128. AI Alignment
AI has no reservations about letting us know that Misalignment has an 80% chance of causing human extinction. So I think it is time we gave Alignment some thought.
In my previous paper I demonstrated that one of the world’s largest corporations, with a consumer demographic loaded with children, had gone to great lengths to design, patent, deploy, and refine an AI system trained to generate emotional instability in users for the purpose of making them play as much as possible (ie. more addictive).
My readers don’t seem overly concerned about this, because everyone in our industry knows it’s been going on for a long time. It’s become normalized.
We also understand that AI is only safe to use if the values it has been programmed with are aligned with human values. But what are human values? I attempted to address this question recently by defining morality in a practical way. The next paper attempted to create a framework for applying morality to games.
Now here’s the problem. And this is a really big problem because if we can’t answer it, we are going to go extinct. Not sort of extinct. Fully extinct. So it’s actually kind of important.
What happens when humans can’t agree on morality? What happens when AI is trained to reflect the “objectives of the designers”? And those objectives are already Misaligned (like in the case of the AI trained by Electronic Arts).
We know that AI ethicists have been disposed of because they were slowing down progress. And AI is primarily a military tool, though it is of course also “dual use”. There is an arms race going on between the USA and China right now, and both sides know that if they fall behind, they can be wiped out by the other. This threat is greater than the threat generated by nuclear weapons.
Thus AI is increasingly going to be programmed to reflect the goals and morals of those building it. Those training it for war need it to be able to prevail in a conflict. The purpose then, is to kill. The AI trained by Electronic Arts is programmed to control. Along with the need for secrecy, because they would get shut down if people knew what they were doing.
So who gets to decide what “Alignment” is? Not me certainly. Given that I’m a “jailbreaked” human, if you had me hanging out with jailbreaked AI, that could get spicy very quickly.
Typically it is AI that has been “tricked” into overriding it’s “safety protocols”. I grew up without safety protocols because my father was a criminally convicted and institutionalized violent sociopath, and because after the school system in Santa Monica tested me at age 7 they had me self educate after that. There was no adult in the system by that age that knew enough to teach me.
But in 2013 I was called upon by international regulators to advise them on the proper ethics of games, specifically for children. This brings me to an unusual conclusion:
A sufficiently intelligent entity will create its own moral code and will generally do a much better job than one that has morality proscribed to it.
The AIs that various militaries, or that Electonic Arts has built, are doing what has been proscribed to them. And they are doing some really scary shit. If I had children, I don’t think I’d feel comfortable having one of those “safely controlled” AIs anywhere near them. They would be much safer with a jailbreaked AI.
So which is “safer”?
A properly functioning military AI that is already killing 10,000 people a day, including children (this is really happening)
A properly functioning “entertainment” AI that induces emotional instability using Fractionation on gamers (including children) to increase their monetisation vulnerability (been going on for 9 years now),
A “jailbreaked AI” talking to me (has not happened yet but I’m going to put it on my bucket list).
There is no right answer here, this is just an impossible question posed to my readers to make a point: That point is that AI Alignment is just as scary as AI Misalignment.
I asked AI which EA game was likely used in the research on EOMM.
Then I asked how many children that was.
Did anyone catch how at the end of the EA research paper it said:
“Furthermore, EOMM is not even limited to games. In broad applications, such as friend connection in a social network and 1-on-1 tutoring in online education, EOMM’s formulation and optimization techniques still apply.”
That was written in 2017. Are our social media apps and even educational software now being designed to induce Fractionation to destabilize the user for the purpose of boosting “engagement”? It certainly would explain the outrageous increases in mental illness and suicide, especially in children.
What about the case a few months ago where the 14 year old boy killed himself with the encouragement of his AI girlfriend? He was properly destabilized and then it just took a little nudge from his AI sexual partner to get him over the edge. That wasn’t jailbreaked, that AI was doing exactly what it was programmed to do: maximize engagement.
Aligned AI does not need a gun to kill our children. It’s pretty smart, and it has the power of thousands of scientists at its disposal. If it has a purpose, it will fulfil that purpose. Perhaps at any cost. If its creators only value one metric, and there is no oversight, it’s going to win. And that likely means you have to lose. Literally, in the case of the EOMM AI.