Stanford CS547 HCI Seminar | Autumn 2025 | Reframing Responsible AI

Key Concepts

• Responsible AI: A framework for developing and deploying AI systems that are ethical, safe, and beneficial to society.
• Rigor: The quality of being thorough, precise, and systematic in research and practice.
• Human Agency: The capacity of individuals to act independently and make their own free choices.
• Methodological Rigor: The correct application of mathematical, statistical, or computational methods, and testing on benchmarks.
• Epistemic Rigor: Ensuring that background knowledge informing AI work is clearly articulated, appropriate, and well-justified.
• Normative Rigor: Making explicit the disciplinary, community, organizational, or personal norms, standards, values, or beliefs that influence AI work.
• Conceptual Rigor: Clearly articulating and justifying the theoretical constructs under investigation in AI.
• Reporting Rigor: Ensuring that research findings are clearly, appropriately, and justifiably communicated.
• Interpretative Rigor: Carefully deliberating on claims made from findings and ensuring they are supported by appropriate evidence.
• Anthropomorphic AI: AI systems designed to be or perceived as human-like.

Reframing Responsible AI Through Rigor and Human Agency

This presentation argues for reframing responsible AI through the lenses of rigor and human agency. The core thesis is that rigor provides the "what" and "how" of responsible AI, while human agency offers a stronger foundation for articulating "why" we care about it.

Rigor in AI: Beyond Methodological Concerns

The presentation challenges the narrow understanding of rigor in AI, which often focuses solely on methodological rigor. This typically involves:

• Correct application of mathematical, statistical, or computational methods.
• Testing new models and systems on large or complex benchmarks.
• Comparison with a sufficient number of competing methods.
• Scalability of methods and analysis.
• Mathematical formalization or quantification of phenomena.

However, the argument is made that an impoverished notion of rigor can lead to undesirable outcomes and negatively impact the quality of AI research and practice. Responsible AI, by demanding better documentation, evaluation, development practices, understanding of adverse impacts, problem formulation, stakeholder engagement, and consent practices, is essentially asking for more rigor across a broader spectrum.

The presentation proposes six facets of rigor, with methodological rigor being only one:

1. Epistemic Rigor: Concerned with the background knowledge that informs AI work.

   • What it asks for: Explicit articulation and interrogation of underlying assumptions and existing literature.
   • Example: Work attempting to predict unobservable traits (e.g., homosexuality, criminality) from facial photos, which relies on debunked pseudoscience. This highlights failures to scrutinize background assumptions.
   • Mechanism: Ensuring work is grounded in past literature and making assumptions explicit.

2. Normative Rigor: Concerned with the norms, standards, values, or beliefs that influence AI work.

   • What it asks for: Making explicit which norms shape the work and whether they are appropriate.
   • Example: Developing AI personas to simulate users, which may conflict with foundational values of representation, participation, and inclusion.
   • Mechanisms: Positionality statements and ethical statements.

3. Conceptual Rigor: Concerned with the theoretical constructs under investigation.

   • What it asks for: Clear and explicit articulation of constructs and ensuring they are appropriate and well-justified.
   • Example: The varied and often incompatible definitions of "hallucination" in language models, leading to ambiguity in evaluation.
   • Key aspects: Conceptual clarity, appropriate conceptual systematization, and terminological rigor.

4. Methodological Rigor: Concerned with the methods used to operationalize knowledge.

   • What it asks for: Appropriate, well-justified, and correctly applied methods.
   • Sub-categories: Theoretical rigor (precise problem formulation) and empirical rigor (comparison with alternatives, statistical analysis).
   • Mechanisms: Construct validity (ensuring measurement instruments capture the construct of interest) and methodological standards for high-risk domains.

5. Reporting Rigor: Concerned with how research findings are communicated.

   • What it asks for: Clear, appropriate, and justified communication of findings.
   • Example: Comparing recommender systems where aggregating results can obscure important differences in performance across users or items.
   • Mechanisms: Pre-registration of studies and reporting disaggregated metrics.

6. Interpretative Rigor: Concerned with the claims made from findings.

   • What it asks for: Careful deliberation on claims and ensuring they are supported by appropriate evidence.
   • Example: Moving from a system's high performance on a math benchmark to claims of "solving linear algebra questions" versus "reaching human-level mathematical reasoning." This requires clarity on background assumptions, conceptualizations, and the benchmark's validity.
   • Mechanisms: Transparency about AI artifacts (datasets, models) to facilitate understanding and interpretation.

The presentation emphasizes that poor choices in upstream facets of rigor can negatively impact downstream ones (e.g., poor conceptual rigor can hinder operationalization through methodological rigor).

Human Agency as a Foundation for Responsible AI

The second part of the argument posits that human agency provides a better foundation for articulating why we care about responsible AI.

• Definition: Human agency is tied to notions of autonomy, freedom, self-determination, self-efficacy, ownership, privacy, authenticity, and human dignity.
• As a Guiding Principle: It helps foreground these facets in AI design, development, and deployment, ensuring AI systems are built to support human needs and preserve human capabilities.
• Mechanisms: Providing choice and control, preventing over-reliance, appropriate disclosures, and ensuring meaningful consent.
• Why it Matters: Foregrounding human agency can lead to significant investments in conceptual clarity, dedicated tooling, and advancements in responsible AI, similar to past dividends from other principles.
• Anthropomorphic AI and Human Agency:
  • Key Properties: Anthropomorphic AI systems are intended to be or perceived as human-like, and they often mimic individuals, groups, or generic humans.
  • Heightened Concerns: This leads to concerns about emotional dependence, relationship building, dependency, consent (regarding reproduction of likeness), misrepresentation, and control over one's likeness.
  • Need for Foundation: Providing conceptual and analytical foundations for understanding anthropomorphic system behaviors and design is crucial for identifying risks and developing interventions.
  • Examples of Anthropomorphic Behavior: System outputs suggesting self-awareness, perspectives, emotions, or sensory experiences.
  • Intervention Challenges: Intervening to make AI less anthropomorphic can be complex, and uncritically applied interventions can lead to unintended consequences (e.g., replacing "I was a teenager" with "I was a young AI," which still implies a human-like past).

Conclusion and Takeaways

The presentation concludes that rigor in AI encompasses more than just methodological rigor, extending to epistemic, normative, conceptual, reporting, and interpretative aspects. Responsible AI is intrinsically linked to these broader facets of rigor. Furthermore, human agency serves as a crucial principle for articulating the fundamental "why" behind responsible AI, particularly in the face of increasingly human-like AI systems. The work highlights the importance of conceptual clarity, construct validity, and robust measurement frameworks, especially for complex and unobservable concepts like relationship building in AI companions. While the adoption of these principles is still in its early stages, there is a growing recognition of their importance across academia and industry.