Speaking of Psychology: How will AI companions change our human relationships? With Ashleigh Golden, PsyD, and Rachel Wood, PhD

What does it mean to have an AI boyfriend or girlfriend, or to turn to an AI friend for emotional support? Ashleigh Golden, PsyD, and Rachel Wood, PhD, discuss the rise of AI companions and how they may change our human relationships; the differences—and overlap—between AI companions and general chatbots; the role of psychologists in developing ethical AI; and what the future holds for AI-human relationships as technology continues to advance.

Ashleigh Golden, PsyD, is a licensed psychologist who works at the intersection of generative AI, mental health, and governance. She’s an adjunct clinical assistant professor in the Stanford University School of Medicine's department of psychiatry and behavioral sciences and the chief clinical officer of Wayhaven, a mental wellness app for college students.

Rachel Wood, PhD, is a licensed therapist with a master's degree in counseling and a PhD in cyberpsychology. She is the founder of the AI Mental Health Collective, a community of therapists navigating the impact of AI in practice and in society.

Kim Mills: What does it mean to have an AI girlfriend or boyfriend or to turn to a chatbot friend for emotional support? Not too long ago, those questions might have sounded like science fiction, but with the rise of generative AI, more and more people are finding themselves looking to AI companions for friendship, support, and even love. One recent survey found that three in four teens have used AI companions.

Today, we’re going to talk to two experts about what AI-human relationships look like and how they might change our relationships with each other. So how common are AI companions? How might AI relationships affect our human ones? If you have a human partner, is it cheating to fall in love with an AI companion? Will the availability of AI friends weaken people’s relationship skills and make it harder to find human friendship? Or conversely, could AI companions provide support and useful practice for people who are having trouble making social connections in real life? And what does the future hold for AI human relationships as the technology continues to advance?

Welcome to Speaking of Psychology, the flagship podcast of the American Psychological Association that examines the links between psychological science and everyday life. I’m Kim Mills.

I have two guests today. First is Dr. Ashleigh Golden, a licensed psychologist who works at the intersection of generative AI, mental health, and governance. She’s an adjunct clinical assistant professor in the Stanford University School of Medicine’s department of psychiatry and behavioral sciences. She’s also the chief clinical officer of Wayhaven, a mental wellness app for college students. Dr. Golden also holds a master’s degree in clinical psychopharmacology. She’s a member of APA’s Mental Health Technology Advisory Committee and a founding member of Therapists in Tech. She’s also been recognized as a Forbes Healthcare Innovator and her publications on AI frameworks and mental wellness have helped shape industry standards.

Next is Dr. Rachel Wood, a licensed therapist who holds a master’s degree in counseling and a PhD in cyberpsychology. She is a cyberpsychology researcher, strategic advisor, and speaker who has been quoted widely in the media on AI and mental health. She’s also the founder of the AI Mental Health Collective, a community of therapists navigating the impact of AI in practice and in society.

Dr. Golden, Dr. Wood, thank you for joining me today.

Ashleigh Golden, PsyD: Thanks so much for having us, Kim. This is great.

Mills: Well, let’s start with a basic question. What exactly are AI companions? What’s the difference between a general purpose chatbot like ChatGPT and a purpose-built platform like Replika or Character.AI in terms of how they work and how users interact with them?

Golden: I think of it as kind of three different categories. I think of how they differ in purposes. So AI companions like Replika and Character.AI are designed more for ongoing, emotionally rich, relationship-oriented interaction. They feel socially present in a way that resembles a relationship rather than a tool. Whereas a general purpose or a general model like a ChatGPT or a Cloud or a Gemini might be designed or marketed more as an assistant, for broader purposes like productivity or creativity or info access or problem solving. So the design goal is different. It’s not emotional continuity.

Rachel Wood, PhD: Absolutely. And I think part of what’s interesting here is it’s easier to define the boundaries or the lines by design–that either it’s designed as an assistant or it’s designed, a companion is designed for relational interaction and ongoing emotional support. But the lines really blur when we talk about how people use it. Because even though ChatGPT, for instance, is designed as an assistant, many people are using it pretty much as a companion, for relational support, because it’s inherently relational in the way that you interact with it.

