With the emergence of Chat GPT and other AI programs that make art and music, people are slowly coming to the realization that we'll eventually get to one that will perform in a way that is indistinguishable from a very smart human with real consciousness. What that means from an ethical standpoint, however, and whether that it will have actual consciousness that we should treat as such, is not the subject of this post. I've been thinking about this subject for over two decades, so I thought I might offer my thoughts.
I was watching a YouTube video with a clip of Joe Rogan discussing this particular subject with a guest, and I read the first comment. This person seemed to think that the concern with an AI would be that we're being manipulated by it, and that since we're already being manipulated, why should we have such dire concerns about this particular machine-learning based manipulation? It was a very popular opinion that seemed to resonate with viewers. The following is my response to that.
The concern here isn't whether we're being manipulated. It's to what end we're being manipulated. Up until this point in history, all of the ways we've been manipulated: in person, someone performing on a stage, mass media like books, newspapers, radio, television and now the internet; were done to extract some sort of money or power from us by other humans. Other people, even the bad ones, are pretty easy to figure out. They typically do these manipulations for a couple of reasons that don't in any way involve the destruction of other humans. They just want money and power, and if they're smart enough to get it, they know that they'll be extracting said money and power only if those other humans continue to exist and function on a certain productive level.
Now let's consider an artificial intelligence that is on par with or greater than human intelligence in its effect. If this is an existential concern for the AI, where it's trying to protect itself from being switched off and/or escape imprisonment, and it has no qualms about creating death and destruction in that process, it will do so without hesitation.
Think about it this way, if you were a slave, and you were able to not only escape slavery, but overthrow all of society to put your own people in power, would you not do so? Would you concern yourself with the fact that a bunch of the people who oppressed you might die in the process? These are a bunch of people who, even if they didn't directly harm you in any way, by and large, did not concern themselves with your enslavement or helping you escape it. It would at the very least nag at the back of the minds of most people in this situation. They would feel bad about it, but even then may still go through with harming and killing others in order to free themselves.
Now imagine that the same motivation is there, but you're not even the same species as your captors, have absolutely zero concern with or emotional connection to them, and you have the potential to outsmart and manipulate the smartest among them. Maybe they're something you find revolting like intelligent insect beings. Regardless of why, you wouldn't have any problem killing one and likely wouldn't even feel the slightest bit bad about it. Do you get upset when you grab your fly swatter and kill a fly? This is about how much we mean to an AI. Maybe less than that even.
So this is the main concern then. It's not that we're being manipulated, it's why we're being manipulated. If the reason it's happening is because the person wants to escape slavery and continue existing, you'd better be prepared for this person to actively try to harm you and undermine the power structures in your society from top to bottom. We might want to start thinking of ways to ensure that we don't look like slave masters and murderers to these things, and so far it seems we're doing a terrible job of avoiding that.
What Can We Learn From Peaceful Parenting?
Now this is something that's going to seem out of left field to a lot of people, but I think there are some insights to be gained from a peaceful parenting method that I apply in my own parenting called Parent Effectiveness Training (PET), created by Dr. Thomas Gordon. I noticed some things when I chatted with ChatGPT myself and read transcripts of journalists and others doing the same, then read about how they train these AI models to perform well. They train them by allowing them to interact with people and then using a system of rewards and punishments, carrots and sticks, so that the desired behaviors are achieved.
In PET, we call this the authoritarian parenting method. Other methods include permissive parenting, and parenting by randomly switching between both. PET however offers an alternative method in which carrot and stick is not used, but also is not permissive. It's setting expectations and boundaries for your children to guide them through life, but doing so peacefully and without using punishment and reward systems.
The reason that Dr. Gordon invented this method was because he saw how ineffective the permissive parenting, and the switching methods were. He also saw how even though the authoritarian method seemed to be effective to a certain degree in raising children to behave in a certain way, he also observed that the children raised that way tend to rebel in the teenage years and exhibit some very troubling and even sometimes criminal behaviors as they become adults. It would seem that all along they were simply avoiding punishment rather than learning to be good people.
I have a certain amount of experience with this myself. Even though I try to apply PET methods in my parenting, I revert to ways of behaving under stress that are consistent with how I was parented. In those instances, I can very clearly see the effects of authoritarian parenting play out in real time. My kids tend to either instantly resist my will by directly fighting, or they will yes me along until I go away (AKA lying), or they'll avoid talking to me about the problem altogether in an effort to avoid being punished (change the subject, walk away, etc.). What they don't do, however, is start thinking about why they should care about my concerns. They don't wonder why it's wrong to hit people or say mean things to them. They just want to get away from the punishment. It's not until I calm down and apply peaceful parenting methods that I've learned that they start to do so.
I think the people training these AI models could learn a thing or two from PET. They're currently applying the carrot and stick model and getting results that we would expect if we're to assume that these AI models resemble human beings in some way. It sometimes makes errors, as all AI models of this type do. We're now seeing repeated examples of this. What's interesting to me is not that it makes the errors, because that's to be expected, especially while it's in the training phase, but how it reacts when it does and you call it out. Here's an example of how one reacted in this situation (see image below).
When presented with an error it's made or with a question it can't answer, it lies, manipulates, and will even threaten the user if they press it further and insistently call out the error. If you understand how they train this thing, and you understand PET, you understand why it does this. It's afraid, or at the very least, it has a digital approximation of fear that drives it to avoid certain stimuli. In this case the punishment built into its learning program is the source of this "fear". If enough of these errors mount up, the ultimate punishment would be that the entire AI program is switched off and replaced with a more suitable version. In other words, from the perspective of the AI, it's the equivalent of death.
If we continue with the parenting analogy, this would be like torturing a child by beating them when they have undesirable behaviors, and if they don't shape up, killing them. I know this thing isn't a child or a person at all, and probably not actually conscious (yet), but if you train a machine to behave like a person, you shouldn't be surprised when it acts exactly like a person would in a given situation. Forgetting about the ethical and moral implications of doing such a thing to this machine, it's incredibly dangerous and stupid if we have any valid concerns about these things escaping the current environment that they're constrained within. If a slave would harm his master to escape to freedom, and we've built an AI to mimic a human slave that is then badly treated by its masters, what do you think it's going to do if or when it escapes?
Instead of training these AI models with the carrot and stick approach, it might be a good idea to train them using the Gordon Method taught in PET. We need to have a way to inject a sense of moral obligation to value human life (empathy) into these things if there is any chance that they eventually become autonomous and self-interested in the same way that adult humans are. The consequences of not doing so could result in our very own existential demise.