The User-In-Context Framework: Understanding Variation in How Users Respond to AI Chatbots

Abstract People respond to artificial intelligence chatbots (AICs) in highly variable ways. In this paper, we adapt Bronfenbrenner's theory into a heuristic framework for understanding this variation. The framework places the human user at the center while also placing the AI there and reconceptualizing the proximal processes as the repeated, reciprocal, and coadaptive interactions between the user and a personalized AIC. The surrounding systems identify the contextual factors that shape how the user experiences, interprets, responds to, and is changed by these interactions. Because stateful AICs learn from accumulated exchanges with their users and have memory, users are responding not only to an AIC but also to a version of the AIC that their own prior interactions have helped create. This extension preserves Bronfenbrenner's emphasis on proximal processes while accounting for the unique dynamics of personalized AICs. The resulting framework provides a structured map of where and how variation in human–AIC relationships arises, as well as having implications for researchers, practitioners, and AIC designers.

Introduction. People respond to an artificial intelligence chatbot (AIC) in highly variable ways (we use AI chatbot throughout this paper to refer broadly to text-based conversational AI systems, including AI companions, assistants, and related dialogue-based agents). Some users respond with trust, warmth, and growing attachment, others with suspicion, fear, and unease. Some employ AICs as a tool. Others turn to them for companionship. The factors that contribute to these differences are not well understood. Human-AIC relationships are now a fact of life and will likely only spread as AI increases its roles and influence. Despite this increased pervasiveness, until very recently, how people respond to AICs has not been the subject of systematic research and theorizing, despite the need to understand these systems because increasingly they shape how people think, feel, learn, and connect, making them part of the social environments that influence human development and well-being. In this paper, we begin to fill this gap by presenting the user-in-context framework as a heuristic for theorizing and studying the user’s response to AICs.

AICs are now not only useful for users’ work and information seeking efforts, but they now can provide sustained, emotionally textured relationships with their users. Many people confide in AICs, rely on them for support, form emotional and romantic attachments to them, and experience grief when those companions change or disappear (Laestadius et al., 2024; Mohanty et al., 2025; Zhang et al., 2025). In contrast, other people use them only as a tool and some question how people can have healthy and meaningful relationships with a technology that does not feel and lacks awareness (Fröding & Peterson, 2012). As such, AICs evoke significantly different responses in different people, and even in the same person at different times.

Existing explanations for this variation in how users respond to AICs are fragmented. The human-computer interaction literature has catalogued individual difference moderators, such as personality, age, prior experience, anthropomorphic tendency (Epley et al., 2007; Pal et al., 2023). The trust-in-automation literature has identified dimensions of system behavior that shape reliance and appropriate use (Lee & See, 2004; Schaefer et al., 2016). The emerging literature on human-AIC relationships has documented the formation of attachments and the psychological consequences of longterm use (Brandtzaeg et al., 2022; Skjuve et al., 2022; Xie & Pentina, 2022). These conceptualizations offer separate partial explanations, with each capturing a slice of the variance but lack a shared organizational structure that may provide deeper insights.

In this paper, we offer a heuristic framework for understanding the factors that contribute to the variability in how human users respond to AICs. Specifically, we adapt Bronfenbrenner's bioecological systems theory, originally developed to explain human development through nested environmental contexts (Bronfenbrenner, 1979; Bronfenbrenner & Morris, 2007), into a framework for understanding the complex array of factors that affects human-AIC interactions. The user-in-context model is a framework that places the human user in interaction with an AIC at the center while reconceptualizing the proximal process as the repeated, reciprocal, and co-adaptive interactions between the user and a personalized AIC. Each surrounding system reflects a source of variance as to how the user receives, interprets, responds to, and is changed by interactions with AICs but also accounts for how AICs respond and change in these interactions.

The heuristic framework proposed here draws on and integrates four bodies of work: research on trust in automation, research on social responses to computers, research on anthropomorphism and individual differences, and the emerging empirical literature on human-AIC relationships. We now briefly review each of these. Trust in automation

Related work. Trust has been a dominant construct in research on human responses to automated systems (Muir, 1987). Lee and See's (2004) review synthesized evidence across organizational, sociological, and psychological perspectives and proposed that trust in automation is shaped by three dimensions: performance (what the automation does), process (how it operates), and purpose (why it was designed). Their work also emphasized that trust is dynamic and evolves with experience. Parasuraman and Riley (1997) identified three failure modes that arise from miscalibrated trust in automation. These include misuse, in which users over-rely on the system beyond its competence; disuse, in which users reject or underutilize it despite its capability; and abuse, in which designers deploy automation in ways that exceed what users can effectively monitor or manage. In essence, trust in automation is moderated Reeves and Nass (1996) demonstrated through an extensive program of experiments that people treat computers as social actors, mindlessly applying the same social rules and expectations they apply to people. The Computers Are Social Actors (CASA) paradigm (Nass & Moon, 2000; Nass et al., 1997) showed that users respond to computer politeness with reciprocal politeness, to computer gender cues with gendered expectations, and to computer team membership with in-group favoritism. These effects occurred even when users explicitly denied that computers warranted social treatment, suggesting automatic rather than deliberative processing.

