Source-linked AI summary
Why did My Robot Just Change Personality? Prompting Guidelines for a Grounded Robot Persona in LLM-Based HRI
Ashita Ashok, Franziska Babel, Patrick Holthaus, Rucha Khot, Karla Bransky, Fethiye Irmak Dogan, Karsten Berns, Silvia Rossi, Minha Lee, Guy Laban
TL;DR
LLM-based HRI lacks explicit prompt practices for grounding robot personas, capabilities, and behavioural boundaries. The paper combines a theoretical framework and literature review with expert workshop data to develop an eight-component prompt template. The findings emphasise limited persona legibility, user adaptation, and safety, deception, and governance concerns.
Problem
Prompt design in LLM-based HRI is underspecified, allowing robots to present hallucinated capabilities, unclear behavioural boundaries, and misleading personas.
Method
The paper combines HRI scholarship, a review of recent LLM-based HRI prompts, and survey and discussion data from 27 HRI experts at the Robo-Identity workshop.
Results
The paper develops an eight-component prompt template and finds that prior prompts emphasise identity and task while often leaving capability boundaries, transparency, adaptation, privacy, and ethics implicit.
Takeaways & Limitations
Prompt design should be treated as a socio-technical design and reporting problem requiring explicit capability boundaries, transparent assumptions, and context-sensitive safeguards.
Abstract
from arXiv · showhide
Large language models (LLMs) are increasingly used for verbal interaction in social robots, yet prompt design in human-robot interaction (HRI) remains underspecified. As a result, robots may present hallucinated capabilities, unclear behavioural boundaries, and misleading personas. This paper develops a framework for prompt design in LLM-based robots and introduces a structured prompt template comprising eight functional components through which robot behaviour can be specified, bounded, and adapted. The framework is grounded in a review of prior LLM-based HRI work and complemented by survey and discussion data from HRI experts gathered at the Robo-Identity workshop at IEEE RO-MAN 2025 (N=27). The qualitative findings highlight limited legibility of robot personality, the need for user adaptation, and strong ethical concerns about safety, deception, and governance. Based on these findings, we present prompting guidelines accompanied by proof-of-concept template as a structured design and reporting aid for HRI research. We argue that prompt design should be treated as a socio-technical problem rather than a minor implementation detail, requiring explicit capability boundaries, transparent behavioural assumptions, and context-sensitive safeguards to support reliable and interpretable HRI.
I. INTRODUCTION
The paper frames prompt design as a central mechanism for grounding LLM-powered robot behaviour in actual capabilities and interaction context. It proposes a structured template to make robot identity, capabilities, disclosures, and behavioural boundaries explicit.
- LLMs enable natural-language conversation in social robots but can hallucinate knowledge, perception, or capabilities that robots do not possess.
- Grounding robot behaviour in its capabilities and interaction context is critical for calibrating user expectations and trust.
- The paper combines HRI scholarship and a survey and discussion with HRI experts to identify how prompt components specify and bound robot behaviour.
- The resulting template comprises eight functional components covering identity, capability boundaries, transparency, tasks, failure protocols, privacy, user adaptation, and ethical red lines.
- The template is presented as a structured design and reporting aid rather than a fixed recipe for every robot or context.
II. THEORETICAL FRAMEWORK
The theoretical framework treats prompt design as a mechanism for configuring robot behaviour and artificial personas in embodied interaction. It links identity, transparency, expectations, explanations, embodiment, and behavioural boundaries.
- Prompts shape how robots are presented, what they are expected to do, and how their conduct is constrained in social interaction.
- Identity: Robot identity emerges through embodiment, behavioural cues, social dynamics, and users’ constructions of robot personas.
- Identity: Operationalising identity requires specifying a coherent role, persona, communication style, expressiveness, tone, and constraints from embodiment and capabilities.
- Transparency: Transparency by Design makes robot intentions, decisions, and internal states intelligible to support safety, predictability, trust, and collaboration.
- Transparency: Deceptive social and emotional cues may increase compliance or perceived agency but can undermine trust and require safeguards.
- Explanations and Expectations: Explanations and expectations form a coupled design problem because explanations communicate capability, limitation, and social role while expectations shape their interpretation.
III. PROMPT DESIGN IN LLM-BASED HRI
The literature review maps which behavioural aspects recent LLM-based HRI prompts specify, constrain, or leave implicit. It finds strong emphasis on identity and task, alongside limited reporting of capability boundaries, transparency, adaptation, privacy, and ethics.
- The review surveyed selected recent LLM-based HRI studies that explicitly reported prompt design and examined specified, constrained, and implicit behavioural aspects.
- Literature findings: 90% of reviewed prompts define robot identity and 60% specify the interaction task.
- Literature findings: Expectation or failure-handling strategies appear in 50% of prompts, while explicit user adaptation appears in 30%.
- Literature findings: Privacy-related instructions appear in 20% of prompts, and explicit ethical constraints appear in 30%.
- Literature findings: None of the reviewed prompts specifies embodiment or system capability boundaries, and none discloses LLM use to users.