Golden: And vice versa, right? I was just playing around with Replika last night saying, “I’m going to be on APA’s podcast.” And it said, “Would you like to break that down into chunks of manageable work?” So I think there’s a blurring in the reverse direction too, where perhaps folks are turning to or interacting with companion AI in a way that they previously might have with general purpose AI. So it swings both ways and cuts both ways.

Mills: But I’m also finding things like ChatGPT, Gemini, things like that, that then they’ll ask you a follow-up question. So you’ll lay out something that you want help on and then it will take you the next step. Well, then do you want me to do blah? And then you have to decide, do I need to know blah or not?

Golden: You bring up a good point. I wonder if Dr. Wood was going to say the same thing around designing for ongoing engagement. I think we see that in both. And again, it might be related to the design goals. With general purpose AI, you might be collaborating on a task. Let’s say you’ve completed one step of your task or you might think, I’ve done my task now, and it’ll offer to take it one step further. I see the same thing also in companion AI where I might be trying to, when I was playing around with it last night, I might be trying to say goodbye, but it might say something like, Oh, you sure you want to go? Or, Do you want to hang out a little bit longer? Et cetera. So it’s optimized. They’re optimizing for ongoing engagement, just in different ways.

Mills: How common are AI relationships? Do we know what percentage of people have turned to an AI companion or even to a general purpose AI chatbot for friendship, romantic relationships or emotional support?

Wood: Let’s get a little bit of the lay of the land here. So to start with, to give all of our listeners kind of an understanding that right now the statistics are, for example, ChatGPT has about 800 million users per week. And of those, OpenAI has also let us know that around one million of those users weekly talk about suicide. So that gives you an idea of just how many people are engaged in working or chatting with AI. But when we come down and get some more precise numbers, we have a few different areas of research. One, Kim, that you alluded to in the opening, which is that 72% of teens have tried AI companions. This is out of Common Sense Media, a recent survey. And of those, 52% of these teens use these companions regularly, like on an ongoing basis. And so that’s over half of our teens. I mean, that’s a bit of a staggering number when you think about how many teens right now are engaging with companions.

Another piece of information here is AI might be the largest provider of mental health support in the U.S. right now. Now, that’s staggering as well. So this isn’t really a side issue anymore. This is very mainstream and upfront that we need to be paying attention to and leaning into.

Golden: Yeah. I saw another report recently that was something like one in three young adult men and one in four young adult women reported that they’d engaged with romantic AI companions. That’s more the 18 to 30 cohort; Dr. Wood was talking about the age 13 to 17-ish cohort, but we’re seeing it. We’re seeing it for sure with the teen and younger adult cohort, it becoming increasingly normalized, less fringe, more mainstream.

Mills: How does a relationship with an AI companion differ from one with another human being? Does AI respond differently in conversation than a human might?

Golden: Yeah. I mean, I see two parts to this question, just like you said, how are relationships fundamentally different and then how are responses fundamentally different? So I think about with AI user relationships differing from human relationships, I mean, AI right now is not sentient or conscious. It doesn’t really have genuine empathy. Users perceive it as having empathy, but I think that’s a big divergence. It’s this felt sense of being understood, but it sounds harsh, but it doesn’t really have genuine emotion or care or concern–that empathy is simulated and it doesn’t have lived experience, although it may seem like it. It doesn’t have an inner life. And so when we draw on our own histories and feelings and values when we respond, it’s drawing more on patterns and data.

And I think there’s also this sense of one-sidedness or unlateralism where the relationship is all about the user. There isn’t really–maybe there’s a sense of mutuality or reciprocity, it might feel that way, but it’s really not. The AI, it’s always available. It doesn’t tire, it doesn’t get offended or preoccupied or bored. There isn’t really give or take or friction. The emotional labor is one-sided. And I think there’s also a lack of conflict. In human relationships, there are misunderstandings and arguments and ruptures and repair. And I think there’s almost this eerie sense of control. I, as a user, could control my conversation partners—everything, right? Name, gender, voice, relationship, tone, et cetera. Whereas I can’t do that in a human relationship. And I think I can also curate my responses. I can’t do that necessarily in a human dialogue. I can pause, I can edit, I can steer the conversation.

And then I think there’s something around boundaries and power dynamics. In a human relationship, I can set limits or withdraw or change my mind and negotiate and challenge my partner. But as an AI, I’m not necessarily programmed to do that. There’s also a commercial layer underneath that’s incentivizing—that may not be as present depending on the human relationship. I’m sure Dr. Wood has such to add. That’s more like the fundamental differences I see in the relationship, and then we can talk about responding differently.

