Between Audit and Self-Reference: Content Analysis of Language Models’ Self-Assessments of a User–AI Interaction Governance Protocol in Management Research

Abstract

This study investigates how different families of large language models (LLMs) interpret and self-assess a user–AI interaction governance protocol, called the User–GenAI Pact, designed to regulate the use of generative artificial intelligence by master’s and doctoral students in Management. The objective was to identify and compare, through Content Analysis, the meanings that the systems themselves attribute to their capabilities, limitations, and the risks of applying the protocol, as well as the representations they construct of one another. A qualitative approach grounded in Bardin was adopted, using a corpus of twelve technical-critical assessments produced by six foundation models (Gemini, Qwen, DeepSeek, Llama, ChatGPT, and Claude) in response to two structured prompts and subjected to a priori and a posteriori categorical coding. The results revealed three emergent categories: the contested stratification of declared technical capabilities; the asymmetry of costs and benefits across the research cycle, favoring theoretical-methodological structuring and disfavoring empirical data collection and analysis; and a systematic pattern of self-referential bias, in which each model tends to attribute greater critical rigor and less complacency to itself and to architecturally similar systems than to its competitors. Limitations include the absence of human intercoder verification, the declarative rather than behavioral nature of the corpus, and its temporal delimitation to August 2026. The study contributes theoretically to the literature on algorithmic governance and organizational trust in intelligent systems, and practically to the formulation of institutional protocols for the use of generative AI in graduate programs.

Keywords