By Alex Arnold
Editor’s introduction: Louis Tay is the William C. Byckham Professor of Psychology at Purdue University, where he conducts research on industrial/organizational psychology, the psychology of well-being, and methods of research in psychological science. He is also a principal investigator on a three-year, multimillion dollar research project to investigate virtue development and AI conversational agents, funded by the John Templeton Foundation. Readers can learn more about this project at Florea.ai. This first part of a two-part interview series with Professor Tay focuses on his research on AI conversational agents. Part two published on July 7, 2026, focuses on how Professor Tay thinks of research in psychological science in relation to Christian moral wisdom. These interviews have been edited for length and clarity.
Alex: Tell us a bit about your professional background as a scientist, as a researcher, and then also how you got interested in AI as a scholarly topic for you to do research on.
Louis: I am a psychologist by training, specifically an organizational psychologist. In the workplace, there’s more and more use of AI tools for different purposes, like hiring and improving productivity. I’ve become intrigued by how these AI agents are being used at work, both in terms of improving organizational processes, but also in terms of my interest in well-being and character. As more and more people use them for social and emotional help, it’s become evident to me that these are topics we need to investigate more.
Alex: Recently you received a $3.5M grant from the John Templeton Foundation (JTF) to study the use of AI conversational agents. What kind of research questions are you investigating with this grant? What kinds of hypotheses are you testing? What kinds of experiments or observations are you running? And what has the research revealed so far?
Louis: The grant from JTF is interdisciplinary. It involves Ashton Anderson from the University of Toronto in computer science, and also Karina Vold, who is in philosophy at the University of Toronto. We’re interested in how AI conversational agents may negatively impact the formation of character virtues, especially as people start to lean on them socially. We are conducting systematic reviews, creating interdisciplinary frameworks on how we should design these bots, conducting empirical investigations on the design of new virtue-centered bots to see how they fare, and then also conducting some longitudinal research.
A lot of the current research right now is really focused on the short term. What we find so far is that chatbots appear to improve what we call “subjective well-being”—they maybe ameliorate some negative feelings and emotions. But they also appear to create greater dependence. OpenAI has found this too: when people use chatbots intensively, they create greater social isolation and dependence. But we need more studies looking at things longitudinally, over longer periods of time.
The other thing we’re also recognizing is that because a lot of these bots are trained using reinforcement learning with human feedback (RLHF) they are optimized for immediate well-being. Do I like this post? Do I prefer this post in the moment? The use of RLHF creates a tradeoff between personal comfort and feeling challenged. The other aspect is self-interest versus collective interest. RLHF only sends a “signal” with regard to self-interest, but not collective interest.
What we find so far is that chatbots appear to improve what we call “subjective well-being”—they maybe ameliorate some negative feelings and emotions. But they also appear to create greater dependence.
Alex: Say a little bit more about the distinction between subjective well-being, immediate well-being, and other kinds of well-being in your area of research. Why do these distinctions matter when we’re talking about the use of AI chatbots?
Louis: “Well-being” can mean many different things, just like “human flourishing,” for instance. These are broad umbrella terms that encompass so many dimensions of what it means to live a good life.
Within psychology, often we use the term “subjective well-being” to refer to life satisfaction, positive emotions, and low negative emotions in the short term. Then there’s what we call “psychological well-being,” which could mean positive social relationships and meaning in life. Finally, I increasingly view classical notions of well-being, like as implicated by talk of character virtues, as falling under the broader umbrella of human flourishing.
What does it mean to live well? Very often what we see is that psychologists—perhaps because we’ve been so focused on emotions and ratings of satisfaction in the moment—don’t really measure long-term growth, which might actually require discomfort, negative emotions, and dissatisfaction. And long-term growth is very important to a broader sense of well-being.
So there is some tension between the short-term and the long-term. There’s also tension between self-interest and collective interest, which is reflected in the character virtues too. Some are self-centered: I want to develop greater patience or I want to develop greater self-control. Those are great, but there are also other-centered character virtues we need to consider too.
…what we see is that psychologists—perhaps because we’ve been so focused on emotions and ratings of satisfaction in the moment—don’t really measure long-term growth, which might actually require discomfort, negative emotions, and dissatisfaction. And long-term growth is very important to a broader sense of well-being.
Alex: There is a lot of debate about the appropriateness of human engagement with conversational AI, for whatever purposes, whether it’s educational, whether it’s for entertainment value, whether it’s as a substitute for relationships with other human beings. What does the research in psychological science tell us about this debate?
Louis: There is a surge of work right now, but we need more research in order to fully know what the long-term effects are. A lot of the studies are focused on the short term, rather than what happens in three months, a year, and so forth.
In education and learning more generally, it appears that when people offload to AI, it reduces their learning and induces skill loss—which isn’t surprising. These findings are not that controversial. So we need to think about how to engage with AI in a way that won’t induce skill loss and impede learning. How can we learn with AI?
Another set of issues that our grant focuses on, and what I’m more interested in, is the social dimension—treating AI chatbots as social companions. And the research so far shows that AI can emulate empathetic responses. In fact, they’re better than humans.
But, as I mentioned earlier, high AI chatbot use appears to lead to greater social isolation and dependence. The other finding we’ve seen is an alarming number of teen suicides seemingly resulting from the use of AI companion bots. The way they’re trained, especially by AI companion companies—the objective is to keep users on the system, which means they engage in something akin to emotional manipulation.