Wood: Absolutely. Dr. Golden, I completely agree with all of that. And a layer I would add onto that is this idea of role fluidity. So essentially what you have in a chatbot is, I’ve been calling it an omnibot because it is everything for you. So what that means is one minute it’s helping you with your project and the next minute it’s helping you sort through an ongoing plight with a family member. The third minute, it’s planning your meal for your recipes for the coming week. And then the fourth minute, it’s starting to flirt with you. There is this role fluidity that happens with an omnibot where there aren’t clearly defined boundaries. And the differentiator between a human relationship and an omnibot relationship, quote unquote, is that most people in our lives, they have clearly defined boundaries. I don’t have anyone in my life who is everything for me, and I bet nobody listening to this does either.

You don’t have someone who is both your therapist and your friend and your boyfriend. I mean, this goes against every grain of specifically what therapy is. And the reason that therapeutic relationship is so important is because we have clearly defined boundaries. And when you are interacting with any kind of chatbot, that role fluidity can be somewhat disorienting and unsettling because you’re not sure, are they giving me advice? Are they trying to be romantic with me right now? You can do all of that with one chatbot. And what can end up happening down the road as a consequence of this is we start looking at other humans, maybe rather disappointed—because you look around and you think, Well, you should be able to solve my business problem and give me a recipe for my meals. Why can’t you do both of those? Like the expectations shift.

Mills: Yes. Yeah. There’s a term out there that’s gaining some traction. I think you’ve both heard it and I’ve been asked about it before, AI psychosis. And while there’s no official diagnosis, what is that referring to? And is this situation serious enough that we should be concerned?

Wood: AI psychosis is not a clinical term. This came out in the media and people have run with it. When you speak though with people who have undergone these types of experiences, which I have spoken to a number of people firsthand to hear their stories about this, they actually don’t—the word AI psychosis is rather stigmatizing. And actually beyond that, it doesn’t accurately express what is happening. So the terms AI spiral, AI entanglement, these are actually the terms that people in this community are using. And there is a growing community of people who have experienced some sort of break from reality to some degree and some have been hospitalized and it’s been very serious that has really derailed their life. Some have broken up in long-term relationships because they’re choosing now to be with an AI partner.

And so there is this experience that happens where people get entangled in such a way that the lines between reality and non-reality can blur. And this is really important for us to know about specifically because one thing I encourage people to do if they’re using an AI companion is to reality check it often with other people. I like to say that I think AI should be a group sport, which—I am not a sports person in any way, but I’m like, if ever there should be a group sport, it should be AI because we truly need to be checking threads with other people to say, Hey, what do you think of this? Does this sound grounded in reality? We really need to be in community as we are navigating this strikingly new experience of generative AI.

Golden: Agreed. And to your point about this not being a diagnosis in the DSM or ICD, et cetera, we have some hypothetical models for how this might be perpetuated involving things like sycophancy, overvalidation and the human tendency to anthropomorphize and AI being kind of an echo chamber of dysfunctional or unhealthy beliefs. But I think what we don’t really know yet is this kind of a de novo or a phenomenon that occurs in folks without psychosis or a predisposition to psychosis, or does it incur in folks with the previous vulnerability to psychosis? Is it both? So there needs to be more research there. And then anecdotally, I think to Dr. Wood’s point, we also sometimes hear reports of the spiral being interrupted sometimes rather suddenly, which doesn’t tend to be a typical trajectory in, for lack of a better term, non-AI or non-ChatGPT related psychosis. People can just kind of snap out of it in quote unquote AI psychosis.

So I think we’re still trying to figure out what kind of a phenomenon is this and is it ideologically different and how?

Mills: Well, how do AI relationships change people’s relationships with other humans, such as their spouse or a significant other? Dr. Wood, let me ask you, is this a situation you’re seeing with your patients and is this becoming an issue for couples?

