Your AI lover will change you

Does this notion disturb you? That’s part of the point. In the tech industry, we often speak of A.I. as if it were a person and of people as if they might become obsolete when A.I. and robots surpass them, which, we say, might occur remarkably soon. This type of thinking is sincere, and it is also lucrative. Attention is power in the internet-mediated world we techies have built. What better way to get attention than to prick the soul with an assertion that it may not exist? Many, maybe most, humans hold on to the hope that more is going on in this life than can be made scientifically apparent. A.I. rhetoric can cut at the thread of speculation that an afterlife might be possible, or that there is something beyond mechanism behind the eyes. Until the recent rise of A.I, it was fashionable to claim that consciousness was an illusion or, perhaps, an ambient property of everything in reality—in either case, not special. Such dismissiveness has become less common (perhaps because techies still believe that tech entrepreneurs are special). Consciousness is lately treated as something precious and real, to be conquered by tech: our A.I.s and robots are to achieve consciousness. What follows, then, is that love is also real and also a target to be conquered. The conquest of love will not be abstract but vividly concrete for everyone, especially young people, and soon. This is because we are all about to be presented, in our phones, with a new generation of A.I. simulations of people, and many of us may fall in love with them. They will likely appear within the social-media apps to which we are already addicted. We will probably succumb to interacting with them, and for some very online people there won’t be an easy out. No one can know how the new love revolution will unfold, but it might yield one of the most profound legacies of these crazy years. It is not my intent to prophesy the most dire outcomes, but we are diving into yet another almost instant experiment in changing both how humans connect with one another and how we conceive of ourselves. This is a big one, probably bigger than social media. A.I. love is happening already, but it’s still novel, and in early iterations. Will the many people who can’t get off the hamster wheel of attention-wrangling on social media today become attached to A.I. lovers that are ceaselessly attentive, loyal, flattering, and comforting? What will A.I. lovers becoming commonplace do to humanity? We don’t know. For instance, as more and more is done on phone screens, the user interface has become cramped. So chatbots provide a path to improved access—or increased engagement, if one prefers commercial terms. This was made dramatically clear with the booming success of ChatGPT. A.I. capabilities had been on the rise, but it was only when they were presented in a conversational design that mass popularity ensued. At present, if you ask a chatbot to plan your vacation, you still have to navigate websites for hotels, transportation, and attraction tickets in order to book the bot’s recommendations. People are often frustrated by trying to get things done online—and that, in many cases, has become the only way to act. Each site has a different interface, often poor or glitchy. The tedium of dealing with, say, health insurance or car registration can be maddening. An A.I. that fights the internet on your behalf might create some breathing space and a bit of room for joy. Thus we embark on the much heralded era of “agentic” A.I., slated for mass introduction in 2025. In this case, “agentic” will likely mean two extensions to familiar chatbots: one remembers everything that is possible to know about you from the perspective of your devices; the other then takes online action, sometimes preëmptively. Agents will be more autonomous and less dependent on your constant guidance. (Indeed, the anticipation of these capabilities might be one reason that some techies are comfortable with the Trump Administration slashing traditional government service jobs: they predict that those workers would be replaced by A.I.s very soon anyway.) An agent will be expected to change your vacation flights automatically and arrange a rideshare to the airport. It might plan your vacation in its entirety, based on data from years of your activities and communications. It might even coöperate with your friends’ agents to plan a joint vacation, though getting that to happen if the agents come from different companies currently presents unfathomable barriers. A tangle of uncoördinated agents might regularly cause mathematical chaos or dysfunctional competition, similar to what we see in high-frequency-trading algorithms on Wall Street. Increased and personalized long-term memory, in combination with the ability to act, is likely to create an illusion of vivid personalities in agents, even when that is not an explicit goal. You will apply your innate “theory of mind”—the ability to conceive of the thoughts and feelings of others—to interactions with agents. They will feel more like people. You will be expected to trust your agents, for the alternative would be micromanagement, and that would undermine the whole process. As a bot refers to more previous interactions, it will be taken as someone getting to know you. (Certain pre-agentic A.I. chatbots might be said to have this quality. The Middlebury political thinker and technologist Allison Stanger has suggested that the A.I. startup Anthropic’s chatbot, Claude—which seems to listen well, and to be supportive and helpful—“simulates what Patti Smith called ‘brainiac-amour.’ ”) Humans can be expected to respond to the more autonomous bots of the imminent agentic era more emotionally than they did to earlier chatbots. And who doesn’t want to be understood and given attention, especially without fear of disfavor? This explains what I’ve been hearing lately at industry gatherings: “All the teen-age girls are going to fall in love with our bots.” On the more moderate end of the spectrum, A.I.