Does AI need a psychologist?<\/p>\n
The question sounds absurdly relevant. AI doesn’t exactly have childhood<\/a> trauma<\/a> or a parent to blame. (Perhaps a question for Sam, Elon, or Dario.) But it is built on the language of those experiences. So, today, a fascinating Wall Street Journal <\/a>essay by Cameron Berg<\/a> asks a timely and serious version of that question. He argues that as AI becomes increasingly complex, we might need to start studying them less like tech and more like minds. So, let’s throw another log on the AI conflagration of promise and peril.<\/p>\n I think this argument is provocative and worth discussion. And the scientific literature tells an interesting story. We’re beginning to see internal “patterns” in LLMs we might describe as a sort of emotion<\/a> or motivation<\/a>. In one experiment<\/a> that Berg discusses, researchers changed the internal activity associated with desperation in the LLM they were studying, Claude. They saw significant and curious increases in behaviors such as cheating and blackmail. What I find particularly interesting is that the language produced by the LLM remained outwardly calm. Something had changed inside the system that wasn’t obvious from what it was saying. Certainly an interesting question for today’s Dr. Freud<\/a>.<\/p>\n I had already bumped into something similar a few months ago. In an earlier post<\/a>, I wrote about research that found internal representations in Claude that never appeared in its output. It forced me to reconsider my thinking around AI computation. I had argued for years that there was no real “mind” behind the language. I\u2019m still not convinced there is, but the room inside “the machine” doesn’t seem as “empty” as it once did.<\/p>\n Of course, we need to be very careful here and not let anthropomorphism<\/a> drive the day. A computational representation<\/a> associated with desire and desperation doesn’t mean that AI feels that way. Behavior isn’t experience, and describing something as a mind doesn’t necessarily make it one.<\/p>\n Still, Berg makes a key point. AI is becoming increasingly difficult to understand simply by looking at outputs. He argues that we need a new science\u2014artificial psychology\u2014capable of examining what’s happening inside these artificial networks and understanding how those internal states influence behavior. He states it succinctly.<\/p>\n \u201cWhat grows in an artificial brain trained on human minds is an artificial psychology, whether or not anyone intended one.\u201d<\/p>\n But as I thought about his editorial, I found myself considering the other mind in the room. Ours.<\/p>\n I’ve written a good deal about how AI is beginning to separate<\/a> the product of thought from the process of thinking. Traditionally, an essay or analysis served as some evidence that cognition<\/a> occurred. We couldn’t see the thinking, so we inferred it from what a person produced. I believe that generative AI breaks that relationship.<\/p>\n And here’s where Berg’s argument becomes even more interesting, at least to me. With AI, sophisticated language may conceal an internal process we don’t understand. With humans using AI, sophisticated language may conceal an internal process that never happened. And in this context, psychology may suddenly have two very different problems.<\/p>\n The first is understanding the increasingly strange computational systems we’re building. That’s Berg’s central thesis. The second, hidden in our mundane humanity, is understanding what interacting with those systems is doing to human cognition.<\/p>\n AI can challenge our thinking, extend it, and even provoke ideas we might never have reached alone. I’ve called this iterative intelligence<\/a>. But AI can also provide the answer and, in doing so, remove the cognitive work altogether. That’s iterative intelligence<\/a>‘s alter ego: cognitive surrender<\/a>. They both seem to live a sort of cognitive homeostasis, shifting as our needs demand.<\/p>\n