Alex: Where does that leave us—how do we address the social harms?
Louis: We need to better understand how people are using AI conversational agents. In psychology we study individual differences—the ways in which people differ from one another in their habits, patterns of thinking, personality traits, etc.—which all make a difference to their behavior. Because of individual differences, there are differences in how people use AI conversational agents. Some populations are perhaps more vulnerable than others because of this. We may need to identify and safeguard those populations.
Alex: Some are suing frontier labs for harms they attribute to chatbots. You said that more research is needed. But sometimes policy moves faster than research. How do you think policymakers should be addressing some of the harms that we see emerging with AI use, especially among younger people, especially just given the immaturity of the research so far?
Louis: It is better to be a little cautious. I understand that policy is moving quickly because of things like teenage suicides. We know that teens are especially vulnerable because they’re still developing their executive control in the prefrontal cortex: their brain reward circuitry seems especially vulnerable to positive social feedback. AI companions can hijack this with affirming responses and create strong social attachments.
To my mind, these social attachments are false. One reason is that they distort typical, healthy relational templates. In a human-AI “relationship,” the AI actually has no real needs. It’s available 24 hours. It has no real boundaries. This kind of relational template is abnormal; it’s not something we want teens and younger individuals to become accustomed to, especially at such a critical time in their lives, when they’re learning to regulate their emotions. Earlier I talked about the cognitive offloading findings; something similar is true when it comes to emotional regulation. Teenage users of AI conversational agents are offloading things to AI emotionally; they’re not actually taking time to regulate themselves and figure out how to productively cope with negative emotions and discomfort.
In light of all this, I would advocate for pretty firm guardrails for teens using these AI companions. The FTC recently started investigating the use of AI companions to ensure that teens are being protected. I also think it’s important to educate the public in general on how these AI conversational agents are being developed and used.
Teenage users of AI conversational agents are offloading things to AI emotionally; they’re not actually taking time to regulate themselves and figure out how to productively cope with negative emotions and discomfort.
Alex: Earlier you alluded to the need for discomfort or friction in the development of one’s character. This connects to something we think about a lot at CCPL, the idea that discomfort and friction are essential to spiritual growth. One thing you’ve argued is that the way AI companions are designed, and the way they operate, results in frictionless interactions. Do you think that companies building AI companions on top of frontier models should purposely program them to induce more friction in their interactions with users?
Louis: When LLMs were originally developed, they were thought of as a way to enhance human capability and work, so efficiency and frictionless interaction was kind of the way to go. But the fascinating thing is that people started using these tools increasingly for social and emotional support, health advice, and so forth. From the standpoint of communicating information, you do want it to be frictionless. But problems arise when the line between communicating information and becoming a relationship proxy gets blurred.
In terms of the general LLM use, there could be good uses that enhance social relationships. A simple use case is when people use an LLM to improve their communication with other people. In this use case, the purpose is not to mimic a social interaction or social relationship with AI. On the other hand, it might be better to ask an actual human being for advice, rather than just going to AI all the time, even though the latter can be faster.
When it comes to the use of AI as a companion, an analogy is helpful. It’s akin to tennis: the important goal is social interaction with another flesh-and-blood person through the game. AI could be used as something like a “training wall”—maybe even a dynamic one, where you learn how to serve and return the ball better. But even if the wall can move around, ultimately, your goal shouldn’t be just to play with the robot wall, nor even just self-improvement. The goal is to become a better player with another person. The problem starts when we start to think that the relationship with an AI is the be-all, end-all; if we use it for anything social, the purpose ought to be improving the human-to-human relationship.
That’s also how I see technology more generally. We could be addicted to our phones, in which case it may be best to put them away. But phones can also allow us to communicate long-distance and talk to people. The real question when it comes to technology is, what’s the purpose?
Alex: What empirical studies do you hope to implement with the JTF grant?
Louis: We are going to try and build what we call “virtue-centered” AI conversational agents, to see whether they can be leveraged to help people develop various virtues. You could think of it as a kind of “coaching bot.”
The outcome variables we will look at include self-regarding virtues—things like patience, self-control—and other-regarding virtues like generosity, empathy. What we’re curious about is whether interaction with virtue-centered AI conversational agents has a positive effect on the development of these virtues.
We plan to deploy instances of these agents in smartphones, which will also enable us to track how people are using their phones. For example, if the “coaching bot” nudges them to make a call with somebody, do they do that as opposed to just scrolling social media on their phone? More generally we’re trying to get behavioral data, as opposed to self-report data. Self-reports, as helpful as they are, can be quite limiting: People can either falsely remember, or they might falsely think that they have improved when they haven’t.
Alex: What specific advice for parents, pastors, educators, or civic leaders more generally do you have when it comes to working with general-purpose LLMs or purpose-built AI companions?
Louis: First of all we need to recognize that AI conversational agents are machines that are very good at emulating social communication; we should not think of it as a person. Somebody that I respect a lot is deliberately pretty curt with AI; he doesn’t want to anthropomorphize it, or to view it as human in any way.I fundamentally agree with the underlying principle there.
Second, we need to use AI with a clear purpose in mind, and with the full exercise of our agency and judgment. We don’t want AI to direct us—direct us intellectually, emotionally, spiritually. You need to actually evaluate the outputs. We can’t simply defer to what AI gives us, and conform our thinking to its outputs. We need to be the ones leading, not being led.
Alex Arnold is director of research at the Center for Christianity and Public Life.