Evolving Ethical Communicability in Generative AI Interfaces: A Longitudinal Comparative Analysis of Ethical Principles Across System Versions
Background: Generative Artificial Intelligence (AI) systems have become central interactive artifacts in activities such as information seeking, creative work, education, and decision support. As these systems increasingly mediate human-computer interaction, concerns regarding transparency, accountability, autonomy, fairness, and social impact have intensified. Although ethical principles for AI are well established in regulatory documents and conceptual frameworks, prior research has shown that their communication through interface elements is often inconsistent and opaque. Furthermore, the rapid evolution of generative AI systems raises challenges for understanding ethical communicability as an evolving, rather than static, property of interactive systems. Purpose: This paper investigates how ethical communicability in generative AI systems evolves over time by conducting a longitudinal comparative analysis of successive system versions. Extending prior work presented at IHC 2025, the study compares recent and earlier versions of three widely used generative AI systems, examining changes in how ethical principles are communicated at the interface level. The goal is to identify patterns of continuity, improvement, and regression, contributing empirical and conceptual insights to the design and evaluation of generative AI based interactive systems. Methods: The study adopts the Semiotic Inspection Method (SIM) in a scientific context, supported by a semiotics-based epistemic tool for ethical reflection, to analyze how ethical principles are communicated through interface signs. The analysis is guided by the AI4People ethical principles of Beneficence, Non-Maleficence, Autonomy, Justice, and Explicability. Using the same inspection protocol, ethical scenarios, and analytical categories employed in the original study, we re-inspected the current versions of the three systems. The results from both inspection phases were systematically compared using a structured analytical framework, enabling the identification of additions, refinements, and inconsistencies in ethical communication across versions. Researcher triangulation was applied to ensure analytical rigor and consistency. Results: The longitudinal comparison reveals that ethical communicability in generative AI systems is dynamic and uneven. Some systems show incremental improvements, particularly in interface elements related to explicability and risk mitigation, such as clearer disclaimers, refined feedback mechanisms, and expanded data control options. However, other ethical principles - especially beneficence and justice - remain weakly communicated or largely implicit across versions. The analysis also exposes persistent inconsistencies between metalinguistic commitments expressed in policies and the ethical cues available during interaction. These findings suggest that system updates do not necessarily lead to systematic or holistic improvements in ethical communication, but rather to fragmented and principle-specific changes. Conclusion: By shifting from a static to a longitudinal perspective, this study demonstrates the value of analyzing ethical communicability as an evolving property of generative AI interfaces. The findings highlight that updates to generative AI systems can both strengthen and undermine the communication of ethical principles, underscoring the need for continuous and systematic ethical evaluation in the design of interactive systems. The paper contributes empirical evidence on how ethical communication changes over time, methodological insights into the use of SIM for longitudinal analysis, and design implications for fostering more transparent, accountable, and human-centered generative AI systems. This extended investigation reinforces the relevance of ethical communicability as a core concern for HCI research and practice.