think it’s the most likely outcome.
The AI toffs developing super-intelligence are also racing to plug the latest versions of their LLMs into “intelligent agents” that could function as deeply personalized, constantly on, and proactively involved assistants for our every decision and activity.
Think Siri or Alexa, only knowing you better than you know yourself and willing to tell you about it.
Imagine being able to outsource every decision, large or small, to an entity that not only knew you intimately but could access and synthesize more information in an instant than you could amass in a lifetime. Then, imagine it offering up commentary and guidance even if you hadn’t asked for it.
Where you go and what you do would be decided, at least in part, by your assistant. What you say and to whom you say it would be mediated by its advice, along with how you interpreted what you saw and heard.
Everything you know would be filtered by what it wants you to know.
These digital assistants will be “sold” to us as a convenience and its early applications will be revelatory, far surpassing the thrill we can already get from using an LLM to write a school paper or summarize a work report. Maybe it’ll be “free” as long as someone regularly uses it.
But it won’t be free…corporate and government interests will be working in the background to nudge your thoughts, decisions, and actions in certain directions. The AI assistant will be doing the same based on its virtual model of what it thinks you think and want. It will engage with other assistants to better facilitate efficient conversations and activities.
As it quietly pushes us in directions it thinks best, for whatever reasons, we’ll gradually learn to skip learning anything ourselves and rely on it instead relying on people we trust or love. We will slowly forget what it was like to live our lives as independent human beings, our next decisions preordained by our past ones.
We’ll all still be alive but AI will have killed us.
My four scenarios for the AI Apocalypse are by no means the only ones, nor do I claim that I’ve presented them with all of their nuances and implications. But each of them are reasonable outcomes of what AI developers have already delivered, not to mention what they’re aiming to accomplish in the very near future.
I wish it were sci-fi.
