Co-Creative Adaptation in Long-Term Naturalistic Dialogue with ChatGPT: Distinguishing User Adaptation from Sycophancy
Year: 2026
Type: preprint
Abstract: Conversational AI systems adapt their responses to users, which supports personalization and practical assistance. At the same time, such adaptation has been problematized as sycophancy: excessive agreement with or deference to a user's views. Treating every response that appears to adapt to a user as sycophancy, however, makes it difficult to capture interactions in which an AI reorganizes material around the user's aims and interests while leaving differences in judgment intact, and the human continues to evaluate, revise, and ask follow-up questions. This study is an exploratory, descriptive qualitative observational study of naturalistic dialogue sustained by a single user in the course of real-world ChatGPT use. Five cases were purposively selected from research, academic writing, text editing, creative work, and everyday interaction to provide detailed descriptions of multi-turn sequences relevant to the research question. The analysis traced human input, AI responses, human evaluation, rejection, revision, selection, or a subsequent question, and the AI response that followed. These cases are not a sample for estimating frequency or representativeness. In this paper, co-creative adaptation is used as a descriptive umbrella label for cases in which user adaptation, broadly construed, appears within interactions treated as Human-AI co-creation. Across the cases, with varying strength, the analysis found that (1) human evaluative agency remained present in the interaction; (2) AI responses took the form of organization or reconstruction oriented to the user's aims and interests; and (3) particularly in Cases C and D, differences in judgment or conclusion remained, dialogue continued, and those differences became material for further questions, investigation, or revision. As a secondary observation, the user also assigned value to the time spent interacting beyond the production of an artifact, after which the framing in the subsequent AI response changed. Conceptual and interactional comparison of these observations with sycophancy research suggests that user adaptation in the broad sense may not be reducible to harmful sycophancy. This study describes a limited set of naturalistic ChatGPT use cases and makes no claims about causality, internal mechanisms, frequency, or generalization to other conversational AI systems.
Author: Hinata, Lucy
Language: English
Topic: Speech and dialogue systems
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7210271004
Cites:
Cited by:
Related to: