When Single-User-Oriented LLM-based Assistants Involve Others: A Scoping Review of Pathways, Risks, and Responses

LLM-based assistants are increasingly extending into multi-party contexts, while core operational processes for context management, personalization, identity attribution, authority attribution, and action execution often remain organized around a single user. Existing work examines particular multi-party settings, but lacks a systematic account of how these single-user-oriented assistants begin to involve additional human parties and what risks emerge. To address this gap, we conducted a scoping review of 58 studies. We identify five operational pathways spanning direct and indirect involvement, five recurring risk domains, and five areas of implemented and proposed responses. Based on these findings, we argue for governance that attends to changing cross-person roles and relationships in practice, and for assistant designs that preserve person-specific boundaries throughout interaction.

Introduction. Powered by advances in large language models (LLMs), AI1 assistants are becoming increasingly integrated into people’s everyday lives as long-term companions, personalized co-workers, and agents that carry out tasks on users’ behalf in the external world [23, 50, 70, 81, 92]. Much of this work treats the individual user–AI assistant interaction as the basic unit of design. An assistant’s tone and voice can be tailored to one user’s preferences, recurring interactions can help it remember that user’s habits and prior conversations, and more agentic assistants can plan or carry out tasks around that user’s goals and requests [23, 85]. We refer to this organization of information and authority around a primary user as a single-user assumption. Despite this progress, the single-user assumption becomes increasingly difficult to sustain. Earlier research has already shown that intelligent assistants in settings such as families and organizations inherently encounter multiple users [41, 61, 79, 80]. More recent work shows that LLM-based assistants designed around an individual user are increasingly used in shared and collaborative settings, including family conversations and group planning among friends [14, 58, 91]. Additional people can also become involved without directly using the assistant [33, 90], as their speech or other information enters the assistant’s context through shared environments or another user’s resources. In these settings, information or permissions organized around one user can extend to others, creating consequential effects beyond the initiating user by exposing stored data to co-users, enabling sensitive inferences about non-users, or treating third-party speech as the primary user’s own input [30, 33, 57]. Existing research, however, captures only parts of this transition. Multi-party dialogue reviews focus mainly on computational challenges within conversations involving multiple speakers [19]; group conversational agent research examines systems intentionally designed to intervene in synchronous group interaction [88]; and recent multi-user LLM work formalizes direct multi-principal settings involving conflicting objectives, authority, privacy, and coordination [85]. These perspectives provide limited coverage of broader operational settings in which single-user-oriented assistant operation begins to involve other people through shared use, cross-user delegation, shared state or infrastructure, information about non-users, or sensing in shared environments. They also do not systematically connect these forms of involvement to their resulting risks, existing responses, and remaining gaps. To address this gap, we conduct a scoping review of LLM-based assistants operating beyond the single-user assumption. We use multi-party operation as an inclusive term covering both additional users who interact with an assistant and individuals who become involved indirectly through its information processing, sensing, resource use, or actions. We reviewed research published between

Related work. Large language models have increasingly been incorporated into conversational assistants and agentic systems that extend beyond one-shot text generation. LLM-based assistants can maintain conversational context and persistent memory, reason over longer task sequences, invoke external tools and APIs, and perform actions on behalf of users [50, 70, 83]. These capabilities have expanded the role of LLMs from conversational interfaces toward assistants that support information seeking, planning, decision making, workplace tasks, and other forms of delegated activity [70, 85]. Despite this expansion in capability, many contemporary LLM assistants remain organized around an individual user. Conventional conversational systems have historically modeled interaction as a dyad between one user and one system [19, 56, 93]. Contemporary LLM interfaces largely inherit this structure. Standard chat templates typically represent human input through a single “user” role, while instruction tuning conditions the model on a unified conversational context and optimizes for an assistant response to that user [85]. Preference-based alignment further commonly reduces feedback to an aggregated or scalar objective rather than explicitly representing multiple independently situated users with distinct objectives and authority [44, 85]. Our review builds on this account of the single-user assumption to examine how it is challenged as assistant operation extends beyond one user to involve additional human parties.