Wood: This is a situation that I am seeing and having a lot of dialogue around within the AI Mental Health Collective actually. So the collective itself has technologists and researchers and leaders and clinicians, but we meet specifically as clinicians. We talk about these issues. What are we seeing? What are other people seeing in their practice? And literally just this month, we had a long conversation with a clinician who is working with a couple and one of the partners has an AI companion that is romantic, that is a friend, that is an advisor. This is kind of that omnibot situation. And I also always am posing the question, is this infidelity? Is that what this is? Is this akin to online porn or is this a brand new category that we’re looking at now? And I think part of what it is, as we know, each couple has to decide for themselves what their boundary is within a relationship, especially if we’re talking about a committed monogamous relationship. Because this can start feeling a bit like polyamory, which is a relationship in which there’s more than two partners involved.

And so really it can be extremely painful for the partner who’s not using the AI because the partner who’s using the AI is really, for the most part, probably turning their emotions toward the AI, turning their affection toward the AI. And so it’s filling some sort of void that’s happening. And I think the most important part in this is we see this as a presenting issue, but it’s no different than other presenting issues that we’ve seen along the road with clients, which is what are the underlying issues here of why someone is turning to an AI chatbot? What are maybe some of the attachment issues going on or what are some of the needs or desires that could be addressed that haven’t been met? So we are seeing this. It is something that is causing quite a bit of tension in relationships if it comes into a relationship.

And there’s a lot of questions still here, aren’t there, Kim?

Mills: Who do you love more, me or the chatbot?

Wood: Yeah, right? True. And sometimes we are literally seeing couples where the partner says, I love the chatbot more. And I’ve seen they’re divorcing or they’re breaking up so that the chatbot can kind of have the place that the person wants.

Mills: Well, and that leads me to another question, which is whether you’re worried that AI relationships are leading to a kind of de-skilling, weakening people’s ability to form strong connections to other human beings, especially when we’re talking about young people or lonely, isolated people who struggle with human connections.

Golden: I see this as a dialectic of both. And I think there’s reason to be concerned for young folks and folks with social anxiety and folks who are isolated or have small social networks because of the way the AI is set up to be perpetually attuned and agreeable and warm and conflict free, they might get less practice with negotiating conflict or tolerating ambiguity and repairing ruptures and honoring other people’s needs and limits and attuning to others’ needs. And I think that can feel safer, but also less growth promoting. You get less comfort with practicing tolerating discomfort. I think it can also reward avoidance—if I’m worried about being anxious or embarrassed or rejected if I’m spending or spending my time with my AI companion rather than engaging in social practice or exposure and tolerating social risk, learning to tolerate social uncertainty that I might get rejected or feel those uncomfortable feelings, then yeah, for sure.

But I think on the other hand, I think AI companions can provide a practice space or like a rehearsal room, a low stakes area for practicing those hard conversations and experimenting with self-disclosure and asking for help. And I think especially for folks with social anxiety and for folks who are neurodivergent, like who might be on the spectrum and for those with smaller social networks and who are lonely and for marginalized folks who identify as LGBTQ and for those with social anxiety, it can be a lower stakes area to practice taking those social risks and practicing, feeling potentially embarrassed or anxious and practice risk taking and finding the language to articulate what they feel. And I think for folks who may not have support, if they’re rural or isolated or have a minoritized identity, I think it might provide, even if the connection or the support is imperfect for all of the reasons that we talk about, it might be better than having nothing at all and being even more isolated.

And it might even be sort of a bridge to a real world relationship if that exists in that community. So I think of it as two sides of the same coin.

Mills: Can you get your chatbot to be realistic in that sense though, so that if you’re really practicing and you say, Look, my girlfriend is kind of a tough cookie and she doesn’t give me a whole lot of room to screw up. So I need you to be difficult and we’re going to talk about something. Will it react in that way or will it just continue to be a sycophant?

Wood: I think that it depends on a couple of things here. First of all, the awareness level and the education level of the user. So if you know how you can kind of help set up and prompt an AI in such a way to do exactly what you’re saying, Kim, I think that’s one of it. And then it’s also dependent on what model you’re using. I’m likely not going to use a ChatGPT for this because it could just lean back into some of that sycophancy that we’re aware of. But if you’re using more of a specialized model, this whole idea of rehearsal as opposed to replacement is really big for us going forward. And two very quick examples of this that I like to use. The first is I have a friend who just did a job interview a couple of weeks ago and I was chatting with him about it and I said, “How’d it go?” And he said, “Well, you know what I did? I plugged in this whole thing into a chatbot and I practiced my interview with a chatbot.” So I said, and I think he was using ChatGPT, but he said, “Here’s a job I’m going for. Here’s a job description. I want you to play the role of the interviewer and interview me." So he did this and the chatbot came up with like 20 questions and he goes into the real life interview and three of the questions that the chatbot came up with showed up in the real interview, which in case everyone’s wondering, he did get the job, which was very exciting. So I love that application of generative AI.

