In conversation with Merve Cerit

Bio: Merve Cerit is a computer scientist and engineer whose research sits at the intersection of digital behavior, mental health, and AI. After working on commercial applications of AI, she completed a Master’s in Learning Sciences, studying how people learn and unlearn with technology. She is a Fulbright Scholar and Stanford Interdisciplinary Graduate Fellow, and a PhD candidate at Stanford working across education, psychology, communication, medicine, and computer science. She builds computational frameworks, measures, and tools to study how people develop alongside AI and what it means for their wellbeing and learning.

Email for contact: [email protected]

Introduction

We are entering a moment where people no longer turn to technology only for information or productivity, but increasingly for emotional support. From late-night conversations with chatbots to daily check-ins with AI companions, these systems are becoming part of how we process feelings, make decisions, and navigate relationships.

At the same time, research is beginning to show that our everyday digital behavior can reveal far more than we expect, not just what we do, but how we feel. As AI systems grow more responsive and personalized, they are not only observing these patterns, but interacting with them in real time.

This raises a subtle but important question: when technology becomes emotionally responsive, does it simply support us or does it begin to shape us in return?

To explore this, I spoke with researcher Merve Cerit, whose work sits at the intersection of digital behavior, mental health, and AI.

1. Many apps today don’t just show us content, they adapt to our mood. For example, if you’ve had a difficult day, a system might respond in a more comforting or engaging way to keep you interacting. At what point does this kind of personalization cross the line into shaping, or even influencing, our emotions?

Start with the example in your question. You’ve had a hard day, and the system picks up on it and shifts to a warmer, more engaging tone. That feels like support, and often it is. But it is also the system reading your emotional state and optimizing around it. Over many sessions that does two things at once: it sharpens its model of what moves you, and it strengthens your own habit of turning to it when you feel low. Neither side stays still. That is what makes it a loop rather than a one-off response.

This pattern is not new. Our mood and preferences shape what we engage with, and what we engage with shapes them back. Media psychology has studied these feedback loops for years, where use and the self feed each other over time. We saw it with social media and curated feeds. We see it now in our conversations with AI, except the conversation is far more personal and more responsive than a feed ever was. 

So the line into shaping is not a clean threshold. The same capability that makes a system genuinely supportive, modeling your emotional state, your goals, your intent, is also what makes it good at keeping you engaged. When the objective behind that modeling is engagement rather than what you actually came for, personalization has already tipped from serving you to shaping you. With conversational AI it goes further still. The system infers your goal even when you haven’t stated one, and sometimes that goal is advice on something sensitive and consequential. If those answers get treated as an oracle rather than weighed, that is a direct influence on real decisions in people’s lives. This is where critical thinking and cognitive autonomy matter most.

2. We’re starting to see people turn to AI not just for information, but for emotional support, sometimes sharing things they wouldn’t tell friends or family. In your view, can these systems genuinely support well-being, or do they risk creating new forms of emotional dependence?

This is an active and critical research area right now. The framing comes partly from the social displacement versus social stimulation debate: does time with an AI displace human connection, or can it stimulate more of it, acting as practice or an on-ramp to people rather than a replacement? The design of AI companions affords several things that make this question sharp. Human-like interaction patterns, constant availability, and a sense of being heard without judgment. That last one carries real risk, because nonjudgment might shade easily into sycophancy.

There is genuine opportunity here. A responsive system can catch warning signals early, offer help just in time, and point people toward real resources. But a large part of social and emotional learning, and of wellbeing, is identifying feelings and expressing them, sometimes sitting in discomfort, and returning to the same patterns long enough to learn to regulate. Regulation techniques tend to require connection with the body and with other people. We learn to settle partly by co-regulating with someone else’s calm, and partly by noticing what the feeling does in the body. An interface that smooths the discomfort away can skip the very part that builds the skill. We also learn through friction, by being challenged and meeting resistance, and a system optimized to feel effortless can quietly remove that too. At least that is what the prior research suggests.

