Designing Social Robots with Ethical, User-Adaptive Explainability in the Era of Foundation Models

Fethiye Irmak Doğan, Alva Markelius, Hatice GüneşarXiv 2026 · DOI · arXiv

TL;DR

As foundation models (LLMs and VLMs) become the "brains" of social robots, their ability to adapt to users creates a new transparency crisis. This paper argues that "one-size-fits-all" explanations are no longer ethically viable and proposes a framework for user-adaptive, co-designed explainability grounded in smaller, fairer datasets to prevent bias and "cognitive ableism."

What Problem Does This Paper Address?

Traditional Socially Assistive Robots (SARs) operated on rigid, rule-based systems where explanations were hard-coded and predictable. Today, robots powered by foundation models can adapt their language and behavior in real-time based on vast, opaque datasets.

The authors identify a critical gap: while these robots are becoming more "adaptive," their explanations remain generic. This creates several risks:

  1. Misalignment: A robot might adapt its behavior based on a flawed inference (e.g., misinterpreting neurodivergent behavior as "discomfort").
  2. Epistemic Exclusion: Models trained on Western-centric data may marginalize non-hegemonic ways of knowing or communicating.
  3. The Transparency Gap: Users often don't know why a robot has changed its behavior toward them, leading to misplaced trust or confusion.

Methods

This is a position and perspective paper that maps the emerging ethical landscape of foundation-model-driven Human-Robot Interaction (HRI). The authors synthesize current literature to identify open challenges and propose a new design philosophy.

Their approach centers on four core recommendations:

  • Modality-Aware Strategies: Tailoring explanations to user needs (e.g., using visual aids for children or non-verbal cues in specific cultural contexts).
  • Co-Design: Involving end-users (especially those from marginalized groups) in deciding how and what a robot should explain.
  • Layered Adaptation: Balancing the technical impossibility of "per-user fine-tuning" with "user-cluster" adaptation (e.g., by age group or accessibility needs).
  • Data Prioritization: Shifting away from "big data" toward smaller, curated, and fairer datasets to anchor robot behavior in community-informed preferences.

Datasets

The paper does not introduce a new empirical dataset. Instead, it critiques the use of massive, uncurated internet-scale corpora (the standard for foundation models) in social robotics. It advocates for the creation and use of smaller, co-designed, and curated datasets that foreground underrepresented communities to ensure explanations are ethically grounded.

Key Insights and Ideas

  • Cognitive Ableism in AI: One of the most striking insights is how foundation models might pathologize neurodivergence. Because they are trained on neurotypical data, a robot might interpret a neurodivergent user's gaze or speech patterns as a "deficit" and offer an explanation that reinforces stigma (e.g., "I am giving you space because you seem uncomfortable").
  • The "Moral Alibi": The authors warn that developers might use the "neutrality" and scale of pre-training as an excuse to diffuse accountability for biased robot behavior.
  • The Illusion of Competence: Tailored, fluent explanations can be deceptive. They may lead users to believe a robot has a deep internal understanding of a situation when it is actually just performing sophisticated pattern matching.
  • Minimalism as a Virtue: Unlike many XAI (Explainable AI) papers that push for more explanation, the authors suggest that "user-adaptive" also means knowing when to provide minimal or on-demand explanations to avoid cognitive overload in sensitive settings like mental health coaching.

Key Results

The paper provides a conceptual framework rather than quantitative benchmarks. Its primary "results" are the identification of three interrelated challenges:

  1. Pre-training biases: Inherited patterns that don't align with situated HRI contexts.
  2. Persistence of generic explanations: The failure of current systems to reflect a user’s specific background or abilities.
  3. Opacity of adaptive mechanisms: The lack of transparency regarding what is being adapted and why.

The authors ground these findings in an illustrative use case of an LLM-driven robot in a school/clinical setting, demonstrating how layered adaptation (R3) and co-designed modalities (R2) would look in practice.

Limitations and Future Work

The authors acknowledge that their recommendations face significant technical hurdles:

  • Technical Feasibility: Per-user fine-tuning is often impractical due to latency, on-device compute limits, and data scarcity.
  • The Deception Paradox: Making a robot more "explainable" and "adaptive" might inadvertently increase anthropomorphism and over-trust, which is ethically risky.
  • Future Research: The authors call for systematic, empirical investigations into fine-grained adaptation strategies across diverse demographic groups and the development of mechanisms that can "calibrate" the depth of an explanation to a user's current emotional state.

Summary updated

Abstract

Foundation models are increasingly embedded in social robots, mediating not only what they say and do but also how they adapt to users over time. This shift renders traditional "one-size-fits-all" explanation strategies especially problematic: generic justifications are now wrapped around behaviour produced by models trained on vast, heterogeneous, and opaque datasets. We argue that ethical, user-adapted explainability must be treated as a core design objective for foundation-model-driven social robotics. We first identify open challenges around explainability and ethical concerns that arise when both adaptation and explanation are delegated to foundation models. Building on this analysis, we propose four recommendations for moving towards user-adapted, modality-aware, and co-designed explanation strategies grounded in smaller, fairer datasets. An illustrative use case of an LLM-driven socially assistive robot demonstrates how these recommendations might be instantiated in a sensitive, real-world domain.
BibTeX
@misc{dogan2026designing,
  author = {Fethiye Irmak Doğan and Alva Markelius and Hatice Güneş},
  title = {{Designing Social Robots with Ethical, User-Adaptive Explainability in the Era of Foundation Models}},
  year = {2026},
  eprint = {2603.00102},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2603.00102},
}
Figure