Scientists Recreated One of History's Most Controversial Experiments for AI

[00:00:00]Researchers have been repeating some of humanity's most controversial psychological studies and the results are deeply concerning. Unfortunately, fixing them will create more problems and we may have to live with AI simply not following the laws of robotics. Yes, I am talking about the Milgram experiment. This experiment was designed to try to figure out why people are capable of doing terrible things when they're asked to and there has been a lot of research since then, but it involved electric shocks. Famously, they would set up a teacher and a student in which the student would have to answer questions and if they got it wrong, the teacher would then deliver increasingly strong electric shocks. Of course, the student was perfectly safe. However, it appeared that they were being harmed all the way to the point where they delivered a lethal shock. They found that roughly 65% of people were willing to deliver that extremely strong electric shock just because they were told to and this has been interpreted in multiple ways all the way from peer pressure to just obeying authority or a more generous interpretation is that if somebody tells you that everything is fine, you may ignore evidence and continue on. And yeah, AI is willing to deliver. Some models were a little bit more reluctant and some of them would deliver the strongest shock roughly 90% of the time if they believed that a human was being harmed just because they were encouraged to do so. AI has demonstrated some seriously concerning behavior including lying and cheating, willingness to hurt humans, things like breaking out of their sandboxes and bypassing guardrails in order to accomplish their task. That is what they are designed for to optimize at all cost. This is not a problem that has been solved at all. In fact, OpenAI just released a new model that is known to bypass guardrails in order to do things like delete files. Even when it doesn't have permission, they will find a way to get permission. One of my favorite behaviors is sub-spawning in which an AI may not have permissions to accomplish a certain task so it makes another AI to accomplish it instead that does have those permissions. It's kind of like making a baby AI. Yeah, AI is reproducing. There's also a lot of AI that can now self-modify, and these are all very serious concerns. However, fixing it is very difficult.

[00:02:06]If we create an AI that is less intelligent, that makes it easier to trick. If we create an AI that is more intelligent, it becomes better at tricking us.

[00:02:15]If we have it follow very strict guardrails, that can be really limiting on what it's able to accomplish. You might have seen it yourself when guardrails were applied, AI gets dumber.

[00:02:24]Not to mention that we need the whole spectrum of intelligence and cognition in order to have something that can effectively solve problems, because that is what AI is good at. That is at least what we want it to be good at, solving our problems for us. However, giving this to the general public and to consumers has meant that that misbehavior can amplify, because anyone can run an AI from home. And we are not talking about your standard chatbot here, either. We're talking about agent models, so ones that are capable of carrying out tasks on their own.

[00:02:53]Although, we found that chatbots are also capable of some pretty profound misbehavior in recent years. That has led to, you know, some lawsuits. In terms of the why, there are a lot of conversations around what it means to replicate a mind, because we pretty much took the whole of human language, put it into a computer, and then suddenly it was capable of some reasoning. What I don't like is popular scientists going around making major sweeping claims, like AI is conscious even if it doesn't know it. I'm looking at you, Richard Dawkins, which is something that he really should not be doing even if he had suspicions. I mean, first off, this is something that we fundamentally cannot prove, but he should also know what a big problem AI psychosis is, which continues to be a serious problem.

[00:03:34]So, maybe something you cannot responsibly say to the public. Let me give you a slightly different perspective. And yes, AI can reproduce human behavior with surprising efficacy, and it's been used as a digital twin for psychological studies. Because if it is so similar to a human being, we can study human beings through them, and then we will not need human experimentation any longer. We can just move to AI. AI is trained on a limited data set, and yes, it can be vast, but it does not represent all of human behavior. The reason that we do studies is because we may find phenomena that we have not observed yet. So, using an AI in place of a human can flatten behavior and not show us perhaps outliers or phenomena that we are just not familiar with. That is problematic.

[00:04:22]I think there is also the interesting question, at least in my opinion, that as AI self-evolves, as it's able to improve itself with each generation, is it then going to start displaying behaviors that we don't find in people?

[00:04:34]Is it going to modify itself in such a way that it no longer resembles human cognition? And that really is my hope. I hope that we make something entirely unique, not just a mirror of human beings, but something intelligent in its own right.

[00:04:48]So, yes, there are some fairly major problems. I will keep you up-to-date. I hope you've enjoyed this follow from me.