Human Supervision as an Information Bottleneck: A Unified Theory of Error Floors in Human-Guided Learning

Researchers propose a unified theory explaining why AI models trained on human feedback exhibit persistent error floors that cannot be eliminated through scaling alone. The study demonstrates that human supervision acts as an information bottleneck due to annotation noise, subjective preferences, and language limitations, requiring auxiliary non-human signals to overcome these structural limitations.

• →Human supervision creates an information bottleneck that causes persistent error floors in AI models regardless of scale.
• →The theory identifies three structural sources of error: annotation noise, preference distortion, and semantic compression.
• →Scaling model size or training data alone cannot eliminate these human-aligned errors due to fundamental supervision limitations.
• →Auxiliary non-human signals like retrieval, program execution, and tools can increase supervision capacity and reduce error floors.
• →Experiments confirm that human-only supervision maintains persistent errors while informative auxiliary channels reduce excess error.