In a recent post I reflected on the much-discussed potential dangers of autonomous artificial intelligence (AI), and concluded that the anxious hand-wringing about this threat that invades our news feeds on the regular results at least as much from fear of the new, and other emotional human tendencies, as it does from careful scientific analysis. I closed by saying that if we are truly concerned about the “behavior” of these artificial “organisms,” we need to study it systematically using the same approach that has taught us so much about the behavior of naturally-occurring organisms. This proposition makes sense for both general and specific reasons — general because every interesting behavior (whether natural or artificial) will necessarily be the combined product of learning history and current context; and specific because the programming heuristics underpinning autonomous AI were inspired by operant learning.
The notion that a behavioral approach is critical to understanding AI is hardly original to me. For evidence, check out a recent short essay in the journal AI & Society which argues much the same thing, while pulling no punches about what will not help:
Cognitivism, the leading school of thought in psychology, is wrong. [It] misleads us about intelligence, and overshadows the behaviorist approach, which is the framework that enables comprehension of the new AIs.
The argument, by the way, is similar to one Skinner (e.g., Science and Human Behavior) advanced regarding biological organisms like human beings: that by directing attention to fictional causal variables, a cognitive viewpoint undermines our capacity to generate solutions to real-world behavior problems. The argument takes on extra urgency when you consider that, as some argue, the biggest problem of all may be autonomous AI’s threat to human survival. And if that’s so, then the second biggest threat might just be ways of conceptualizing behavior that prevent us from understanding the nature of the threat.
“Cognitivism prevents us from understanding artificial intelligence” (available as full text here) will take you just a few of minutes to read and helps to flesh out the [behavioral] context in which we should think about autonomous AI. For a more-things-change-the-more-they-stay-the-same experience, I recommend pairing it with one of Skinner’s commentaries on cognitivism, such as “Why I am not a cognitive psychologist.”
The essay, which has been downloaded close to 3000 times as of this writing, demonstrates how it is possible to share and defend a behavioral perspective in mainstream publications. Its author, Mehdi Bugallo, who’s based in Lille, France, deserves our support, so check out his essay!