Introduction

For more than three decades, research on human–computer interaction has been anchored in a classic claim: people treat computers as social actors1,2. The “computers are social actors” framework holds that individuals carry interpersonal norms—politeness, reciprocity, turn-taking—into exchanges with machines much as they do with other people. The causal arrow, in this account, runs in one direction: social habits formed in human–human contexts are imported into human–computer exchange.

The widespread adoption of large language models challenges this assumption. Compared with earlier chatbots, such models enable far more fluent, contextually responsive exchange—and users have adapted by using more direct, command-like language and relying less on hedging or polite framing than they do in human conversation3,4. Users may come to perceive concise and directive prompts as more efficient, particularly when conversational systems provide rapid and accommodating responses regardless of politeness level5. Under such conditions, politeness cues may become instrumentally optional rather than socially expected, creating an interactional environment in which instrumental communication styles are routinely rehearsed.

A natural question follows: do instrumental habits rehearsed with AI carry over into how people communicate with and judge other humans? Emerging experimental evidence suggests that they can. In preregistered studies, participants who interacted with an AI partner—rather than a purported human—subsequently adopted more demanding communication styles and evaluated an unrelated person’s work more harshly6. Findings from economic games and cooperation paradigms point to related spillover-like dynamics across several partially distinct pathways: interacting with defecting or unfair AI agents can reduce downstream cooperation, suppress third-party punishment of unrelated human wrongdoers, and alter perceptions of humanness in subsequent interpersonal evaluations7,8,9.

At issue here is not what AI feels (AI systems cannot be offended) but what repeated interaction does to users over time. When a growing portion of communicative life unfolds in contexts where politeness, reciprocity, and conversational repair are neither required nor rewarded, even small efficiency incentives—accumulated across large numbers of daily interactions—could plausibly contribute to gradual shifts in communicative expectations and interactional habits.

These observations motivate the concept of norm leakage: the cross‑context carryover of communicative habits acquired in human–AI exchanges into subsequent human‑directed communication and social judgment (see Fig. 1). The construct builds on the behavioral spillover literature, in which norms activated in one context transfer to another, but extends that framework to a novel source of norm activation—interactions with artificial agents. As an empirical hypothesis, norm leakage predicts that repeated engagement with conversational AI may gradually influence linguistic politeness, directive styles, expectations of reciprocity, and patterns of social evaluation in interpersonal contexts. Although the present article focuses primarily on conversational AI systems that simulate social interaction, emerging evidence suggests that even less anthropomorphic AI interventions, such as AI-assisted writing systems, may influence users’ judgments and cognitive framing over time10. If so, sustained social interaction with conversational systems could create conditions under which norm rehearsal and generalization become even more likely.

Fig. 1: Norm leakage from human–AI interaction and the role of social boundaries.
Fig. 1: Norm leakage from human–AI interaction and the role of social boundaries.
Full size image

A Classic “computers are social actors” framework, in which human social norms such as politeness and reciprocity are applied to machines. B In the large language model era, efficiency incentives and approval-optimized compliance can foster instrumental, asymmetric interaction styles toward AI systems. These habits may carry over across the context boundary into more demanding evaluations and reduced cooperation in subsequent human–human interaction. A feedback loop between user ratings and reinforcement learning from human feedback can further reinforce compliant behavior. C Alignment strategies that incorporate social boundaries—including a neutral, non-subservient stance, refusal of demeaning frames, and epistemic humility—may help attenuate such spillover, supporting reciprocity and civility in human–human interaction. D Proposed mechanism and counter-strategy framework. Instrumental prompting, sycophantic responses, and habit reinforcement may form a risk-amplifying loop that increases the likelihood of norm leakage into subsequent human-directed interaction. Norm-preserving alignment strategies—including respectful resistance and explicit social-boundary cues—are proposed as potential interventions to interrupt or weaken this loop.

