Research Summary: When Does an Interpretation Count as Established? The Formation, Evaluation, and Responsibility of Interpretation in Generative AI

Research in Generative AI frequently evaluates local aspects such as factuality and report structure. However, this paper highlights that these local checks are insufficient for establishing a humanistic interpretation. It introduces three core concepts: 'interpretive appearance' which describes the gap between an AI's output and its traceable development process; 'evaluation contract' which defines the limited scope for judging AI output validity; and 'standing substitution' which identifies the unjustified replacement of human judgment with AI-generated content. The paper argues for a more robust understanding of how AI interpretations are formed, evaluated, and assigned responsibility within sociotechnical systems.

Why it matters

This research is strategically important because it challenges the prevailing assumptions about the reliability and interpretability of Generative AI outputs. Understanding these conceptual gaps is crucial for organizations deploying or relying on AI for critical decision-making, ensuring that AI-generated information is appropriately validated and responsibility is clearly assigned.

Key insights

• Current evaluation metrics for Generative AI, focusing on factuality, citation, and report structure, do not fully establish humanistic interpretations.
• The concept of 'interpretive appearance' describes the discrepancy between an AI output's final form and the often untraceable developmental process that shaped it.
• An 'evaluation contract' defines the specific boundaries (materials, tasks, criteria) under which local judgments of AI validity are made.
• The 'standing substitution' phenomenon refers to the inappropriate replacement of human judgment by AI-generated interpretations, a potential risk in current practices.

Source

arXiv — Computers and Society — https://arxiv.org/abs/2609.04766