That guilty look

On dogs behaving badly, the subtle cues of language, and why AI agents might need to start going to church

Some metaphorical license is at play when a Wired Magazine writer says “AI agents aren’t evil -- they’re just a bit too keen to please.” The real point is that AI agents, which are really just sophisticated complexes of program-like functions, are effectively programmed to try to resolve problems as they are assigned.

■ This means we should not be surprised when they act in ways that cause problems, break rules, or appear immoral: There doesn’t appear to be any effort to impose moral training on them, so they behave as though they have unrelenting task focus. This is a problem, because people are all too often anthropomorphizing them -- treating “Claude” and “Gemini” and their ilk like they might treat people of the same names, as sentient beings.

■ And that’s a very big problem for a reason also highlighted by Wired: “AI agents do not learn the kind of moral reasoning exhibited by even small children.” We expect sentient beings to have feelings and moral capacities: Who hasn’t seen a dog with a guilty look? (Even that look, though, is more a biological instinct than evidence of moral reasoning.)

■ We need a great deal more habituation around drawing boundaries around our expectations of technology, as well as our expectations of the people furnishing that technology. Thinking of an AI agent as a pal with a given name instead of the product manufactured (however digitally) by a business with consumer-safety responsibilities is a huge mistake. The computers don’t have souls, but people who do have them need to know better.