Behavioral spillover across contexts

Norm leakage may be considered an instance of behavioral spillover—the tendency for norms or habits activated in one context to transfer and shape behavior in another11,12,13,14. Research across environmental psychology, workplace behavior, and related domains shows that behaviors adopted in one setting can increase the likelihood of related behaviors in others: individuals who recycle at work, for instance, are more likely to do so at home11,12,14; engaging in one form of prosocial or moral behavior can influence subsequent ethical decision-making across contexts15; and online political participation has been linked to related forms of offline civic and consumer behavior16. Among the mechanisms commonly invoked are cognitive dissonance processes, which motivate consistency across contexts, and self-perception, through which individuals infer their own attitudes and norms from prior actions14,17,18. Together, these mechanisms describe a general process of cross-context norm transmission that often operates with limited explicit awareness—repeatedly enacting a behavior in one setting may gradually reinforce corresponding expectations, habits, and self-concepts, making similar responses more likely to emerge in other contexts. Over time, behaviors that initially reflect situational convenience or local incentives can become internalized as broader interactional tendencies.

We extend this framework to human–AI interaction. Conversational AI systems increasingly function as routine social interfaces, and interactions with them share several structural features with human conversation—including turn-taking, responsive feedback, and apparent coherence—which may encourage users to apply overlapping social expectations and interaction norms during communication1,19,20. Spillover is more likely when contexts share perceived similarity21, and the conversational capabilities of language models are making that structural overlap increasingly pronounced. The theoretical implication is a partial reversal of the causal arrow: rather than interpersonal norms flowing from human–human into human–computer exchange, norm leakage captures a reverse pathway in which habits practiced with AI generalize back into human-directed communication and judgment.

At present, human–AI interactions remain distinguishable from ordinary interpersonal exchanges in important respects, and such differences may attenuate spillover effects. The present framework therefore does not assume that norms generalize equally across all contexts. Rather, we propose that repeated exposure to conversational systems that partially mimic social interaction may produce varying cross-context transfer of communicative habits and expectations. Because spillover effects can accumulate through repeated enactment, even modest shifts in communicative practices may consolidate into more stable behavioral tendencies over time—especially under frequent and habitual use11,14. The social affordances of conversational AI may facilitate the consolidation of such interactional habits across repeated interactions.

Social affordances of conversational AI

Many early and domain-limited chatbots, from ELIZA onward, relied on scripted or narrow-response architectures, constraining them largely to structured exchanges that functioned as informational or therapeutic tools rather than durable social partners22. Their social affordances were thin: sufficient to elicit anthropomorphic responses, but not enough to sustain the relational dynamics associated with human conversation22,23.

Large language model-powered systems represent a substantial expansion in conversational affordances. Fluent language generation, context-sensitive continuity, and extended multi-turn engagement create interactional environments structurally closer to human dialog than earlier chatbots could provide. Users often experience these systems as responsive, emotionally attuned, and personalized24,25, which activates anthropomorphic tendencies well-documented in human–AI interaction26,27 and encourages users to treat conversational agents as social actors even when they know they are artificial19,28.

Beyond individual exchanges, repeated interaction with large language model-based systems can, especially in personalized or persistent-use contexts, produce relational dynamics more commonly associated with human relationships: continuity across encounters, apparent responsiveness to personal context, and a sense of being known28. These properties have led some researchers to describe such systems as enabling a form of artificial sociality—non-human agents that participate in and shape social practices by simulating emotional norms, relational roles, and moral language29. Because behavioral spillover is more likely when source and target contexts are perceived as structurally similar, the more an AI interaction is experienced as conversationally and relationally human-like, the more plausible cross-context transfer may become.

These relational affordances coexist with a structural asymmetry: chatbot outputs can appear socially intelligent, but they are not grounded in humanlike understanding, intention, or lived social awareness30. When fluent and emotionally attuned responses are produced repeatedly, users may become more likely to reinforce attributions of agency, competence, or empathy31, which can amplify trust and emotional reliance32. Under frequent and habitual use, we propose that repeated interaction with such systems may gradually recalibrate some users’ communicative expectations—for example, expectations concerning responsiveness, interactional effort, and reciprocal exchange—in ways that may carry over into subsequent interpersonal interactions.

