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

1. Problem Statement

The integration of Foundation Models (FMs), particularly Large Language Models (LLMs) and Vision-Language Models (VLMs), into social robotics has shifted the paradigm from rigid, rule-based systems to flexible, adaptive agents. However, this shift introduces critical challenges:

• The "One-Size-Fits-All" Failure: Traditional explainability strategies (generic justifications, static visualizations) are insufficient for FMs. These models are trained on vast, heterogeneous, and opaque datasets, making their internal logic difficult to trace.
• The Adaptation-Explanation Gap: While FMs enable real-time behavioral adaptation based on user context, they often generate generic, surface-level explanations that fail to account for user diversity (e.g., cognitive abilities, cultural backgrounds, neurodivergence).
• Ethical Risks: Delegating both adaptation and explanation to FMs creates risks of:
  • Cognitive Misalignment: Pathologizing neurodivergent behaviors or reinforcing ableist biases under the guise of "personalization."
  • Epistemic Exclusion: Centralizing knowledge authority in hegemonic datasets, marginalizing non-dominant ways of knowing.
  • Commodification of Intimacy: Using hyper-personalized, flattering explanations to manipulate user trust and obscure system limitations.

2. Methodology and Approach

The paper employs a conceptual analysis and framework development approach rather than a specific empirical experiment. The methodology involves:

• Literature Synthesis: Reviewing existing work in Human-Robot Interaction (HRI), Explainable AI (XAI), and the ethical implications of Foundation Models.
• Problem Mapping: Identifying the specific friction points between FM capabilities (open-ended generation, probabilistic reasoning) and the requirements of ethical, user-adaptive explainability.
• Framework Construction: Proposing a set of design recommendations grounded in Design Justice principles and Co-design methodologies.
• Illustrative Use Case: Validating the proposed framework through a hypothetical scenario of an LLM-driven Socially Assistive Robot (SAR) operating in sensitive domains (e.g., schools, care homes) with diverse user groups (disabled, older adults, LGBTQ+, Indigenous communities).

3. Key Contributions

The paper makes four primary contributions to the field of HRI and AI ethics:

A. Identification of Open Challenges

The authors categorize the challenges into three interrelated areas:

1. Pre-training Biases: FM behaviors are inherited from internet-scale corpora, leading to misalignment with situated user needs.
2. Persistence of Generic Explanations: Current systems fail to adapt explanation complexity, modality, or content to the user's mental model.
3. Opacity of Adaptation: There is a lack of transparency regarding how and why the robot adapts its behavior, making the adaptation process itself uncontestable.

B. Ethical Risk Analysis

The paper details three specific ethical dimensions:

• Cognitive: The risk of robots misinterpreting neurodivergent cues (e.g., gaze, prosody) as deficits, leading to inappropriate "accommodations" that reinforce ableism.
• Epistemic: The danger of FMs acting as "knowledge authorities" that erase situated, co-constructed knowledge, particularly for marginalized communities.
• Institutional: The externalization of responsibility, where developers use the "neutrality" of pre-training as a moral alibi, concentrating power in tech providers while harms fall on marginalized users.

C. Four Strategic Recommendations

The authors propose a framework for moving beyond generic explanations:

1. Modality- and User-Aware Strategies: Explanations must be tailored to the user's sensory and cognitive abilities (e.g., visual aids for children, simplified text for those with language disorders, avoiding nonverbal cues that may be culturally ambiguous).
2. Co-Design for Adaptation Criteria: Users and stakeholders must be involved in defining what is adapted, how it is explained, and which ethical aspects are prioritized. This moves design from designer assumptions to user-defined epistemic profiles.
3. Layered Adaptation: A technical strategy balancing personalization with feasibility. Instead of per-user fine-tuning (which is often computationally expensive or data-scarce), the system should use transfer learning from broad user clusters (e.g., age, disability type) and refine with parameter-efficient techniques (e.g., LoRA) as data becomes available.
4. Prioritizing Smaller, Fairer Datasets: Rejecting the "big data" paradigm in favor of curated, co-designed datasets that foreground underrepresented users to ensure explanations are ethically grounded and effective.

D. Illustrative Use Case

The paper demonstrates how these recommendations apply to an LLM-driven SAR. It argues that a "one-size-fits-all" explanation fails in clinical or educational settings. Instead, the robot should:

• Adapt explanation format based on the user's emotional state and cognitive load.
• Use layered adaptation (cluster-based initial models refined by feedback).
• Rely on small, curated datasets co-designed with the specific communities served (e.g., Indigenous or disabled groups) to constrain the LLM's output.

4. Results and Discussion

While the paper does not present quantitative experimental results, its "results" are the conceptual framework and the identified research agenda:

• Rejection of Definitive Technical Solutions: The authors argue that technical fixes alone cannot solve ethical issues; the problem requires a shift in design philosophy.
• Warning on "Illusion of Competence": The paper highlights a critical finding: highly fluent, personalized explanations generated by FMs can create a deceptive illusion of deep understanding, leading to over-trust. Users may infer rich internal reasoning where only pattern matching exists.
• Contextual Necessity: Explanations are not universally beneficial; in high-stress or simple scenarios, they may cause cognitive overload. Adaptive systems must be able to withhold explanations when appropriate.

5. Significance

This paper is significant for several reasons:

• Bridging the Gap: It connects the rapid advancement of Foundation Models in robotics with the often-siloed fields of HRI ethics and user-centered design.
• Redefining Explainability: It moves the definition of explainability from "transparency of code" to "ethical, user-adaptive justification," emphasizing that an explanation is only useful if it is accessible and meaningful to the specific user.
• Design Justice: It explicitly centers marginalized populations (neurodivergent, disabled, Indigenous) in the design of adaptive systems, arguing that without co-design, FMs will inevitably reproduce structural biases.
• Future Research Agenda: It sets a clear path for future work, calling for research into layered adaptation techniques, modality-aware XAI, and ethical constraints on LLM-generated robot behaviors.

In summary, the paper argues that for social robots to be truly ethical and effective in the era of Foundation Models, explainability must be treated not as a post-hoc feature, but as a core, user-adaptive design objective grounded in fairness, co-design, and a critical awareness of the limitations of large-scale models.