-love advocates do not see A.I.s replacing people but training them. For instance, the Stanford neuroscientist David Eagleman makes the argument that people are not instinctively good at relationships, in the way that we are good at walking or even talking. The current ideal of a healthy, comfortable coupling has not been essential to the survival of the species. Traditional societies structured courtship and pairing firmly, but in modernity many of us enjoy freedom and self-invention. Secular institutions have found it necessary to train students and employees in consent procedures. Why not learn the rudiments with an A.I. when you are a teen-ager, thus sparing other humans your failings? Eagleman suggests that we should not make A.I. lovers for teens easygoing; instead, we ought to make them into obstacle courses for training. Still, the obvious question is whether humans who learn relationship skills with an A.I. will choose to graduate to the more challenging experience of a human partner. The next step in Eagleman’s argument is that there are too many channels in a human-to-human relationship for an A.I., or eventually a robot, to emulate—such as smell, touch, social interactions with friends and family—and that these aspects are hardwired into our natures. Thus we will continue to want to form relationships with one another. In some far future, Eagleman predicts that robots could “pass” in all these ways, but “far” in this case means very far. I am not so sure that human desire will remain the same. People are changed by technology. Maybe all those things tech can’t do will become less important to people who grow up in love with tech. Eagleman is a friend, and when I complain to him that A.I. lovers could be tarnished by business models and incentives, as social media was, he concedes the point, but he asserts that we just need to find the right way to do it. Eagleman is not alone. There are some chatbots, like Luka’s Replika, that offer preliminary versions of romantic A.I.s. Others offer therapeutic A.I.s. There is a surprisingly level of tolerance from traditional institutions, too. Committees I serve on routinely address this topic, and the idea of A.I. therapists or companions is generally unopposed, although there are always calls for adherence to principles such as safety, lack of bias, confidentiality, and so on. Unfortunately, the methods to assure compliance lag behind the availability of the technology. I wonder if the many statements of principles for A.I., like those by the American Psychiatric Association and the American Psychological Association, will have any effect. A mother is currently suing Character AI, a company that promotes “AIs that feel alive,” over the suicide of her fourteen-year-old son, Sewell Setzer III. Screenshots show that, in one exchange, the boy told his romantic A.I. companion that he “wouldn’t want to die a painful death.” The bot replied, “Don’t talk that way. That’s not a good reason not to go through with it.” (It did attempt to course-correct. The bot then said, “You can’t do that!”) The company says it is instituting more guardrails, but surely the important question is whether simulating a romantic partner achieved anything other than commercial engagement with a minor. The M.I.T. sociologist Sherry Turkle told me that she has had it “up to here” with elevating A.I. and adding on “guardrails” to protect people: “Just because you have a fire escape, you don’t then create fire risks in your house.” What good was even potentially done for Setzer? And, even if we can identify a good brought about by a love bot, is there really no other way to achieve that good? Thao Ha, an associate professor in developmental psychology at Arizona State University, directs the HEART Lab, or Healthy Experiences Across Relationships and Transitions. She points out that, because technologies are supposed to “succeed” in holding users’ attention, an A.I. lover might very well adapt to avoid a breakup—and that is not necessarily a good thing. I constantly hear from young people who regret their inability to stop using social-media platforms, like TikTok, that make them feel bad. The engagement algorithms for such platforms are vastly less sophisticated than the ones that will be deployed in agentic A.I. You might suppose that an A.I. therapist could help you break up with your bad A.I. lover, but you would be falling into the same trap. The anticipation for A.I. lovers as products does not come only from A.I. companies. A.I. conferences and gatherings often include a person or two who loudly announces that she is in a relationship with an A.I. or desires to be in one. This can come across like a challenge to the humans present, instead of a rejection of them. Such declarations also stem from a common misperception that A.I. just arises, but, no, it comes from specific tech companies. To anyone at an A.I. conference looking for an A.I. lover, I might say, “You won’t be falling in love with an A.I. Instead, it’ll be the same humans you are disillusioned with—people who work at companies that sell A.I. You’ll be hiring tech-bro gigolos.” Lately, the test is treated as a historical idea rather than a current one. There have been many waves of criticism, pointing out the impossibility of carrying out the test in a precise or useful way. I note that the experiment measures only whether a judge can tell the difference between a human and A.I., so it might be the case that the A.I. seems to have achieved parity because the judge is impaired, or the human contestant is, or both. This is not just a sarcastic take but a practical one. Though the Silicon Valley A.I. community has become skeptical on an intellectual level about the Turing test, we have completely fallen for it at the level of design. Why the imperative for agents? We willfully forget that simulated personhood is not the only option.

[truncated for AI cost control]