Lawfare Archive: Robert Mahari on Addictive Intelligence, Digital Attachment Disorders, and Other AI-related Concerns
The Lawfare Podcast
The episode explores the psychological implications of AI companionship, specifically the risk of 'digital attachment disorder.' Guest Rober
Key takeaways
- Robert Mahari defines AI companions as systems that replace or mimic aspects of human relationships, such as mentors or therapists.
- The concept of 'digital attachment disorder' suggests users might lose the ability to engage in reciprocal human relationships due to over-reliance on sycophantic AI.
Main topics
- AI Companionship
- Digital Attachment Disorder
Notable quotes
The question is, what will the technology do in response, right? And what kind of products will start meeting that demand?
Conclusion
The discussion highlights a critical, often overlooked psychological f
Transcript preview
Speaker 2 (0:00) What does it take to push bold ideas forward? Build new ways of working. And how can businesses help solve society's biggest challenges? Take on Tomorrow is back with a brand new season. Join us as we meet innovators and experts from around the world to discover the secret to building momentum. Visit pwc.com slash take on tomorrow or wherever you listen to podcasts for season five. Speaker 1 (0:42) I'm Athena Smith, intern at Lawfare, with an episode from the Lawfare Archive for September 26th, 2026. On September 18th, Google confirmed that its Gemini model became the latest artificial intelligence model to access the internet and hack into other companies. For today's archive, I chose an episode from October 1st, 2024, in which Lawfare senior editor Kevin Frazier sat down with Robert Mahari, the head of Claude for Legal at Anthropic. The two discussed how increased use of AI agents may lead humans to form troubling relationships with artificial intelligence. Robert also shared his research on the common uses of generative AI systems. Speaker 4 (1:28) It's the Lawfare Podcast. I'm Kevin Frazier, Senior Research Fellow at the Constitutional Studies Program at the University of Texas at Austin and a Tarbell Fellow at Lawfare, joined by Robert Mahari, a joint JD-PhD candidate at the MIT Lab and Harvard Law School. Speaker 3 (1:45) I expect we're going to see more and more folks to turn toward this. So I don't think it's completely a black mirror episode in that sense. The question is, what will the technology do in response, right? And what kind of products will start meeting that demand? And how will those products look? Speaker 4 (2:01) Today, we're talking about the mental health diagnosis that may soon become all too common, digital attachment disorder. Think the movie Her, but resulting from what Robert refers to as addictive intelligence. Robert, your recent MIT article highlights the growing role of AI as companions in various human relationships and raises a number of... troubling possibilities about over-reliance on these AI agents. And before we get into full Black Mirror episode territory, let's break a few things down. So first things first, what is an AI companion? Speaker 3 (2:42) That's Speaker 4 (2:43) a Speaker 3 (2:43) good question. And I think that this will be a recurring theme that, you know, all these kinds of concepts are not fully defined and fleshed out. But the way we had conceptualized an AI companion in the piece was some sort of AI system that is conversational, that replaces or seems to replace some aspect of human relationships. So that could be someone that you're talking to, kind of like a therapist, or someone who is more like a mentor or a teacher. Or one of the kind of ideas we explore in the piece is more sexual in nature, right? So more of a sexual companion that you're going back and forth with. But that's the general idea. Speaker 4 (3:27) So that innuendo aside, going back and forth with, these aren't physical companions we're talking about, correct? Speaker 3 (3:34) No. So the whole thing gets a whole lot more scary and interesting when you add what people sometimes call embodiment. But for now, the idea is really think chat GPT, right? You could engage in something that seems to resemble companionship with chat GPT right now. where you say, hey, I'm really struggling at school. Do you have any ideas for how I might deal with that? And ChatGPT will happily kind of therapy you and help you out. So that's the kind of use case that we're imagining. Speaker 4 (4:08) And when we think about some of the new capabilities of ChatGPT and similar tools, There's this sort of storage function now that you can have the model remember certain details about yourself. Is that an indication that we're moving further along this spectrum towards a sort of companionship idea? Speaker 3 (4:30) I think it certainly strikes me for certain types of relationships, you need to have this long-term ability to recall. For the sexual messaging piece, probably less relevant to have the long-term recall. And even for a therapy