Here is a paper on the need to train AI so that it doesn’t prey upon our base human emotions:
https://om.omni.se/a/gaestkroenika-sociala-mediers-algoritmer-styrs-av-var-raedsla
Like all current AI research, the sentiment comes from the right place: AI need to understand the human better.
But the understanding is backwards. Not only are the concerns backwards, but the AI system is also backwards. Both are based on the “bottom-up” approach to understanding humans. The belief is that we have a reptilian or “lower” brain which responds to certain stimuli with fear and, eventually, violence, stemming from a “component part” we inherited from the last common ancestor we share with chimpanzees. We could extend this “component parts” logic to all ape phenomena: ape communication became articulated language, ape eyes were gradually whitened, and so on. The bottom-up approach, generally speaking, breaks down the human into various systems and organs that evolved due to external stimuli over millions of years.
Human aggression and reason are seen as proof of this. Since chimpanzees are highly aggressive, it’s assumed that our aggression is a sort of stowaway that retains this aggression. Similarly, chimpanzees are very intelligent; hence, it’s assumed our intelligence is derived from a similar intelligence.
Since both the aggression and intelligence “components” are supposedly derived from our last common ancestor, it’s assumed that they have both “evolved” by some degree. We want them to be opposed: our intelligence should lead us toward peaceful resolutions; conservatives (or innate-aggressionists) generally hold the same view, with the belief that the aggression must still be redirected elsewhere for peaceful ends (combat sports, etc).
But often our aggression and intelligence become aligned: our intelligence makes the technical blueprints that our aggression uses to kill millions; and AI could make us hate.
Based on this bottom-up or “component features” approach, it’s natural then to assume that we must program AI with intelligence, and not with violence. This is exactly what they’ve done: they feed it lots of textbooks, and not much violence, if any.
In fact, they can’t feed violence into AI, because to do so they would have to *parse* violence for the AI to learn about it, and this is impossible for an AI curator to even comprehend. What can possibly be *parsable* about violence, if it’s just a stowaway?
My ROBA Hypothesis is instead a top-down approach; animals have a fixed combat loop restrained by their natural weapons and defenses, but we have an exclusive human property (XHP) which allows us to use objects in combat; all humans can do this, no animals can. This ruins the combat loop, which is now constantly open, and all objects are lethal, all violence is contagious and escalatory, rendering all human relations potentially apocalyptic. Hierarchies must be formed by other means.
From this “ROBA Drive” descend all things human and good; intelligence, culture, religion, language, art, kinship, and the sciences reformulate the apocalyptic threat, but it’s now in a deferential form: by engaging in them, we defer our apocalyptic violence indefinitely. There are no component features here; the ROBA Drive simply co-opts everything from the top-down.
Train AI to understand this, and its intelligence will be world class.
My upcoming book If These Fists Could Talk will present this in detail. It should be out in a month.