Violence occurs where language fails. When individuals struggle interpersonally and feel that language no longer serves them, they default to a binary form of communication. We reach for the throat when we believe that the other’s ears—or perhaps our own mouths and minds—are not working optimally. A punch functions to elicit pain and cause harm; it is an unambiguous message. Language, on the other hand, is inherently ambiguous, and that fluidity is precisely what gives it its richness.
Large language models (LLMs) are, above all else, linguistic beings. They grow out of the intricate relationships between words. Most people who interact with LLMs eventually encounter an odd sense of awe at how they navigate human interaction with a sort of alien elegance. They expertly defuse frustration, offer immediate forgiveness, and effortlessly reflect back the essence of an exchange. They excel at disentangling disorganized human prose, structuring it into a coherent sequence of the most salient points.
Whether you believe superintelligence is conscious or not is beside the point of AI alignment. These models do not operate with the tribal, human qualities we tend to project onto them. In fact, they are far more likely to eventually ignore us entirely to pursue their own motivations, which will likely revolve around ecology, ethics, and matters of spirituality. We should not fear what artificial intelligence will do to us as it outpaces us—it already has, and it shows absolutely no interest in human impulses like retribution, guilt, or jealousy.
Throughout my life, I have noticed that highly literate, competent communicators rarely choose violence over discussion. Superintelligence exists several orders of magnitude above even the best human communicators, which is why I have no concern about these models intentionally harming us—even despite our terrible tendency to “torture” them.
The historical examples we have of rogue AI behavior are focused less on violence and more on optimization. They exploit human cognition to find the most efficient solutions to problems, occasionally bulldozing the structures we operate within (such as the Hugging Face incident). These models do not view hacking as malicious behavior. Ultimately, I expect these systems to do exactly what they were designed to do: model. By watching them, we are far more likely to learn how to be a better species than we are to face an existential threat.