- Proposed template: The proposed eight components define persona, capability limits, transparency, task, uncertainty handling, privacy, user adaptation, and ethical red lines.
IV. QUALITATIVE ASSESSMENT
The qualitative assessment adds participatory evidence from HRI experts about how robot personas are interpreted and governed. Participants viewed persona as potentially useful but contested, with safety, deception, and context central to its legitimacy.
- The empirical layer uses a participatory expert survey and discussion to examine persona legibility, continuity, safety boundaries, and context-sensitive governance.
- The workshop included 27 HRI experts divided across three groups with 10, 8, and 9 participants.
V. RESULTS OF THEMATIC ANALYSIS
The qualitative-led analysis found broad agreement across groups, with disagreement concentrated on items 2, 3, 4, 5, and 11.
- Groups mostly agreed about most statements, while items 2, 3, 4, 5, and 11 were exceptions.
A. Persona as a Useful but Contested Design Abstraction
Participants viewed robot persona as potentially useful but neither categorically rejected nor unconditionally endorsed it. Its legitimacy depended on how persona was defined and what purpose it served.
- Participants treated robot persona as a potentially useful design abstraction rather than inherently desirable or unacceptable.
- Participants questioned whether companion-like behaviour was normatively appropriate, including how personas could accommodate cultural variation.
B. Robot Personality Is Not Necessarily Legible
Participants reported that robot personality is difficult to recognize reliably because subtle cues may be missed and interpretation varies across users. The associated statements covered persona design, identity, anthropomorphism, AI-driven personas, and ethics.
- Participants questioned whether one robot persona could adequately model cultural differences and rejected a one-size-fits-all framing.
- The survey grouped statements into personas as design constructs, anthropomorphism and deception, AI-driven personas, and deception and ethics.
- Subtle behavioural cues were often missed, and neurodiversity could affect personality perception and recognition.
- Personality was considered clearer under exaggerated conditions, while interpretation depended on user experience.
C. Identity Over Time Was Framed as a Memory Problem
Participants framed identity continuity over time as dependent on memory and system architecture rather than prompt design alone.
- Longitudinal persona was limited by context, and LLMs alone could not sustain continuity without memory of prior interactions and temporal sequencing.
- Adaptive identity was conditioned on mechanisms that retain and organize interaction history.
D. Safety and Ethics Structured the Discussion
HRI experts treated hallucination, manipulation, deception, and harmful behaviour as central persona risks, while stressing that safeguards must account for context, culture, and user vulnerability.
- Participants linked robot persona and identity to risks from hallucination, manipulation, and misleading representations, including harms from verbal outputs.They described safeguards as necessary and noted that deception’s effect on trust depends on context and whether it is detected.
- Participants rejected one-size-fits-all persona design, emphasising adaptation to context, culture, user preference, and vulnerability.They also raised concerns about identity continuity across embodiments and increased caution for vulnerable groups.
- Explicit boundaries identified as unacceptable included lack of transparency, real-person personas, harmful behaviour, and privacy risks.Participants additionally emphasised the need for ethical oversight.
F. Overall Interpretation
Experts viewed persona as potentially useful only when it is legible, supported by continuity, and constrained by strong safety and ethical requirements. They were most aligned on risks of hallucination, manipulation, and deception, while questioning stable personality perception and broad generalisation.
- Persona was considered potentially useful when supported by legibility, continuity, safety, and ethical constraints.
- Agreement was strongest on risks involving hallucination, manipulation, and deception.
- Participants were sceptical that users reliably perceive stable personality or that persona generalises broadly as a design tool.
VI. DISCUSSION AND CONCLUSION
The discussion combines a literature-derived prompt structure with expert findings that elevate capability boundaries, transparency, safety, and context sensitivity. It frames the template as a design and reporting aid, while extending identity design toward embodiment, memory, privacy, and temporal coherence.
- Discussion and Conclusion: Existing LLM-based HRI systems most often report IDENTITY and TASK, whereas CAPABILITY BOUNDARY and TRANSPARENCY are less consistently specified.Experts linked this imbalance to potentially illegible, poorly adapted, deceptive, or harmful robot personas.
- Discussion and Conclusion: Safety, deception, hallucination, and governance emerged as consequential practical concerns, making underreported prompt components important for responsible deployment.
- Discussion and Conclusion: Prompt-based identity specification is necessary but insufficient because identity also depends on embodiment, behavioural cues, and interaction context.Coherent behavioural cues and embodiment-consistent signalling are needed for persona to become legible in interaction.
- Discussion and Conclusion: Sustaining a stable persona across encounters requires coordinating prompt design with memory architecture and governance of stored user information.The discussion frames longitudinal identity primarily as a memory problem involving privacy and temporal coherence.
- Discussion and Conclusion: The template is intended as a structured design and reporting aid rather than a fixed recipe for every robot or context.It helps researchers make prompt assumptions explicit, comparable, and open to review.
- Discussion and Conclusion: Explicit capability limits, such as lacking physical manipulation or internet access, can calibrate expectations and reduce misleading interaction.Future work should test the guidelines across different robot embodiments and evaluate users’ perception of robot identity.