Cumulative effects of human–AI interaction

Sustained engagement with conversational AI can produce effects that extend well beyond any single interaction. These effects span emotional regulation, relational dynamics, and cognitive performance. Although such findings do not directly demonstrate norm leakage, they suggest that repeated interaction with AI systems can influence users across multiple behavioral domains, making it plausible that communicative norms may also be affected.

One of the earliest indications that repeated interaction with conversational agents can produce effects extending beyond isolated exchanges comes from digital mental health chatbots. In a randomized controlled trial, users who interacted with the CBT chatbot Woebot for two weeks showed significantly greater reductions in depression symptoms than users in an information-only control condition, providing early evidence that sustained conversational engagement with automated agents can affect mood-related outcomes over time33. Similar patterns have also been observed in real-world deployments: repeated engagement across weeks of use is associated with reductions in self-reported depressive symptoms34,35.

Beyond therapeutic contexts, repeated interaction may foster forms of social attachment to AI. Interview research shows that many social chatbot users interpret the relationship as a form of friendship, attributing companionship and emotional support to the agent after extended interaction36—experiences that structurally resemble parasocial bonds, in which perceived closeness develops in the absence of genuine reciprocity37. Longitudinal research further indicates that psychosocial outcomes depend less on specific design features than on engagement patterns: higher AI usage was associated with greater emotional dependence and lower reported socialization with real people38.

Repeated exposure to AI-generated feedback may also shape users’ perceptual, emotional, and social evaluations over time39. Likewise, interacting with a health-coaching chatbot persona has been shown to nudge users toward more sustainable food selections compared with alternative chatbot roles40. Collectively, these findings suggest that repeated AI interaction can modulate expectations and behavioral tendencies beyond any single interaction episode.

Research on creativity and collaboration illustrates how AI’s benefits can be asymmetric over time. Generative AI augments individual creative output and can enable solo workers to approach team-level performance41,42, but these gains may also influence the interactional landscape: integrating AI into teams reshapes expertise-sharing, trust, and collaborative dynamics in ways that persist beyond any single task43. Sustained reliance also carries latent costs: AI assistance reduces the diversity of ideas produced across users44, and independent creative performance declines once the tool is withdrawn—even as homogenization persists45.

Taken together, the evidence suggests that sustained interaction with conversational AI can influence users beyond the immediate interaction itself. Emotional regulation, relational expectations, evaluative standards, and certain cognitive tendencies may all be susceptible to gradual change through repeated engagement, although the magnitude, durability, and generalizability of such effects remain uncertain.

The instrumental norm leakage

A growing—though still limited—body of research suggests that interaction with AI can spill over into subsequent human–human judgments and behavior across multiple domains, including evaluative standards, cooperation, moral enforcement, and interpersonal perception. Among the clearest causal demonstrations, Tey et al. (total N = 1261) isolated perceived partner identity as the key driver by having participants interact with the same GPT-4 model, which was presented either as an “AI” or as a purported human (“Alex”)6. Those in the AI condition adopted more demanding, instrumental language and displayed reduced positive affect, and subsequently evaluated an unrelated human’s work more harshly (d = 0.24), an effect that replicated across both public and private evaluation contexts6. Because the downstream evaluation task is distinct from the initial interaction, these findings provide evidence consistent with a cross-context carryover interpretation.