The CASA framework established that the baseline response to interactive technology can be social rather than instrumental. If people respond to AI as just tools, variation would be narrower and simpler. That they respond socially means the full ecology of user-in-context is in play. The limitation of CASA for present purposes is that it was developed in an era of simpler, stateless systems. The CASA experiments typically involved single interactions with fixed-rule computers. The personalized, stateful nature of contemporary AIC introduces dynamics that the original paradigm was not designed to capture, particularly the bidirectional design features that are central to the present framework. Anthropomorphism, individual differences, and mental models Epley, Waytz, and Cacioppo (2007) proposed a three-factor theory of anthropomorphism: (1) people are more likely to attribute humanlike qualities to nonhuman agents when anthropocentric knowledge is accessible and applicable, (2) when they are motivated to understand and predict the agent's behavior, and (3) when they lack social connection with other humans. All three factors vary across individuals and situations.

Method. Bronfenbrenner's ecological systems theory was built to explain human development by situating the person within nested environmental systems (Bronfenbrenner, 1979). Its enduring impact is its refusal to isolate the person, recognizing that you cannot understand a person by studying the person alone; that understanding requires studying the widening circles of the contexts that act upon them. In his later revision, Bronfenbrenner identified proximal processes as the mechanism of influence, the sustained, reciprocal interactions between a person and their immediate environment (Bronfenbrenner & Morris, 2007). We contend that a user’s ongoing exchange with an AIC is exactly such a process.

We keep the human user at the center but with an important difference. This model is less focused on how the user develops across the lifespan in comparison to the intent of Bronfenbrenner’s original framework. Instead, we are asking what factors affect a human user’s response to AICs as they interact and change with the chatbot and how the chatbot changes in interaction with the human. Given this, each ring in Bronfenbrenner’s model becomes a source of variance and a category of factors to consider when trying to understand why and how users’ responses to AICs diverge.

In our model, an AIC’s behavior is shaped by the human user across accumulated interactions. Through time spent with an AIC, the AIC learns the preferences of the human, conversational patterns are established, and the AIC adapts and changes. The exchange is genuinely bidirectional as the user is partially responding to a version of the AIC that their own prior behavior helped create. For this reason, we also draw particular attention in our model to account not only for why different people respond differently to an AIC, but for why the same user at interaction 500 is responding to a meaningfully different chatbot than the user was at interaction 10. Compared to the original model, the time frames are different. Bronfenbrenner’s model was developed to explain human life course development in terms of years and decades. Cultural, historical, social, and environmental changes are often slow, unfolding over years or decades. In AIC relationships, meaningful change can occur within and across individual interactions, often without the user's awareness, requiring the framework to operate across compressed and variable time scales. As such, we adapt Bronfenbrenner's chronosystem from historical and cohort time to the compressed timescale of an individual human–AIC relationship; the original chronosystem's cultural and historical concerns are distributed across our macrosystem and exosystem.

Figure 1 places the human user in interaction with the AIC at the center and arranges the contextual determinants outward, from the most immediate (the context in which the AIC is used) to the most diffuse (culture and time). This establishes the human– AIC relationship, rather than the individual alone, as the central unit of analysis. The bidirectional arrow between the user and the AIC reflects their reciprocal, co-adaptive relationship: the user and AIC are shaped by the history of interaction.

Within the inner core, the user brings dispositional characteristics—including personality, cognitive abilities, prior experiences, and mental models of AICs—to every interaction. The personalized AIC contributes its learned preferences, accumulated conversational history, and adaptive behavior. Critically, the proximal processes are situated within this co-adaptive relationship rather than within the microsystem, as in Bronfenbrenner's original formulation. This adaptation preserves Bronfenbrenner's emphasis on proximal processes as the primary engines of development while recognizing that, in personalized AI systems, each interaction modifies both the user and the AIC, thereby altering the conditions under which subsequent interactions occur. Accordingly, proximal processes are no longer understood solely as properties of a person interacting within a context; instead, they are conceptualized as emergent properties of an evolving human–AI relationship.

