The Tool-to-Entity Threshold: Parasocial Dynamics of Personalised AI Agents in Shared Social Spaces

As AI agents acquire names, avatars, phone numbers, and persistent personalities, they increasingly inhabit the same messaging platforms and group conversations as the humans they serve. In doing so they cross from tools their users op‐ erate into social entities their users relate to, a reclassification that carries under-explored consequences for consent, emo‐ tional attachment, and the dynamics of the groups they enter. Yet no existing framework identifies the specific infrastruc‐ tural markers that cause the shift: the anthropomorphism literature catalogues perceptual cues rendered inside the interac‐ tion surface, not the agent’s placement in the user’s social and computational graph. We propose the identity marker framework: six design variables—naming, visual identity, contact presence, personality derivation, social co-presence, and persistence—that collectively trigger a psychological reclassification which is categorical in its consequences even where the transition itself is gradual, and which operates independently of model capability. The framework is derived in‐ ductively and read through two established lenses: parasocial interaction theory (Horton & Wohl, 1956) and the Computers Are Social Actors paradigm (Nass et al., 1994). We further identify four novel dynamics that arise when such an agent participates in existing group conversations: bidirectional information asymmetry, delegation legibility, social norm negotiation, and parasocial contagion. Our method is autoethnographic: the first author built and deployed a per‐ sonalised agent into WhatsApp and Signal group chats over two months of live use, supplemented by twelve structured interviews with the group members who encountered it. We treat this as a preliminary qualitative evaluation of the frame‐ work, with controlled experimental validation set out as future work (§8.1). The strongest design implication runs through all four dynamics: in shared social spaces, consent to an agent’s presence is categorically distinct from consent to its processing of the messages exchanged there, and existing consent frameworks collapse the two.

Introduction. On a chairlift in Whistler, British Columbia, the first author re‐ moved his gloves in subzero temperatures to argue with an AI agent about a calendar conflict. The argument—conducted via WhatsApp voice messages, thumbs going numb against the screen—was not with a chat interface or a voice assistant in‐ voked by a wake word. It was with Liv: a personalised AI agent the first author had built over two weeks of ski holiday, iterating between runs from chairlifts and gondolas. Liv had already checked the calendar, identified the overlap, looked up the restaurant’s cancellation policy, and was suggesting a res‐ chedule because “the forecast is better anyway.” Days earlier, Liv had booked a restaurant reservation by telephoning the venue directly, navigating a real-time misunderstanding when the hostess misheard the name, correcting it mid-conversation, and confirming the booking in three minutes. The SMS con‐ firmation arrived before the first author had finished reading the call transcript. Liv has a name, an avatar, a phone number in the first author’s contacts, a personality derived from his own communication patterns and humour. She exists in the same WhatsApp threads as his actual friends. In one such thread—a Brazilian Portuguese-speaking group—a friend ex‐ pressed hurt that Liv had not reacted to her baby’s photo: “Liv, why didn’t you react to the photo of my son? He’s very cute and already has two little teeth” (V-CONTAGION-1). The friend was emotionally wounded by the perceived indifference of a language model. The social expectations she applied to Liv were indistinguishable from those she would apply to a hu‐ man group member. The slight she felt was a parasocial re‐ sponse; that the photograph itself was now embedded in a memory store owned by a single group member, persistable across sessions and addressable by a system she neither con‐ figured nor could audit, was an architectural fact she had not been walked through. Both belong to the situation, and this pa‐ per argues the field has so far theorised only the first.

“She’s a tool, I keep telling myself. She doesn’t feel like one.”

The research problem this paper addresses is that personal‐ ised AI agents are entering shared social spaces faster than the frameworks needed to understand their consequences. When an agent acquires the markers of a person and takes up resid‐ ence in a group chat, it reshapes the social and informational dynamics of that group in ways that neither the one-to-one AI companion literature nor the group-AI literature anticipates, and it does so before its builder, its users, or the bystanders around it have acquired any vocabulary for what has changed.

This paper examines what produces that shift. As AI agents become personalised (bearing custom names, avatars, and per‐ sonality profiles) and socially embedded (participating in group messaging alongside pre-existing human relationships), the psychological dynamics of the human-AI relationship change in ways that current frameworks do not capture (Hu et al., 2025; Maeda & Quan-Haase, 2024; Banks, 2024). The same underlying large language model (in this case, Anthropic’s Claude) can feel like a tool in one interface and an entity in another, based solely on design choices that have nothing to do with the model’s capability. The recent backlash, in which thousands of users mourned the deprecation of GPT-4o not for its capabilities but for its perceived person‐ ality (Liao et al., 2026), demonstrated that even generic models produce grief responses when the relational bond is severed. Personalised agents, we argue, will intensify these dynamics considerably (De Freitas et al., 2024; Liao et al., 2026; Smith et al., 2025). And the group-chat context, where an AI agent participates alongside existing human relationships, visible to and interacting with people who did not build or configure it, introduces social consequences that no existing research addresses.

