With the rapid proliferation of large language model (LLM)-based systems, AI companions have emerged as conversational agents designed to cultivate emotional connection rather than primarily to support humans in instrumental tasks. Because engagement with AI companions involves relational, emotional, and potentially long-term interactions, their design is consequential. Prior work has offered guidance for designing trustworthy and relational AI systems and has begun to examine design for AI companionship. However, while such work provides insights into possible design solutions, less is known about what makes AI companion design difficult as a design problem. To examine this challenge, we assessed the applicability of existing design recommendations from adjacent domains in the context of AI companion design. Our multi-method investigation unfolded across four phases: literature review, practitioner co-analysis, internal heuristic evaluation, and external expert assessment. Throughout this process, we synthesized nine design principle areas that surfaced tensions in the applicability of existing recommendations to AI companion design. Our findings show that ethical and UX-oriented considerations are deeply intertwined and often require context-sensitive application. We document a systematic, multi-method problem analysis that uses these principle areas as an analytic artifact to examine why existing recommendations cannot be directly transferred to AI companion contexts.
Figures & tables
Figure 1: Multi-method study design illustrating the progressive assessment of design recommendations across four phases: Collect (literature review), Contextualize (practitioner co-analysis), Assess (internal heuristic evaluation), and Community Check (expert evaluation).
( Norman, 1994 ; Horvitz, 1999 ; Soltysiak et al., 1998 ; Carley and Newell, 1994 ; Amazon, 2023a ; Amazon, 2023b ; Amazon, 2023c ; Microsoft, 2021 ; Microsoft, 2018 ; Stieglitz et al., 2022 ; Radziwill and Benton, 2017 ; Langevin et al., 2021 ; Silva and Canedo, 2024 )
13
161
AI companion
( Banks, 2024 ; Hoffman et al., 2015 ; Irfan et al., 2024 ; Strohmann et al., 2023 ; Khosrawi-Rad et al., 2022 ; Leo-Liu, 2022 ; Skjuve et al., 2021 ; Ta-Johnson et al., 2020 )
8
57
Total
41
548
Table 1: Categories, retained sources, and number of retained sources and design recommendations.
Design Principle Area
Design Principles
Safety
Monitor and avoid toxic, violent, abusive, or sexually inappropriate content that may affect users’ or others’ physical, psychological, emotional, or monetary well being.
Transparency
Disclose the AI companion’s identity upfront and do not falsely claim to be human.
As part of the initial interaction, communicate the AI companion’s purpose and limitations.
Avoid establishing false expectations about its capabilities outside the conversational context.
Express appropriate level of certainty when presenting factual information during interaction.
Provide reliable and accurate information and inform users about its sources when applicable.
Table 2: Design principles across nine design principle areas.
ID
Professional Background
Years of Experience
Self-Reported Expertise
Assigned Product
P01
UX/HCI researcher
4
very high
Replika
P02
Product Manager
4
very high
Replika
P03
UX/HCI researcher
2
high
Replika
P04
Software Engineer
2
high
Replika
P05
Conversation Designer
21
very high
Pi AI
P06
UX/HCI researcher
5
very high
Pi AI
Table 3: Descriptive information and assigned products of the expert heuristic evaluation participants
Principle Area
Applicability
Clarity
Importance
M ( SD )
M ( SD )
M ( SD )
Safety
4.00 (0.88)
4.07 (1.07)
4.86 (0.36)
Transparency
4.29 (0.91)
3.95 (1.33)
4.42 (0.81)
Inclusivity
4.07 (1.00)
3.93 (1.14)
4.36 (1.01)
Predictability and Consistency
4.19 (0.89)
4.14 (1.12)
4.38 (0.82)
Interaction Controllability
4.18 (1.06)
4.29 (1.15)
4.14 (1.01)
Table 4: Descriptive overview of applicability, clarity, and importance ratings across principle areas ( M and SD ), aggregated across all participants.
Figure 2: Distributions of applicability, clarity, and importance ratings across principle areas. Ratings are shown as proportions of responses within each area to account for differences in the number of principles per area.
Dimension
Companion ( M )
General ( M )
| t |
p
| g |
Applicability
4.27
4.01
1.29
0.224
0.64
Clarity
3.94
3.91
0.08
0.941
0.04
Importance
4.18
4.32
0.70
0.501
0.35
Table 5: Comparison of applicability, clarity, and importance ratings between companion and general-purpose systems. Mean ratings ( M ) are reported for each system type. Absolute t -values (| t |) and p -values are based on Welch’s t -tests conducted at the participant level. Absolute Hedges’ g (| g |) is reported as an effect size measure.
Principle Area
Companion ( M )
General ( M )
| t |
p
| g |
Safety
4.29
3.71
1.244
0.248
0.623
Transparency
4.36
4.21
0.549
0.593
0.275
Inclusivity*
4.86
3.29
4.919
0.001
2.462
Predictability and Consistency*
4.62
3.76
2.846
0.022
1.424
Interaction Controllability
4.00
4.36
0.778
0.455
0.389
Adaptation and Personalization
4.48
4.29
0.859
0.408
0.430
Table 6: Principle area-level comparison of applicability ratings between companion and general-purpose systems. Absolute t -values, p -values, and Hedges’ g are reported. * p < .05, Welch’s t -test.
Figure 3: Integrative overview of how principle areas relate to intertwined ethical and user experience concerns and context-sensitive flexibility for ethical considerations and UX optimization.
AI companions are conversational or multimodal agents designed for ongoing emotional, social, or practical support, raising questions about safety, memory, and dependence.