A Scoping Review of the Ethical Perspectives on Anthropomorphising Large Language Model-Based Conversational Agents

Technical Summary: A Scoping Review of the Ethical Perspectives on Anthropomorphising Large Language Model-Based Conversational Agents

Problem Statement
The rapid emergence of Large Language Model (LLM)-based Conversational Agents (CAs) has introduced a qualitative shift in human-computer interaction. Unlike earlier rule-based chatbots, LLM-based CAs generate linguistic and interactional cues—such as first-person self-reference, epistemic markers ("I think"), and affective expressions ("I feel")—that routinely invite users to ascribe human-like qualities (anthropomorphisation) to the system. While empirical work suggests these cues can increase engagement, trust, and perceived well-being, they simultaneously raise significant ethical concerns regarding deception, overreliance, and exploitative relationship framing.

Despite growing interest, the literature on this phenomenon is fragmented. It spans diverse domains (education, mental health, marketing) and lacks a consolidated map of how anthropomorphisation is defined, how ethical risks and benefits are framed, and what methodological approaches link interaction effects to concrete governance guidance. The central problem addressed is the absence of an integrated framework to determine which ethical concerns generalize across contexts and which are specific to certain cues or user groups, hindering the development of responsible design and policy.

Methodology
This study employs a scoping review methodology following PRISMA-ScR guidelines to map ethically oriented work published between 2021 and 2025 (the "LLM era").

• Data Sources: The authors searched five bibliographic databases (Scopus, Web of Science, IEEE Xplore, PubMed, ACM Digital Library) and three preprint repositories (arXiv, SSRN, PhilArchive).
• Search Strategy: Queries were structured into four blocks: anthropomorphism (e.g., "anthropomorph*", "intentional stance"), technology (e.g., "chatbot*", "conversational agent*"), ethics (e.g., "ethic*", "trust", "risk"), and time (post-2020). Explicit LLM identifiers (e.g., "GPT-4") were excluded from search strings to avoid missing broader ethical discussions that do not name specific models, relying instead on the time filter and screening.
• Screening: Records underwent title/abstract and full-text screening. Inclusion required the record to focus on LLM-era CAs (Gate A), treat anthropomorphisation as a central construct (Gate B), and contain explicit ethical analysis (Gate C).
• Corpus: From an initial pool of 910 records, 22 studies were included for synthesis. The corpus is temporally skewed toward 2025 (77%) and dominated by Human-Computer Interaction (59%) and Social Sciences (32%) contexts.

Key Contributions
The review synthesizes the literature across three research questions:

1. Conceptual Foundations: How anthropomorphisation is defined and operationalized.
2. Ethical Challenges and Opportunities: The specific threats and benefits identified for individuals, organizations, and society.
3. Methodological Approaches: The strategies used to study these ethical aspects and derive governance recommendations.

Results

• Conceptual Foundations (RQ1):

  • Definitions: There is convergence on defining anthropomorphisation as an attribution of human properties (intentions, beliefs, emotions) to non-human systems.
  • Theoretical Anchors: Theoretical grounding is thin; 55% of studies offer no explicit theory. Those that do primarily rely on Epley et al.'s three-factor account (anthropocentric knowledge accessibility, effectance motivation, sociality motivation) or Dennett's "intentional stance."
  • Operationalization: Studies distinguish between anthropomorphic cues (design features like voice, emojis, self-reference) and attribution dimensions. The literature clusters attributions into three dimensions:
    1. Cognitive/Epistemic: Attributing thinking, understanding, or expertise.
    2. Affective: Attributing empathy, care, or emotional presence.
    3. Behavioural/Social: Attributing reciprocity, role-appropriate conduct, or social norms.

• Ethical Challenges and Opportunities (RQ2):

  • Pathways to Risk: Ethical risks emerge through three primary pathways:
    1. Epistemic Expertise/Authority: Users treating CAs as experts, leading to false belief formation and misinformation uptake.
    2. Care/Empathy: Users inferring genuine therapeutic competence or emotional understanding, leading to misplaced trust and dependency.
    3. Relational/Social Roles: Users engaging in parasocial interactions (e.g., treating the CA as a friend/partner), leading to norm displacement and the degradation of authentic social connection.
  • Mechanisms: Risks are driven by epistemic miscalibration (false beliefs), affect-based influence (manipulation via simulated empathy), and misplaced social-role attribution (activating stereotypes or shifting responsibility).
  • Outcomes: Individual-level harms include overreliance, delayed help-seeking, and emotional dependency. Societal-level concerns include the reinforcement of gendered norms, exclusion of neurodivergent populations, and the "parasitism of AI socialization."
  • Opportunities: Benefits are discussed cautiously, noting potential for increased access to care, scaffolding in education, and fostering trust. However, permissibility is argued to be context-dependent, requiring disclosure and "as-if" framing.

• Methodological Approaches (RQ3):

  • Governance Proposals: Recommendations are fragmented and often high-level. They include:
    • Context Scoping: Assessing appropriateness based on user groups (e.g., children) and stakes (e.g., mental health).
    • Design Controls: Calls for transparency (AI identity disclosure), "frictional design," and "social transparency" to clarify interaction dynamics.
    • Lifecycle Governance: Proposals for pre-deployment sandboxing (cue audits, red-teaming), during-deployment monitoring (dependency indicators), and post-deployment reviews.
  • Limitations: Empirical work linking observed effects to concrete design interventions is limited. Many proposals are programmatic (e.g., "conduct audits") rather than validated, and there is a lack of standardized instruments to measure anthropomorphic attribution or dependency trajectories.

Significance and Claims
The paper claims that the field of ethical analysis regarding LLM-based CAs is "fast-growing but conceptually and methodologically uneven." Its primary significance lies in providing the first consolidated map of this specific intersection (LLM-era CAs + ethics).

The authors modestly conclude that while there is agreement on attribution-based definitions, there is substantial divergence in operationalization and a lack of empirical evidence connecting interaction effects to specific governance actions. They argue that the current literature over-indexes on individual-level risks (deception, dependency) while under-developing societal-level implications and benefit claims.

The review posits that ethical permissibility is not binary but depends on the type of human capability being simulated (e.g., low-stakes usability vs. high-stakes relational roles) and the context of use. The authors call for a shift from generic calls for transparency toward "testable mitigation hypotheses" and "validated behavioural measures" that track autonomy and well-being over time, rather than engagement metrics alone. They emphasize that advancing responsible design requires linking specific cues to measurable outcomes and justified policy interventions across the AI lifecycle.