2.2 Prior Reviews of AI Beyond Single-User Interaction Prior reviews have examined conversational AI beyond dyadic or single-user interaction from several perspectives. Ganesh et al. [19] focus on computational multi-party dialogue modeling, including speaker/addressee recognition, turn-taking, disentanglement, response generation, and information flow, but largely cover pre-LLM approaches. Yeo et al. [88] review agents that sense and intervene in group interaction, with LLMs treated alongside rulebased and Wizard-of-Oz systems. Lei et al. [34] trace multi-party dialogue toward LLM-driven and multi-agent systems, but remain organized around MPD tasks, methods, datasets, and evaluation. Yang et al. [85] more directly examine the transition from singleto multi-principal LLM agents, emphasizing instruction conflict, access control, and coordination in shared-agent settings. These lines of work leave a broader transition less systematically examined. Multi-party dialogue research generally begins with a conversation that already contains multiple speakers [19, 34], while group conversational agent research [88] begins with agents intentionally designed for group-level intervention.

Method. 2.3 Conceptual Scope of This Review In this review, we use an LLM-based assistant to refer to an interactive AI system in which a large language model plays a central role in interpreting user input and contextual information, generating responses or plans, or coordinating actions, tools, and resources on behalf of a user. This scope includes conversational assistants and more agentic forms of assistance that can maintain state, access external resources, or perform delegated actions. It does not refer to AI systems generally, or to algorithmic systems. Unless otherwise specified, the term assistant throughout the remainder of this paper refers specifically to an LLM-based assistant in this sense. We also distinguish multi-user, non-user, and principal. We use multi-user for settings in which multiple people directly use, access, address, or act through an assistant. A non-user does not directly operate the assistant but becomes implicated through represented information, sensing in a shared environment, or exposure to assistant outputs or actions. Following prior work on multi-principal LLM agents [85], principal is used to refer to a user whose objectives, instructions, authority, or resources the assistant is expected to represent or act upon. Rather than treating multi-party AI as a predefined system category, this review starts from LLM assistants ordinarily organized around a single user and examines their transition into operation involving additional human parties. We use operation broadly to include interaction and context processing, inference and sensing, resource or tool access, output generation, and delegated actions. Our central unit of analysis is this single-user-to-multi-party transition and its identifiable consequences for the primary user, additional direct users, indirectly involved individuals, and their relationships, allowing responses to be examined in relation to the specific cross-person operation and risks they address.

We conducted a scoping review of how LLM-based assistants organized around a primary user extend beyond that user to involve When Single-User-Oriented LLM-based Assistants Involve Others: A Scoping Review of Pathways, Risks, and Responses , , additional human parties. We followed Arksey and O’Malley’s fivestage scoping-review framework [4], covering research-question development, study identification and selection, data charting, and synthesis. Study identification combined database searches with a supplementary Google Scholar search and backward and forward snowballing following Wohlin [78]. We used the PRISMA Extension for Scoping Reviews (PRISMA-ScR) [68] to guide transparent reporting of the search, study selection, data charting, and synthesis procedures.

3.1 Research Questions The review was guided by the following research questions:

• (RQ1) How do multi-party contexts extend AI assistant operation beyond the single user? • (RQ2) What risks are reported when AI assistants operate in multi-party contexts beyond the single user? • (RQ3) What design responses have been implemented or proposed to address these risks, and what research gaps remain?

3.2 Databases and Corpus Selection 3.2.1 Databases. Research on LLM-based assistants in multi-party contexts spans HCI, NLP, AI and machine learning, security, robotics, and agent systems. We therefore searched five complementary sources: ACM Digital Library for HCI, CSCW, interaction design, and social computing; IEEE Xplore for AI systems and robotics; OpenReview for major machine-learning and related AI venues; ACL Anthology for broader archival coverage of NLP, conversational systems, and language-agent research not fully represented in OpenReview; and arXiv for rapidly developing preprint research across these areas. We considered research from January 2022 to September 2026. The starting point was chosen to capture the emergence of publicly accessible LLM-based conversational assistants, particularly around the release of ChatGPT in late 2022, when such systems began to enter widespread use [46, 76].