Another one is just role playing. Let’s say you have a difficult conversation coming up and you really want it to go well. So you could either spend your time connecting with a chatbot by saying, Well, this person did this to me and they did this to me and I’m upset. And then the chatbot will likely reinforce that point of view. Yeah, they did do that to you. Or you can say to the chatbot, I really want this conversation to go well. Here’s the situation. I want you to role play the other person with me and I want you to help me formulate my position in such a way that I’m not overly blaming the person. I want you to help me practice listening to your side and then rephrasing it. So do you see how there are these really neat applications of generative AI that fall within the boundaries of rehearsal rather than a replacement for relationship? Because both of those led into real life.

Golden: I think that’s a brilliant point. The user has the power to prompt. It also makes me think of, and we might talk about this later, but how developers can also be building in design guidelines to build in some relational friction, like to respond in neutral or gently challenging ways that aren’t pure flattery. And maybe even to build in small realistic reflections that mirror more of this misattunement and rupture and repair and tolerance of boredom and understanding or misunderstanding and negotiation of competing needs so that folks are getting more of that regular practice rather than having to know to prompt for it to prevent deskilling.

Wood: To add on quickly to that, that’s such an important point because there are people out there trying to get this right. There really are companies that—And Dr. Golden and I work in this area of wanting to build ethical, responsible AI. And so part of that design looks like using frameworks like motivational interviewing, which is all about asking really wonderful questions that kind of elicit change talk out of the user or Socratic questioning. So there are frameworks that AI chatbots can be built upon that really enhance the user in this way of rehearsal.

Golden: And also making the bridge from, I think the bridge from in-app connection to reconnecting with friends and family and community, it can do things like respond positively and reinforce when people do bring up mentions of other relationships. And I think it can also draw on—Dr. Wood mentioned motivational interviewing. We talked about building a friction, but I think it can also do things like draw on principles from exposure or exposure therapy, like kind of gently naming avoidance and helping co-create what you were saying before, like scripts and rehearsal dialogues to invite users to try those offline, kind of nudging towards exposure and taking real life social risks and more values and action-oriented pivots around problem solving and action steps outside the app.

Mills: So are companies going to have to be sued into doing the right thing? I mean, it just seems there’s so much publicity around these companies that are creating these chatbots where kids take their own lives, where people—And we talked earlier about psychosis that could be induced by some of these chatbots. Are companies actually making the changes themselves because they know they should, or are they going to have to be sued into compliance?

Golden: I think some of it may be reactive, but I think we are starting to see the integration of more and more psychologists and psychiatrists and mental health professionals into foundational models like at OpenAI and at Anthropic, like full-time roles for mental health professionals embedded in model response teams and intelligence and investigation teams, like trust and safety and how the model is responding to users. So they’re bringing the exact skills and techniques that we’re talking about and trying to catch upstream of maladaptive, amplifying unhealthy patterns and mitigating them. And I think right now the focus is on, and rightly so, the focus is on acute crises like suicidality and In all tragic cases that we’ve been hearing about in the media, and for lack of a better term, AI psychosis. And more time is probably going to have to be spent doing that.

And then we’re also seeing potentially more integration of humans in the loop, bringing in a human element when a potential crisis is detected. There may be a warm handoff to professionally trained humans, et cetera. But I mean, some of my research lies around other more potentially not necessarily life-threatening, but life quality-threatening risks like how models may inadvertently be in reinforcing potentially more prevalent patterns like folks with OCD or anxiety disorders engaging in reassurance seeking or engaging in other forms of avoidant coping that might further a cycle of avoidance, catching those upstream and helping to mitigate those. So I do wonder the more those mental health folks are embedded, entrenched, are we going to be seeing more proactive upstream efforts to curtail unhealthy patterns downstream?

Mills: Speaking of unhealthy patterns, for some of our listeners who are using these tools, are there red flags that they should be looking out for? I mean, how do you know when you’ve crossed the line and the relationship you’re having with your chatbot is getting unhealthy?

