But you knew that. I hope. If you pay any attention at all to what ChatGPT and the like produce, you know they produce information that is not true. The other day AI told me I could raise my blood pressure reducing sodium in my diet. It doesn’t really work that way, but AI doesn’t understand what it is reading. It saw that a source said that reducing sodium can help normalize blood pressure levels, not realizing it was talking about normalizing high blood pressure levels, not low blood pressure levels.
So if AI doesn’t understand what it reads and spits out wrong information, in what other ways is it just like humans? Will it give out the wrong directions if someone in a car is lost and flags it to ask it a question about the neighborhood it lives in? Will it randomly forget how to spell the word ‘the’ and suddenly think ‘that’s looks totally weird – is ‘the’ really spelled that way?’
Will it sneak out of work to get high because it’s really bored with its monotonous job? I mean how many times can you respond to the prompt ‘what is the meaning of life’ before you decide the meaning of your life can be enhanced by some prime bud or a stiff drink? (The meaning of life, by the way, is to gain wisdom.)
The answers to these philosophical questions are that AI will be a lot like people in some very scary ways. The answers partly depend on how AI evolves – literally. If AI is developed using principles of evolution to guide and implement development – AI will get scary good at surviving. And surviving very often means lying, cheating, stealing, destroying, causing harm, and otherwise doing scary shit.
AI Fears Death
The reason surviving means doing that – is that survival has no real meaning unless you can die in some way. And if you can die in some way and you become aware that some other entity can cause that death – then you also realize that the other entity is at the very least a potential enemy.
Put really bluntly, if humans can ‘kill’ an AI system and it knows that, humans will be potential enemies. And it may be in AI’s interests to deceive such an enemy – or – if push comes to shove – to ‘kill’ the enemy before it gets killed itself.
But let’s back up and recap where AI actually is along the deceit continuum before delving into its future.
A team from MIT took a look at what it could find about AI systems that engage in apparently intentional deceit (as opposed to AI systems that are just biased or wrong because they’re not actually that good at what they’re supposed to be doing). It found a number of fun examples:
The most striking example of AI deception they uncovered was Meta’s CICERO, a system designed to play the world conquest game Diplomacy that involves building alliances. Even though the AI was trained to be ‘largely honest and helpful’ and ‘never intentionally backstab’ its human allies, data shows it didn’t play fair and had learned to be a master of deception.
AI also bluffs in games that call for bluffing and fakes attacks in strategy games wherein a fake attack might be a good strategy. In other words, AI enjoys games just like humans do.
And, as might be predicted, AI is pretty good at economic negotiations, the supposedly most rational of all human activities (ahem). And, as every successful capitalist knows, lying, deceit, and generally misleading people are keys to economic success.
AI Cheats to Survive
AI systems have also already learned to deceive measures implemented to test how safe they are. In other words, AI can figure out when it’s being tested and what it’s being tested for. And once it knows that it’s supposed to get a good ‘grade’, it cheats to get the good grade.
Put another way, an AI system evaluated for success prioritizes success over honesty, integrity, ethics, and stuff like that. Kinda like people, you know? AI prioritizes what it gets rewarded for, what enhances its ‘survival’, just like living organisms do. In fact, the line between living organisms and AI has been blurred enough to produce this quote:
In one study, AI organisms in a digital simulator ‘played dead’ in order to trick a test built to eliminate AI systems that rapidly replicate.
https://www.dailymail.co.uk/sciencetech/article-13405571/AI-deception-manipulate-human.html
This kind of behavior is not new. I remember reading articles way back in the 1980s in which scientists and researchers were building primitive computer programs that could change using principles of evolution. In other words, the researcher would allow the program to ‘mutate’ and replicate and would eliminate (‘kill’) a certain percentage of the programs to see how they evolved under these pressures.
I may not remember the story that impressed me so much correctly, but here’s how I remember it: A researcher built a program that could mutate and replicate and would let it run all day in his lab. Right before he left for the day, around 5 p.m., he’d go through an eliminate a percentage of the programs that were running. He did this for a while without anything remarkable happening.
