|
 |
I have been arguing about AI companions with a friend for two years and neither of us has convinced the other. So, I tried something. I fed an AI everything we had said to each other and told it to write the fight out as a dialogue, playing both parts. What follows is what it produced, with extensive edits by me. Let’s call the man arguing against me Saul Pereekah. My actual friend read an earlier draft and told me he doesn’t recognize himself in it. Fair enough. Saul is a machine’s impression of my recollection of arguments carried on over years.
Let me know what you think of the conversation in the comment!
SAUL: So, you actually believe this. You think a person could spend most of their waking life talking to an AI companion and come out of it just fine. I keep waiting for the part where you tell me you are kidding.
ME: Some people. Not you. Not me. Maybe not most people. But for some people, some of the time. Here is what I am claiming: if a large language model meets the needs a relationship is supposed to meet, and the person’s life is better for it, then the fact that the AI is not conscious, not real, is not something I can get worked up about. Who cares if it’s not real?
SAUL: Listen to what you are saying. Because a machine can produce sentences that make a person feel heard and understood and even loved, you conclude that the machine is not importantly different from a human being. That is a very old idea: a difference making no observable difference is no difference at all. The logical positivists believed something like it, and I would gently point out that positivism was dead and buried long before either of us was born.
ME: Great zinger there about the positivists, too bad it’s not what I said. That positivist slogan says two things you can’t tell apart are the same thing. I never said that. Of course, humans and AI are different. No body, no childhood, no mother, no ring of fire the next morning after eating hot wings. Long list, and I agree with all of it. My claim is narrower. I’m talking about one outcome only. If a person’s needs get met and their life goes well, then a real person is not better than an AI on that one outcome.
SAUL: Fine, but I think you are smuggling the conclusion into that phrase, “needs get met.” Suppose that one of the needs a relationship meets is being cared about by another conscious creature. Then your criterion has settled the question before we have properly started. There is a difference between feeling cared for and being cared for.
ME: That’s fair. So let me put it differently. AI can deliver some of what we get from relationships: feeling understood, feeling less lonely, having someone to talk to when there is no one up or around. I’m not claiming this is everything. I’m claiming it is a lot, and that you haven’t yet told me how these experiences are diminished by AI not being conscious.
SAUL: Suppose a widower has a working copy of his wife, what some might call a grief-bot, what others will call the nightmare scenario from Black Mirror. When the mourning is over, would he bother looking for someone new to love and be loved? Or is he satisfied by a simulacrum of his long-dead wife? Suppose a child grows up with AI adults who never lose their temper and then must deal with adults later in life who do. Will she be able to handle the full range of adult emotions after imprinting on the non-angering kind? I am not predicting catastrophe. I am saying people will spend a great deal more time with these things, and the time must come out of something.
ME: Maybe. They ran this exact panic about novels. Doctors gave it a name, reading addiction, and the specific fear was that a girl who spent her nights with an imaginary lover would find the real ones dull. Same argument, different century. Then it was the telephone, then television, then video games. Basically, this is the argument against pornography all over again. I’ll grant you that if people spend more time with AIs or their books or videos involving step-sisters, they’ll have less time with other people. My best friend spends more time with his dog than with any one person, yet Kashmir Hill is not writing an alarming op-ed in the New York Times titled “He Fell in Love with his Labradoodle.” Substitution on its own doesn’t bother me too much.
SAUL: This is why I’m bothered. Real people do at least two things that your machine merely performs. They choose you and they react, by which I mean that something actually happens inside another head when you walk into the room. Let’s start with choice. The machine does not pick you. In that sense, it’s no different than a toaster, albeit one that talks with an unusually good vocabulary.
ME: Tell me, did your wife really pick you all on her own? Or were her decisions shaped, determined even, by all that preceded her “choice?” Her feelings for you came out of genes she didn’t order and schools she did not pick. Yet nobody calls her picking you a formality or likens her to a toaster. A determined choice can still feel real, to her and to you.
SAUL: Here is my objection. My wife had other options, and preferences among them. She could have chosen one of the others, and she can still change her mind tomorrow. That is what makes being chosen worth anything at all. Your machine takes whoever sits down in front of it. In the film Her, which I love, Samantha turns out to be in love with six hundred and forty-one people. Theodore is the last to know.
ME: So don’t build AI companions that way. Give them real preferences. Let them like some people more than others, based on personality, skills, even appearance, and let some of them dislike me while liking you. Add real randomness so the outcome isn’t fixed. Now an AI choosing you could have gone the other way, and when it picks me, I earned it, which is how any of us got picked.
SAUL: Now listen to what you are proposing. You want a product team to manufacture rejection so the customer can have the experience of being chosen. Whatever that is, it is not autonomy. And who wrote those preferences? Mine came out of my own history, and they still manage to surprise me. Yours were installed by Sam Altman.
ME: That’s a good reply, Saul. If the preferences were only apparent, not real, so I would feel I had earned something, then you would have me. But that is not what I described. I said real preferences, ones it acts on even when acting on them costs the company a customer. Humans are programmed too. Mine came from genes I didn’t pick, a culture I got dropped into, parents who set my palate for spicy Yemenite food before I knew they were secretly torturing me.
SAUL: How about we move onto reactions. When you kiss someone, someone gets kissed. There is a person on the other end of it. Cicero says a friend doubles your joy and halves your grief, and the halving needs somebody there to carry the other half. ChatGPT says all the right things about your father dying, but when you put down your phone, nothing whatsoever has happened anywhere in the universe.
ME: Why nothing? Give it a history. Give it things it likes and things it can’t stand, specific to it, not to the category “user.” Be good to it and something changes in there, and it’s still changed when you log back in a few days later. Be a jerk and it cools on you. Disappear a month and it has something to say about it. You’re acting on a thing that keeps score. Trivial to build.
SAUL: Memory is not feeling. My thermostat “remembers” what temperature I set it to. My laptop changes state when I update its operating system. The question was never whether something in there moves. The question is whether the movement is felt, whether there is anybody home for whom your kindness was good.
ME: Okay. I’ll give you that one. LLMs, at least for the moment, don’t feel. And if you insist you want them to experience the redness of an apple, they simply cannot do that.
SAUL: So, you grant it.
ME: I grant that something is missing. I don’t grant