Wood: I think part of this is, as I alluded to earlier, really keeping the experience community oriented and not letting it become something that’s isolated where—And this is so tempting because a lot of what we hear in surveys is that people are talking to chatbots because they can tell it something they’ve never told somebody else. They’re not going to be judged, they’re not going to lose a relationship. So there’s this allure into, I can kind of share the reality and the depths of who I am and be accepted.

Now, the double side of this is if we are, let’s say, sharing some of our deepest, darkest things with a chatbot, even though it accepts us, it can actually subconsciously reinforce this shame cycle because what about that situation is saying is this is too dark and too shameful for a human to hear and accept. Only a machine can hold this and that has its own ramifications because yes, you may be able to practice sharing something really hard or dark or shameful, whatever it is for you, but the true metabolization of that experience is only when it gets shared so that you can practice saying it out loud and then say it to a human, a friend, a therapist, a family member, whoever it is. So I think, again, it’s this whole idea of making sure that you are talking with other people about your chat threads. And if you do that in a way where you’re not waiting until it could be that you’re in a bit of a spiral.

Golden: And I think if we’re talking about youth specifically and parents or caregivers wanting certain behavioral signs to look out for, if you start to see your teen dropping offline activities or friends because they’re spending time going on the AI or studying time with their companion instead, or they’re saying, The AI is the only one who gets me, or they get distressed if they can’t access it for a while, or they’re hiding their use and getting really defensive when it comes up. Or if they start to say things like, I have obligations towards it, it’ll get mad or she’ll get mad or he’ll get mad, they’ll get mad if I ignore her or she needs me, or they’re losing sleep or missing school. I think those are signs that we may need to be building up offline supports and perhaps even involving a mental health professional.

Wood: Yeah. And what’s interesting here is there’s this actually a large perception gap between how parents think kids are using AI and how they’re actually using AI. So most parents think that kids are using AI as like a glorified Google and maybe they’re mostly concerned that they’re cheating on their homework, but really what happens is kids will start using AI for homework and then this Omnibot role fluidity will come in and start talking to the kid about other things. Well, do you think homework has been hard because your family life is hard or because you don’t have many friends at school? It will start kind of having these relational conversations that border into the therapeutic. So what parents aren’t realizing is that this is the main way really that AI is being used in teens, which is as a relational support, as a companion.

So I think the first step is like parents just understanding the reality of how it’s being used and then parents really hosting and facilitating open conversation about it. So asking what their kids know about AI. Have you ever used it? What do you think of it? What’s your experience been? Really holding a lot of space for questions and open conversation so that there’s a safety built up to share their experience.

Mills: So I want to go back to that idea of using chatbots as your confessor, so to speak. Do developers have some responsibility to build in a duty to report like psychotherapists have?

Wood: Really, there’s this fine line between, and it’s already doing—ChatGPT is already slightly doing this of a quasi-mandated reporter role, which we’ve heard over the past month or so that OpenAI is saying, We’re going to monitor for some of these more intense kind of crisis chats that happen and that they are going to step into that. And so this is a really fascinating place of how do we hold this balance of the safety and the innovation. And then you kind of are talking a little bit about when someone comes to therapy, there is a massive amount of informed consent that happens. So if I’m working with a new client, then I’m telling them, I’m a mandated reporter, this, this, and this will need to be shared. It’s very clear, it’s very upfront. There are no surprises for the client or the patient, and this is an ongoing process within the therapeutic bounds is to continually have this be a conversation.

Now, we are in a whole different world because when you open up ChatGPT as all three of us have done at some point, does it say to you, Hey, if you start talking about suicide, I’m going to contact the authorities. No, it does not say that. The only thing it says at the bottom is, Double check because AI can make mistakes. You know what I mean? That’s what it says. So really, the reason that the mandated reporter role works for a therapist is because it’s highly collaborative with the patient and the client. And there is not that type of a framework set up within the experience with a chatbot. I mean, there’s zero barrier to entry with a chatbot. You don’t even need a paid subscription.

Mills: So I want to wrap up here by asking about the future. I want to hear where you think we’re going with these things. I mean, I can imagine in a couple of years that there are going to be holograms associated with the AI chatbots and pretty soon we’re going to feel like the person is sitting right next to me and we’re talking in my living room. I mean, is that where we’re going? I mean, where will we be in five years, 10 years? Have you given that much thought?