Discussion. 8.1 Why Single-User-Oriented Assistants End Up Operating in Multi-Party Contexts Our RQ1 findings (Section 5) show that LLM-based assistants organized around a single user extend to others through shared use, connected resources, sensing, and delegated interaction. These pathways echo prior work on group conversational agents, multi-party dialogue, and multi-principal LLM systems [19, 85, 88]. These operational pathways arise from different drivers. Some are functionally motivated because tasks depend on information, preferences, or actions distributed among several people. Information-gathering agents, for example, rely on collaborators’ knowledge when the initiating user’s local context is insufficient [28], while cross-user memory sharing reuses information acquired from one user for another [52]. Other extensions arise from practical use, such as sharing LLM accounts to divide subscription costs [62]. Still others occur incidentally when connected resources, persistent state, or ambient sensing bring other people’s information or presence into assistant operation without intentional participation. Despite these different drivers, they produce a common shift that functions organized around one user begin operating in relation to others. This helps explain why RQ2 risks (Section 6) recur across different pathways. A speaker may not be the intended user, information in one person’s context may concern another, a collaborator may contribute without controlling the resulting action, and an affected person may never interact with the assistant. The challenge is thus not simply the presence of more people, but the extension of user-scoped functions into relationships where identity, information, preferences, authority, and interests are no longer aligned. This shifts the design question from whether an assistant is formally “multi-user” to how it accounts for the additional people it comes to involve.

8.2 Policy Implications for AI Assistant Governance Existing AI governance frameworks already provide foundations for considering people beyond the immediate user. The OECD Framework [18] distinguishes users from directly and indirectly impacted stakeholders and relates AI actors to different system dimensions and lifecycle stages. The NIST AI Risk Management Framework [2, 5] similarly emphasizes context, affected stakeholders, organizational responsibilities, monitoring, and feedback throughout the AI lifecycle. The EU AI Act [16] translates some of these concerns into requirements for specified systems, including risk management, transparency, record-keeping, human oversight, and post-market monitoring. For specified high-risk deployments, its fundamental rights impact assessment considers categories of people and groups likely to be affected in the context of use. These frameworks therefore recognize that affectedness can extend beyond direct users and that relevant roles vary by lifecycle stage and deployment context. Our RQ1 findings (Section 5) echo this framing by showing distinct forms of direct and indirect human involvement in assistant operation. Multi-party assistant risks can turn on finer cross-person relationships within ongoing operation. The risks identified in , , RQ2 (Section 6) often depend on whose information is accessed or reused, whose preference is represented, whose instruction carries authority, whose resource is acted upon, and who is exposed to the resulting output. These relationships may coexist or change within the same interaction and are less explicitly captured by broader categories such as user, affected stakeholder, deployer, or provider. Multi-party governance therefore also depends on representing these relationships where information, state, and action are handled.

Conclusion. This scoping review clarifies how LLM-based assistants extend beyond the single-user assumption as additional people become involved in assistant operation. Across 58 studies and 118 risk instances, we identified five operational pathways, five recurring risk domains, and five response groups that together map how multi-party involvement emerges, what risks it produces, and how existing work has responded. This review provides a structured account of these extensions and their associated risks, while highlighting the importance of governance and design that preserve person-specific relationships as assistants sense, remember, communicate, coordinate, and act. As LLM-based assistants become more persistent and agentic, accounting for multi-party involvement should become a routine part of assistant design rather than an exceptional case.

Limitations. 8.4 Limitations This review has several limitations. First, our study-identification strategy focused primarily on scholarly databases and research repositories and did not systematically cover grey literature or legal materials, such as industry reports, policy guidance, or legal cases. Relevant risks and implementation practices documented outside academic research may therefore be underrepresented. Terminology also remains heterogeneous across research communities, particularly for LLM assistants, multi-user interaction, and indirectly involved actors, so some relevant studies may have been missed despite our broad search strategy and supplementary citation searching. Second, we included arXiv alongside peer-reviewed sources to capture emerging work not yet archivally published. Thirty of the 58 included studies were arXiv preprints, and their findings may change through further revision or evaluation. The evidential basis of the identified risks also varied. Of the 118 risk instances, 94 were empirical, meaning that the risk was supported by evidence generated within the study itself, while 24 were conceptual or author-asserted, meaning that the risk was articulated through problem framing, design rationale, discussion, or conceptual analysis rather than directly demonstrated by the study’s own empirical results.