Executive Summary
When a communication channel is lossy and the receiver must produce structured output, the receiver is forced to generate: to fill the gap between what was received and what must be produced. This paper formalizes the generative lossy channel, a channel where compression, strategic misalignment, and conformity pressure combine to force reconstruction that systematically diverges from the source.
The mechanism is substrate-independent. It appears in organizational hierarchies, where reports compress reality and institutional norms constrain output, and in AI systems, where training compresses human values and reward signals constrain behavior. The paper argues the claim is not analogical: both domains share the structural cause. Compression creates information gaps, and selection operating in those gaps determines whether the divergence is functional or pathological.
Five sufficient conditions (C1–C5) are proved to guarantee net-beneficial noise in any two-level system with a generative lossy channel. Computational validation over 500 Monte Carlo configurations yields zero counterexamples to sufficiency (75 of 75 configurations where all conditions hold), across six mechanistically distinct domains.
Key Contributions and Methodology
The formal foundation draws on three bodies of theory: Crawford and Sobel (strategic communication under preference divergence is endogenously lossy), Blau and Michaeli (lossy compression under a perceptual quality constraint forces the decoder to generate), and Atlan (noise at one level of organization can produce information at a higher level).
Theorem 1 establishes an inverted-U relationship between noise amplitude and system performance under conditions C1–C5 (suboptimality, nonlinear integration, accessibility, asymmetric weighting, and graceful degradation), with three independent proof paths: threshold stochastic resonance, Jensen-gap rectification, and opposing-monotone mechanisms.
Cross-domain evidence maps the conditions onto published work: Vaughan's stages of normalized deviance in the Challenger disaster, Berliner's and Pressing's analyses of jazz improvisation, and sycophancy in language models as measured by Perez et al. and Sharma et al. The paper's scope statement is explicit: these instances verify qualitative satisfaction of C1–C5, and full formal verification within each domain is future work.
Key Findings
- Five sufficient conditions (Theorem 1): C1–C5 jointly guarantee net-beneficial noise, with an inverted-U relationship between noise amplitude and performance
- Computational validation: zero counterexamples to sufficiency across 500 Monte Carlo configurations
- Dual valence (Proposition 2.10): the generative residual produces drift toward internally fit outputs under convergent selection and creative emergence under divergent selection
- Inseparability (Corollary 2.11): a system cannot eliminate dysfunction potential without also eliminating creative potential
- Compression ratchet (Proposition 2.12): once compression-driven drift crosses a self-reinforcing threshold, recovery requires external intervention
- Substrate independence: the same mechanism operates in organizational hierarchies and in AI systems
Key References
Strategic Information Transmission. Econometrica.
Thinking in Jazz: The Infinite Art of Improvisation. University of Chicago Press.
Improvisation: Methods and Models. In Generative Processes in Music. Oxford University Press.
The Challenger Launch Decision: Risky Technology, Culture, and Deviance at NASA. University of Chicago Press.
Towards Understanding Sycophancy in Language Models.