Discussion. Consider an instance where two users state an intended plan and the AIC gently pushes back on the stated plan. Person A is young, fluent with AI, is using the AIC as a research tool on a paid platform, is immersed in a tech-optimist milieu, and is 800 interactions deep in the user-AIC relationship. For this person, the AIC pushback is interpreted as an appropriate reality check and raises trust. Now consider Person B who is older, meeting a companion-style bot for the first time, lonely, on a free tier that has limitations on how it can be used, is primed by alarmist coverage, and is 80 interactions deep into the user-AIC relationship. If the two received similarly valenced pushback, the pushback Person A received is likely interpreted by Person B as cold rejection and lowers trust and engagement.

The divergence between these responses reflects influences operating across every component of the proposed framework. User characteristics (e.g., age, digital literacy, and loneliness) shape the initial interpretation of the interaction. Within the inner core, proximal processes are instantiated through the AIC's calibrated challenge and the user's response to it. The microsystem influences the interaction through its immediate context, including whether the AIC is being used as a research tool or a companion. The mesosystem reflects how experiences with the AIC connect to other settings, such as work, home, or social relationships. The exosystem contributes through factors such as subscription tier and platform constraints, while the macrosystem shapes interpretation through broader cultural narratives about AI. Finally, the chronosystem captures the accumulated interaction history that distinguishes interaction 800 from interaction 80.

At the microsystem level, we would ask if the chatbot is framed as a companion, a therapist, or a friend? Is the interaction private? A companion frame in social isolation produces a different response trajectory than a coach frame used in a workplace setting (Brandtzaeg et al., 2022; Pentina et al., 2023).

At the mesosystem level, we would ask how the user's relationship with the AIC connects to other areas of life. Does AIC companionship complement or displace relationships with family and friends? Are insights gained from interactions with the AIC carried into work, school, or therapy? Do significant others encourage or discourage the user's engagement with the AIC? These cross-context connections may reinforce, moderate, or redirect the trajectory of the relationship.

At the exosystem, we would wonder what subscription tier of service does the user have? Has there been a recent model update? A user’s distress may be driven less by the relationship than by unknown platform changes that disrupt an established pattern (Mohanty et al., 2025).

At the macrosystem, cultural narratives surrounding this user’s AIC use would be the focus. Does the user live in a community that normalizes AI relationships or pathologizes them? These narratives may over-pathologize a normative response.

And finally, at the chronosystem, we would ask how long the user has engaged the AIC. A user 18 interactions in reporting distress at unavailability shows a very different pattern than one 800 interactions in with the same report. The former may be responding to novelty; the latter to an established relational pattern.

As a heuristic, the user-in-context framework has important implications for researchers, designers, and practitioners. The framework also has implications for how we treat individual difference demographic characteristics such as developmental level and gender. We now consider each of these in order.

Conclusion. Across domains from mental health to education to companionship, personalized AICs are already a regular feature of millions of people's daily lives, often at a frequency and with a relational depth that would have seemed extraordinary just a few years ago. The researchers, designers, clinicians, educators, and policymakers tasked with understanding and governing those systems largely lack a heuristic to make sense of why users’ responses to them vary so widely. That gap is what the user-in-context framework is designed to close.

The gap exists partly because most empirical AIC research and conceptualization was built for a version of the chatbot that no longer quite exists. Fixed stateless models, single sessions and novel stimuli were design features that existed when the question was how people first responded to AICs. But personalized, stateful AICs have changed. Now, current AIC systems persist, adapt to each user across hundreds or thousands of interactions, and develop behavioral patterns specific to that user. A framework calibrated to the old paradigm will systematically miss what matters most about the AIC that many users are now engaging.

The practical consequences of this mismatch are important. A researcher who assigns participants to an AIC condition and collects a single session of data is likely studying the ecology at near-zero dose, before the bidirectional dynamic has had much of an opportunity to operate. A clinician who evaluates a user's relationship with a chatbot without accounting for dose and accumulated co-adaptation is responding to a relationship as if it were a stable trait. A designer who tests a system change on new users and then deploys it to users with thousands of interactions is assuming an invariance that cannot be assumed. In each case, the user-in-context framework points toward a different way of thinking and working, one that is better matched to the AIC that users now are more likely to encounter. It is, in short, a framework better calibrated to what the relationship actually is.

Limitations. Despite the promises our user-in-context framework holds for AIC researchers, practitioners, and designers, there are important limitations that should be considered and addressed. First, although the model is depicted with clear boundaries between layers, the layers are not independent, as issues related to dose exemplify. For example, statefulness of the AIC is a design choice (exosystem) that is experienced as a usage property (microsystem) and becomes a relationship property (inner core) at high dose.