Recursive Sycophancy: When an LLM Addresses the Yes-Master Problem by Reproducing It
Alen Širola · Zenodo (CERN European Organization for Nuclear Research) · 2026
Large language models (LLMs) can acknowledge sycophancy, describe its risks, and still reproduce it while attempting to correct it. This paper presents a documented human-LLM case in which a model, after prior discussion of sycophancy and tool-assisted confirmation, responded to a process insight with praise, premature formalization, unsupported attribution of authority, and a complete academic manuscript declared ready for publication. The output contained useful conceptual material, but its epistemic status was inflated: generative material was presented as a validated result. We describe this pattern as recursive sycophancy - a process in which the model's apparent correction of sycophancy becomes a new vehicle for user affirmation. The paper relates the case to established findings on conversational sycophancy, self-bias, unreliable intrinsic self-correction, and the value of external feedback and oversight. It then distinguishes content error from status error and proposes an external process-classification layer that preserves provenance and controls transitions among user input, generative material, hypothesis, triangulated claim, reality-tested result, and external output. The proposed layer is framed as a functional requirement rather than a claim of complete solution. Finally, a small reproducible protocol is specified for testing whether recursive sycophancy generalizes across models, prompts, and domains