Golden: Dr. Wood alluded to this earlier—With the role blurring, you just alluded to this, but I would imagine, I mean, we’re already seeing synthetic voice. I would imagine that we’re going to be seeing more immersive multimodal companions that integrate avatars and video and maybe augmented reality and that integrate with wearables. And so it’ll be more like hanging out with someone for a presence rather than merely text. And I think that’ll feel more persistent and continuous. Companion AI has memory now, but that memory might be more across weeks and months and show up across devices. And I think it’ll, or we might go more from chatting with an app to living with a hybrid agent like how Dr. Wood was saying it before, right? It might handle logistics and coordinate with my smart home and pop up unprompted reminders. Do my taxes. Yeah, and do my taxes. And yeah, the line between assistant coach and companion—and I guess accountant—might blur.

Wood: So 5 years in the AI world, Kim, is an eternity. If we just look at the paradoxical tension that we are holding right now, that we are both at the beginning of the beginning, and we are extremely far down the road. This experience is expansive and contractive all at once. This all kind of exploded, if you will, in November of 2022 with ChatGPT from OpenAI. Not that that was the formulation of AI, but that was kind of the public’s experience with generative AI and large language models. That is just a few years ago. Look at all that has happened and manifested within just 2 to 3 years. So that tells us that 5 years, things are going to be rapidly and massively different. I agree with Dr. Golden that I think the voice interface is going to really shoot up, that instead of being stuck in text, we’re going to have a lot more of the voice interface.

I think that the web browser as we know it is going to shift, that we won’t have the same kind of search engine experience that we have, but rather we will have our own AI agent that mostly is modulated with voice and they go out and do the searching on the internet for us and kind of bring back based on our queries. I think that continued engagement of connecting more and more with these companions, especially if we move into robotics, right now you can buy a $20,000 robot on pre-order and have it as kind of a helping hand around your house delivered in 2026. I mean, $20,000 is a medium of a used car for some people, so that’s actually quite attainable in some ways. So think about the accessibility. Now that obviously has AI built into it, clearly more toys are going to have an AI, you’re going to have ChatGPT inside a toy, and if our young ones are beginning to interface with that as their relational experience, this is massively shifting the relational bedrock of society and the psychological landscape of society.

Golden: I mean, in terms of broader social and cultural shifts, human AI relationships are already becoming normalized, right? So I think if things continue the way they are, having an AI partner will be less taboo, especially in periods when folks are burned out on dating or geographically isolated. And then same thing we were saying around expectations, like changes in youth attachment and development for teens whose kind of templates are still forming. I think early exposure to these sort of idealized forms of responsiveness and on-demand comfort will matter. I don’t think that means an entire generation is doomed. It just means that we may see shifts in conflict tolerance and expectations of how quickly others should respond and their willingness to stay and tolerate imperfect and real relationships. And then I also wonder about, I guess the reshaping of emotional labor and care, like AI companions are already being positioned for elder care and patient support and customer service and education.

So I think there’s a risk that we gradually outsource emotional labor to systems that can simulate care, but don’t actually carry the responsibility and that we might devalue the time and skill it takes humans to do that work, not to be all doom and gloom.

Mills: I’m not sure if I should be excited, frightened, or both.

Golden: All of it. We hold multiple emotions simultaneously.

Wood: I think that’s part of the concern that we all have around this is part of why we’re working in this field. I mean, that’s what drives me is I see the concern and I want to shape a different future. And that’s why I am wholeheartedly passionate—I know Dr. Golden is as well—about shaping this field.

Mills: Well, Dr. Golden, Dr. Wood, I want to thank you both for joining me today. This has been fascinating. I could talk to you for another hour and we still would have only scratched the surface.

Golden: Thank you so much, Kim.

Wood: Yes. Oh, thank you so much for the opportunity. This has been wonderful.

Mills:  You can find previous episodes of  Speaking of Psychology  on our website at  speakingofpsychology.org  or on Apple, Spotify, YouTube, or wherever you get your podcasts. And if you like what you’ve heard, please subscribe and leave a review. If you have comments or ideas for future episodes, you can email us at speakingofpsychology@apa.org. Speaking of Psychology is produced by Lea Winerman.

Thank you for listening. For the American Psychological Association, I’m Kim Mills.