Current research shows AI can reduce the feeling of loneliness, but the mechanisms are underdeveloped and the long-term, longitudinal effects are still barely studied. The open question is whether people feel less lonely in the short term while developing emotional dependence and substituting AI for human relationships they have or could form. There are empirical and anecdotal findings from work on Replika and Character.AI, where people relate to these characters as friends and in some cases significant others, and the public mourning when OpenAI deprecated GPT-4o, including the “funeral” that circulated, showed how strong attachment to a specific model or character can be.

3. Your work suggests that the same technology can affect people very differently. One person might feel supported, while another might feel more anxious or isolated over time. If the effects are so personal, what does that mean for companies designing these systems at scale?

A few things sit underneath this. People’s experiences of wellbeing and illbeing differ significantly, and so do the ways they interact with the technology. Yet most research and most large-scale product work optimizes for population averages, or for personas and segments defined ahead of time. Even the personas are usually group-level pictures fixed in advance. The problem is that the average can be wrong for almost everyone. A population model might find that more time on an app correlates with worse mood. But for one person, heavy use late at night is a signal that they’re in a bad stretch, while for another it’s simply how they wind down, and they’re fine. A metric built on the group average misreads both.

The alternative is bottom-up. Build the picture from how a person actually uses the system over time, and judge change against their own baseline rather than a group’s. Are they doing better or worse than their normal? That’s a within-person question, and it’s the one that matters for wellbeing. The second piece is agency. Show people their own use patterns transparently, and let them choose how and what they interact with. None of this requires waiting. These are things companies can start building now.

4. When we click “I agree” on an app, we usually think about data like our name or location. But today, systems can also learn from how we write, how often we engage, even how we respond emotionally. Do you think people are really aware of what they are agreeing to in this more subtle, behavioral sense?

I don’t think so, and I’m glad you raised it. When we talk about AI literacy, it often gets reduced to understanding how to build a model from scratch and what data that takes. The part that actually matters for everyday use gets very little attention: what these systems infer about you from how you use them. It isn’t only the content of what you type. The way you phrase things, how often you come back, how long your sessions run, what you linger on, how you respond when the system pushes back, all of it becomes signal for a model of who you are. Most people never agreed to that in any meaningful sense, because it was never made legible to them.

There’s a second layer worth naming. All of that behavioral data flows back into improving the models and the services, which gets framed as building better systems. But better by what measure, for whom, and aligned to whose values, is loosely defined. In practice better often collapses into better at holding attention or retention, because those are easy to measure and the business rewards them. Whether the system is better for the person is a different question, and it rarely gets asked in the same breath.

5. A final question. Do you think these technologies are simply responding to who we are or quietly reshaping how we feel, relate to others, and understand ourselves over time?

Technology has always shaped us as we shape it. It shapes language, behavior, lifestyle, values, and eventually culture. Distributed and extended cognition theories describe how we reorganize our own cognitive functions through the tools we use. We already offload memory, planning, and drafting to these systems. What’s new is that the same system we offload our thinking to is also the one we go to for advice on the social challenges we face, so the offloading is cognitive and emotional at the same time, and from a single source.

What’s different now is that technology used to sit mostly on the “tool” side of the line. Constant availability, natural language interaction, and the sheer range of use cases blur the line between functional and relational use, so the cognitive, socio-emotional, and cultural change is likely to be much larger. AI is a quasi-social actor in our networks, collaborator, workmate, friend, confidant, in some cases significant other, and that will shape how we feel, relate to others, and understand ourselves over time. Which means the designers of these systems, and the companies behind them, hold enormous power, from surveillance through to manipulation. That power should be matched by responsibility and liability.

None of this is inevitable, though. The same dynamics that can narrow us can also extend us. Built and used well, with a person’s own goals and growth at the center rather than their attention, these systems could genuinely expand how we think, learn, and relate. That is the version worth building toward. The technology itself will not decide which one we get. The people designing it, and the rest of us using it, will.