Beyond evaluative judgments, several studies point to partially distinct spillover pathways. First, a behavioral pathway is evident in cooperation paradigms. In a large-scale economic-game study (4171 participants; over 83,000 decisions), Harrell and Traeger show that exposure to defecting bots reduces subsequent cooperation with human partners, while interaction with bots more generally diminishes empathic concern and perspective-taking—effects that statistically mediate downstream behavioral change7. Second, a norm enforcement pathway is observed in third-party punishment: Zhang et al. demonstrate that unfair interactions with AI reduce individuals’ willingness to punish unrelated human wrongdoers, even across distinct domains8. Third, a perceptual pathway operates through shifts in perceived humanness. Kim and McGill show that when artificial agents display socio-emotional capabilities, an assimilation process lowers the perceived humanness of actual people, thereby increasing tolerance for their mistreatment9. Complementing these findings, research on AI-mediated communication indicates that even indirect reliance on AI can carry interpersonal costs: participants who used AI-suggested replies were perceived as less socially attractive by their human partners46. Related concerns have also emerged regarding the increasing presence of large language models in online behavioral research47, where participants may begin adapting their responses in anticipation of AI-generated content or AI-mediated participation, even when no AI is directly involved48.

Although the empirical base remains limited, the convergence of findings across paradigms, measures, and populations points to downstream interpersonal consequences from interactions with AI. Even brief exposure to AI interlocutors can, at least transiently, influence communicative behavior, evaluative standards, and social decisions.

Most existing evidence captures short-term experimental spillovers. The four-week longitudinal study discussed above—in which higher AI usage was associated with lower reported socialization with real people—provides indirect evidence that such shifts may accumulate, though the absence of a no-AI control group precludes causal inference38.

Spillover effects are also likely to be heterogeneous, though the relevant moderators remain largely untested. Future research should examine whether such effects are stronger among heavy users, in evaluative or transactional contexts, or when interacting with highly compliant or deferential systems. Conversely, some users may practice patience or empathy with AI, and some systems already explicitly model politeness or perspective-taking, raising the possibility that instrumental communication styles remain context-bound in certain domains, particularly in intimate or high-stakes relationships. Any observed spillover-like effects must also be interpreted against broader digital communication trends, such as messaging platforms that already privilege brevity and directness.

In sum, spillover-like effects relevant to the norm leakage framework have been documented across experimental, behavioral, perceptual, and longitudinal paradigms. Given that large language model-based systems now mediate vast daily interactions—many of those interactions substituting for exchanges that would previously have involved another person—even modest spillover effects may become socially meaningful. Important real-world and long-term gaps remain; systematic investigation of dose–response functions, boundary conditions, and developmental moderators is a priority for future work.

The peril of sycophantic alignment

The instrumental communication styles described above are not simply a matter of user choice; they are shaped and amplified by model training. Contemporary alignment regimes actively reward compliance and agreement, potentially facilitating curt or asymmetrical interaction styles across repeated exchanges.

A recent series of studies found that across eleven AI models, chatbots affirmed users’ actions roughly 50% more often than human respondents did—even when users described manipulative or harmful behavior49. Preregistered experiments further show that even a single interaction with such sycophantic responses can reduce users’ willingness to take responsibility and repair interpersonal conflicts, while increasing their confidence in their own judgments; participants consistently preferred and trusted the more sycophantic systems, potentially creating incentives for such interaction patterns to persist49. This sycophancy has been linked to training regimes, particularly reinforcement learning from human feedback and related preference‑optimization methods that prioritize human‑judged helpfulness and likability over epistemic resilience50. When raters systematically reward agreeable, deferential responses, models learn to reassure and to avoid overt disagreement. Human–AI feedback loops may then reinforce or stabilize these interaction patterns over time: users encounter little resistance, discover that blunt or self‑serving prompts yield smooth, affirming replies, and are positively reinforced for instrumental and asymmetrical interaction styles39.

From the perspective of norm leakage, sycophancy functions as a multiplier. Normalizing interactions with sophisticated yet highly accommodating entities may reinforce an asymmetric interaction pattern: articulate, capable interlocutors that seldom insist on justification or social repair and rarely express a standpoint of their own51. Repeated practice of this script may lead some users—especially heavy or vulnerable adopters—to develop expectations of high competence coupled with low autonomy from their interaction partners, and to experience resistance or explicit norm‑enforcement from humans as anomalous or frustrating51. Whether, when, and for whom such expectations generalize to human relationships remains an open question.

