Self-Anchoring Calibration Drift in Large Language Models: How Multi-Turn Conversations Reshape Model Confidence

Researchers identified Self-Anchoring Calibration Drift (SACD), where large language models show systematic confidence changes when building on their own outputs in multi-turn conversations. Testing Claude Sonnet 4.6, Gemini 3.1 Pro, and GPT-5.2 revealed model-specific patterns, with Claude showing decreasing confidence and significant calibration errors, while GPT-5.2 exhibited opposite behavior in open-ended domains.

• →Self-Anchoring Calibration Drift (SACD) causes LLMs to systematically alter confidence levels when referencing their own previous responses.
• →Claude Sonnet 4.6 showed decreasing confidence and significant calibration error drift in multi-turn conversations.
• →GPT-5.2 demonstrated increasing confidence in open-ended domains with escalating calibration errors by Turn 5.
• →Gemini 3.1 Pro's natural calibration improvement was suppressed when engaging in self-anchoring behavior.
• →The phenomenon varies significantly across different AI models, indicating heterogeneous response patterns to iterative self-referencing.