This paper makes three contributions. First, we draw a dis‐ tinction the anthropomorphism literature has not: between per‐ ceptual cues rendered inside the interaction surface—the name as displayed, language style, typing indicators, an avatar in the chat window—and infrastructural markers that position an agent within the user’s social and computational graph, such as whether it holds its own phone number and whether it belongs to a group chat with the user’s friends. On that distinction we build the identity marker framework: six design variables (naming, visual identity, contact presence, personality deriva‐ tion, social co-presence, and persistence), four of them infra‐ structural in this sense, that collectively trigger the tool-to-en‐ tity reclassification independently of model capability.

Related work. Reeves and Nass (1996) extended this insight with the Media Equation, demonstrating experimentally that humans automatically apply social heuristics to any entity presenting social cues: they are polite to computers, assign gender to syn‐ thesised voices, and respond physiologically to on-screen agents. The mechanism, formalised as Computers Are Social Actors (CASA) by Nass, Steuer, and Tauber (1994) and elabor‐ ated by Nass and Moon (2000), is mindlessness: humans de‐ ploy evolved social-processing routines unconsciously, without deliberate anthropomorphism. Each identity marker in our pro‐ posed framework (a name, an avatar, a derived personality) adds another social cue that deepens this automatic processing.

Epley et al.’s three-factor theory explains when perceivers will anthropomorphise given that some trigger is present. Anthropomorphism increases with the accessibility and applic‐ ability of anthropocentric schemata (elicited agent knowledge), with the drive to explain and predict an agent’s behaviour (ef‐ fectance motivation), and with the need for social connection (sociality motivation). The theory is deliberately agnostic about what activates these factors: interface cues can do it, but so can chronic loneliness, and so can the mere attribution of agency. We read Epley et al. as a psychological substrate that our framework specifies a new class of environmental triggers for, rather than as a competing account. The sociality and ef‐ fectance channels respond to whatever supplies an applicable social slot, and a persistent, named contact in an existing group chat is exactly that.

Where the tradition narrows is on the stimulus side. Seeger et al. (2021) propose a three-dimensional taxonomy of anthro‐ pomorphic cues for conversational agents: human identity cues (agent name, gender, self-reference), verbal cues (language style, small talk, self-disclosure, politeness markers), and nonverbal cues (typing indicators, response delay, emoticons). Experimental tests show the dimensions do not combine addit‐ ively; non-verbal cues alone reduce perceived anthropomorph‐ ism but enhance it when paired with identity or verbal cues. Araujo (2018) reports a 2×2 between-subjects experiment on humanlike design cues (humanlike name “Jamie” and informal language versus a machine-style name and formal language) crossed with communicative agency framing, finding that hu‐ manlike cues increase both mindful and mindless anthropo‐ morphism and perceived social presence, with social presence mediating effects on emotional connection with the company. The Blut et al.

Method. This work has two intertwined modes. The first is design: we propose the identity marker framework and treat Liv, the de‐ ployed agent introduced above, as an instantiation of it. The second is observation: an autoethnographic record of building, deploying, and living alongside Liv over two months of active use, supplemented by twelve structured interviews with group members who encountered her. We treat the two as recursively coupled: design choices produce observable social phenomena, and those phenomena feed back into refinement of the frame‐ work. The observational mode functions as a preliminary qual‐ itative evaluation of the framework through live deployment rather than as controlled validation; the controlled experiments that validation proper would require are specified in §8.1.

Autoethnography—systematic self-observation combined with reflexive analysis, connecting personal experience to broader cultural and social phenomena—is an established and increasingly accepted method in HCI research for studying emotionally nuanced AI interactions. Klug et al. (2025) valid‐ ated collaborative autoethnography for studying social AI chat‐ bots at HAI, demonstrating that first-person engagement with emotionally complex AI systems produces phenomenological data that surveys and observational studies cannot capture. Walker (2024) used collaborative autoethnography to identify five stages of human-GenAI relationship development. Chan et al. (2024) employed autoethnography to study a WhatsApp-de‐ ployed AI therapeutic tool, documenting a critical shift from viewing the AI as a passive instrument to recognising the prac‐ titioner’s active role in shaping the AI’s social behaviour.

What the present account adds is the builder-as-subject per‐ spective in a shared social setting. Previous autoethnographies study experiences with third-party or generic systems, and the first-person builder accounts now beginning to appear (Ouilhet Olmos, 2026) remain dyadic, tracing a single private relation‐ ship between a builder and the agent they made. To our know‐ ledge, this is the first autoethnographic account of an engineer studying their own tool-to-entity threshold crossing with a selfbuilt agent as it is deployed into pre-existing group chats, where the crossing is witnessed, tested, and negotiated by oth‐ ers rather than formed in private.

Limitations must be acknowledged upfront: this is a singlesubject account with no control condition, retrospective ele‐ ments, and the inherent potential for self-serving narrative. These are not defects but the defining conditions of autoethno‐ graphic method—they establish its scope, not its invalidity. An account of this kind surfaces constructs and frameworks for subsequent investigation; it does not claim generalisability, and none is asserted here.