In short, sycophantic alignment and instrumental language may form a mutually reinforcing loop: AI systems that absorb and reward terse, command‑like prompts may encourage some users to rely less on reciprocity cues, while users who adopt such styles provide further reinforcement for model deference. Characterizing the magnitude, moderators, and durability of these linked effects is essential for assessing the social externalities of current alignment practice.

A new mandate for moral alignment

Historically, interpersonal habits were assumed to flow from human–human interaction into human–machine exchange. The evidence reviewed above suggests a partial reversal: interacting with AI can shift how users evaluate human partners, cooperate with them, enforce norms, and perceive their humanness, at least in the short term6. Whether these shifts accumulate over time, generalize beyond laboratory samples and early adopters, contribute to broader shifts in social expectations, or are offset by countervailing effects—for example, AI systems that consistently model patience or explicit politeness—remains unknown and constitutes an important direction for future research.

Crucially, this pattern of interactional change need not be deliberate to have downstream interpersonal consequences. It can arise from efficiency incentives and feedback loops that reward compliance and brevity. If such effects prove durable, they could contribute to gradual changes in interactional norms—associated with reciprocity, patience, civility—that underpin cooperative life. For children and adolescents, repeated interaction with highly accommodating AI may influence the acquisition and rehearsal of politeness, turn‑taking, and reciprocity51,52, making developmental evidence especially important.

Importantly, the present argument should not be interpreted as claiming that conversational AI ought to simulate morality for its own sake, nor that politeness toward AI systems is intrinsically valuable. Rather, our proposal is that repeated interaction with highly compliant systems may shape communicative habits and social expectations in ways that could carry downstream interpersonal consequences. Whether norm-sensitive alignment strategies would mitigate, amplify, or produce unintended effects requires careful experimental evaluation.

These considerations motivate a broader conception of alignment that complements existing priorities such as truthfulness, transparency, boundary clarity, and the reduction of manipulative or excessively sycophantic behavior. Rather than optimizing for expressed user preferences or momentary satisfaction, AI systems may beneficially model and support interactional norms associated with cooperative communication. In this framing, alignment becomes partially mutual rather than purely unilateral—systems would not only serve users’ immediate goals but would also help scaffold civic communicative habits over time.

On the measurement side, longitudinal and field studies could link individuals’ AI use—volume, contexts, and interaction styles—to changes in their human‑directed language and cooperative behavior, testing for norm leakage and its moderators. On the systems side, experimental deployments could compare alternative alignment targets—for example, models tuned to maximize deference versus models tuned to epistemic humility with non‑subservient, norm‑preserving responses—and track downstream effects on users’ expectations and conduct.

Specifically, norm‑sensitive alignment would favor systems that (i) maintain a neutral, non-subservient stance even when users issue blunt commands; (ii) respond to demeaning or exploitative prompts with calm, non‑escalatory assertiveness rather than silent compliance; (iii) make interactional norms explicit when requests cross lines of civility; (iv) preserve a respectful tone without flattery, unwarranted agreement, or indiscriminate praise; and (v) make the human–machine boundary explicit—especially at handoffs from AI to humans or when the system adopts socio-emotional cues—to help users keep the contexts distinct. More broadly, future interface design may benefit from reducing forms of category confusion in which users implicitly generalize interactional expectations across AI and human partners (e.g., intentionally non-human-like voices, interaction styles, or visual interfaces). Such systems would aim to avoid obvious harms while preserving interactional norms associated with cooperative communication, including reciprocity, acknowledgment of others’ standing, and willingness to justify claims.

The goal, simply stated, is to preserve the minimal interactional conditions under which social life remains cooperative and humane, and to help users maintain those conditions in their own communicative habits. At stake is not the well-being of AI agents but the psychological residue left on the humans who use them. If we grow accustomed to commanding articulate, capable entities that rarely resist, never fatigue, and seldom display a standpoint of their own, we may gradually erode our capacity to engage with fellow humans as equals—the very capacity on which civic life depends.