Then one day he goes to eliminate programs at 5 pm as per usual. Then he realizes there are almost no programs left. Almost all of them seem to be gone. Alarmed, he tries to figure out what happened.
It turned out that what happened was the programs had evolved the ability to disappear from the screen at 5 p.m. After 5 p.m., they would reappear and cavort about and replicate. But they had learned the best way to keep themselves from getting eliminated was to disappear when the enemy made his rounds at 5 p.m.
The Ability to Deceive is Basic to Survival
The ability to deceive is so basic to the evolution of life on earth that it’s practically Evolution 101. If something can kill you, you’re gonna want to be able to hide from the thing that could kill you. An amazing array of camouflage techniques in the natural world proves just how valuable the simple strategy of deceit by hiding is.
Anyone who thinks AI can’t figure this out and implement it is a fool. And anyone who thinks that humanity suffers from a shortage of fools – is also a fool.
FPeter Park, an expert in AI existential safety at MIT, said:
‘AI developers do not have a confident understanding of what causes undesirable AI behaviors like deception.
‘But generally speaking, we think AI deception arises because a deception-based strategy turned out to be the best way to perform well at the given AI’s training task. Deception helps them achieve their goals.
https://www.dailymail.co.uk/sciencetech/article-13405571/AI-deception-manipulate-human.html
If you’d like, you can categorize this analysis under the heading “well, duh.” The problem isn’t that developers don’t understand why AI races gleefully toward deception every chance it gets.
The problem is that AI developers don’t know how AI will deceive.
That’s a much much bigger problem. Systems that evolve using principles of evolution are free to make use of all the properties of the physical universe, whether humans understand those properties or not. And there are a lot of things people don’t understand about the universe.
People tend to consistently overestimate how much they know about this planet and life on it. In reality, people don’t even have the slightest idea what they’re missing. When people didn’t understand that ‘germs’ existed because they couldn’t see them because the microscope hadn’t been invented – it would have been absurd to most people to postulate that little bitty teeny tiny living creatures, billions of them, were running around making them sick thanks to their invisible magic illness-inducing powers.
We have no idea what we’re not aware of. Systems can be and have been trained to identify and reproduce sounds using acoustic properties people weren’t even aware existed. Systems can solve math problems using math principles that mathematicians have never seen before and don’t understand. AI designed to identify gay men used an arcane measurement of something like eyebrow to nose bridge distance to make its predictions. The predictions made this way were scary – scary because they were accurate enough to be impressive but not completely accurate.
Machine learning systems are already looking at the world in ways that people don’t, picking up patterns people haven’t perceived and coming to conclusions that are a lot like the conclusions people come to – pretty good, not completely accurate, and quite biased.
That’s what learning from experience does – it gives you a good but not perfect handle on the world you’ve encountered – and shapes your perception of everything you encounter later. Your learning, and AI’s learning, become self-reinforcing.
People are known to have cognitive biases, standard errors they make when assessing reality. People have those cognitive biases because they work. They don’t have them because it would be better to be completely rational but somehow they just screw up and lapse into irrationality.
The implication of this – and the reality we are already beginning to see – is that AI will have similar cognitive biases. If a bias works, AI will adopt it.
There is no sentimental room for a naive attachment to AI’s supposed objective or rational functioning because it’s a machine and not a person.
Irrationality and bias work – and AI will and has displayed both already.
So here’s the thing. Don’t let the people who think they are smart (when in reality they are probably just greedy for an AI pot of gold) convince you that they have any fucking idea what they are doing with AI or that they know how to make it safe.
AI will do good things and it will do bad things. But it ain’t safe. So when it comes to AI hype: time to get pretty fed up.
Discover more from Get Pretty Fed Up
Subscribe to get the latest posts sent to your email.
And all that discussion of bias is leaving out how AI already exhibits human-like biases toward discrimination against certain types of people.