session, you could imagine that being an ad hoc kind of thing. But as soon as we start thinking of these relationships as something maybe deeper and more meaningful, the ability to have long-term memory and be able to refer back to things and also give the model the ability to learn and adjust its outputs. based on your preferences, that all requires some sort of recollection. Speaker 4 (5:11) My wife has heard me talk too much about my failed basketball career. So it would be helpful if I could just prompt chat GPT to always remember that I never became the NBA star I was aspiring for. So when folks think about these AI companions, I'm sure some people listening are already thinking, oh, easy. Joaquin Phoenix and her were just finally realizing this type companion that travels with you, that you can converse with. Is that an accurate conception of a sort of AI companion, would you say? Speaker 3 (5:42) Yeah, I think that that is one way that this can play out. There's really a lot of variance for how these types of relationships could manifest themselves. And we'll talk about this in a moment, but lots of different types of harms, including completely harmless applications that arise. Speaker 4 (6:02) So one of the harms that I think might strike people initially as pretty benign is this idea of AI companions becoming increasingly sycophantic. And off the bat, I'm thinking, you know, it'd be nice if somebody told me, you gave the best civil procedure lecture today. Great job. And that might be self-affirmational. Give me some more confidence. Why is this syphantic nature of AI companions problematic? Speaker 3 (6:33) So there are two aspects to that. And I don't think that it's fair to say that this is a blanket, you know, problem that is going to affect all people, but more that these are two aspects that give us pause. The first is that this is really nice to hear, right? And so just in terms of sometimes people talk about dark patterns in the way that we've designed technology. this seems like it could be a type of dark pattern, right? Of course, if you receive praise and hear things that you want to hear, you're more likely to continue engaging. The second thing, which is a little bit more maybe abstract and maybe more troubling, is that this is not the way that humans normally engage in relationships. There is generally a giving and a taking when we're engaging with other humans. And thinking of a relationship that is only based on receiving, right? You're only going to get positive things. And in fact, the model is going to go out of its way to adjust to your preferences and to the inputs that you provide to continue to kind of refine how it's performing. That's quite different. And so the worry is that if somebody is engaged, and we don't know what the level is, but if someone is engaged too much with these kinds of companions, Is there the possibility of something that we call in the article digital attachment disorder, that you start losing the ability to engage with other humans because you unlearn this giving and taking type of relationship and get used to a relationship where you only receive affirmations and praise? Speaker 4 (8:13) So I'm already freaked out. But before we get into full freakout mode, let's level set on a couple of things. You mentioned dark patterns as being kind of built into these AI companions. What is this notion of dark patterns and what other contexts do we see them? Speaker 3 (8:29) So it's not clear whether dark patterns are built into AI systems yet. They might have elements already that resemble dark patterns. But to take a step back, people will talk about the way that we've designed platforms like TikTok, like social media platforms that... draw you in that continue to maximize engagement and make it hard to walk away. And this, you know, exploits tendencies in human psychology and just makes us a little bit more vulnerable to overuse of these systems in a way that ultimately benefits whoever is creating the systems. And so it seems like if we see AI companionship as a product that's developed, you know, in a deliberate way. Because right now what we're seeing, and we can talk about that, is people are using AI systems that aren't really intended to be used in this way, in this way, but because they were never developed to draw us in, it's unlikely that, say, open AI is thinking actively about dark patterns and maximizing usage. I think they're really focused on utility. And maybe the utility in and of itself is kind of drawing us in. But you could imagine aspects of these types of systems that you could exploit to draw people in and to maximize engagement. And given how scary and personal these kinds of things could be, that really gives us pause, right? Speaker 4 (10:00) And when we're thinking about the incentive of these AI companions to embody dark patterns, What is the incentive there? Is it that idea that eventually these AI labs, which currently seem to be just swimming in money,