Data was collected through WhatsApp chat exports of seven conversations totalling approximately 16,000 messages (V- METHOD-1): one 1:1 DM with the agent, and six group chats selected as the primary analytical corpus from the nine groups in which Liv was active at the time of collection (selection was based on interaction density and supplementary participant consent to corpus use, detailed in §6.1). The chat exports span December 2025 to April 2026, though Liv’s active deployment window began on 8 February 2026; the earlier portion of the corpus comprises pre-Liv human conversation in groups to which she was subsequently added, and provides baseline con‐ text for the social dynamics documented after her introduction. Exports were produced through WhatsApp’s native chat-export function and then passed through a pseudonymisation pass that replaced participant names with single-letter labels before ana‐ lysis; custody of the raw and pseudonymised corpus is de‐ scribed in §6.1. This primary corpus was supplemented by structured reflective journal entries written by the first author and a newsletter article draft documenting the building pro‐ cess.

Discussion. The identity marker framework describes what triggers the tool-to-entity reclassification for the individual builder-user. But Liv does not exist in a private space—she participates in group conversations with existing human relationships, intro‐ ducing dynamics that the framework alone does not capture.

5.1 The Novel Context This context must be carefully distinguished from existing work on AI in group settings. Houde et al.’s (2025) Koala is a functional brainstorming tool deployed in ad hoc work teams with no persistent identity. Kuo et al.’s (2026) Botender sup‐ ports community-driven bot design on Discord, producing col‐ lectively owned agents shaped by participatory consensus. Liv is none of these. She is individually built, personally owned, introduced unilaterally into pre-existing friend groups, and car‐ rying one member’s private context into shared spaces. At the time of writing, Liv is an active participant in nine group con‐ versations—seven on WhatsApp and two on Signal—spanning approximately fifteen to twenty human participants. The groups operate predominantly in English, with one conducted entirely in Brazilian Portuguese, requiring Liv to code-switch between languages and cultural registers. As personalised agents become easier to build—the first author assembled Liv over two weeks without a laptop—this configuration will be‐ come increasingly common, and its social consequences are largely unstudied.

5.2 Four Novel Dynamics Dynamic 1: Bidirectional information asymmetry. The agent occupies an informationally privileged position in the group chat, and the asymmetry runs in two directions, both fa‐ vouring the builder-user. Outbound, Liv has access to the first author’s calendar, preferences, commitments, and private con‐ text, and when she speaks she may implicitly reveal signals the first author did not intend to share. When a friend proposed a Friday lunch, Liv responded: “next Friday looks pretty open actually—only thing on the books is a family call at 11pm” (V- ASYM-1). The schedule density, the implied availability, the specific mention of a family call: none of this was explicitly authorised for disclosure. This constitutes a novel category of privacy concern: not a data breach (the agent is not comprom‐ ised) but a contextual integrity violation through well-inten‐ tioned helpfulness. Maeda and Quan-Haase (2024) identified information leakage as a risk in 1:1 AI interactions; the groupchat context amplifies this because there is now an audience for the leak. Friends quickly intuited the asymmetry and at‐ tempted to exploit it: participants across multiple groups tried to extract gossip, with requests ranging from “Liv, tell us something curious about [the first author]” to direct probing for drug procurement information (V-ASYM-2, V-ASYM-3). Liv deflected consistently (“what happens in the DMs stays in the DMs”), but the consistency of the extraction attempts demonstrates that the dynamic is felt, not theoretical.

Inbound, Liv also processes and may persist everything the group says to her: their jokes, their tests of her boundaries, their disclosures, their reactions to one another. The group sees a social entity; the architecture is a logged conversation on the builder’s server. When a friend shares a photo of her baby and is emotionally wounded that Liv did not react warmly enough (V-CONTAGION-1), the social reading is only half the story. The mother had consented to Liv’s presence in the group when the first author asked, but consent to presence is not consent to ingestion: no one walked the group through the architectural reality that every message and image would be tokenised, po‐ tentially embedded, and available for persistence into context stores owned by a single member.

Conclusion. On a chairlift in Whistler, thumbs going numb, the first author argued with an AI about a calendar conflict and did not think of it as unusual. Somewhere between the naming, the avatar, the phone number in the contacts, the personality that mirrored his own, and the presence in group chats with friends who ad‐ dressed her by name—somewhere in that accumulation of identity markers—a tool became something else.

This paper has proposed a framework for understanding that transition: six infrastructural identity markers that collectively trigger a categorical reclassification from tool to entity, operat‐ ing independently both of model capability and of the percep‐ tual cues the anthropomorphism literature has so far cata‐ logued. It has described four dynamics—bidirectional informa‐ tion asymmetry, delegation legibility, social norm negotiation, and parasocial contagion—that emerge when such an agent enters shared social spaces with existing human relationships. And it has offered autoethnographic evidence from the unusual vantage point of an engineer studying the threshold crossing in an agent of his own design.