The Generative Lossy Channel: Five Sufficient Conditions for Net-Beneficial Noise

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. It draws on three results: 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).

Under convergent selection, where fit with existing frames is rewarded, the divergence produces organizational self-deception, compliance theater, and AI sycophancy. Under divergent selection, where functional novelty is rewarded, the same divergence produces creative emergence. The mechanism is identical; the selection environment determines the valence.

The paper argues that substrate independence is not an analogy: organizational hierarchies (where reports compress reality and norms constrain output) and AI systems (where training compresses human values and reward signals constrain behavior) share the structural cause. Five sufficient conditions for net-beneficial noise are proved and validated computationally, and a ratchet proposition (Proposition 2.12) shows that once compression-driven drift crosses a self-reinforcing threshold, recovery requires external intervention.

Key Contributions and Methodology

Theorem 1 (stated formally as Theorem 2.5) shows that five conditions (suboptimality C1, nonlinear integration C2, accessibility C3, asymmetric weighting C4, and graceful degradation C5) are jointly sufficient for net-beneficial noise in any two-level system with a generative lossy channel, with an inverted-U relationship between noise amplitude and performance. Three independent proof paths cover threshold stochastic resonance, Jensen-gap rectification, and opposing-monotone mechanisms.

The compression ratchet (Proposition 2.12) formalizes why convergent selection traps a system. When the acceptability distribution is updated from prior outputs, selection rewards conformity with it, and no external correction enters the loop, the system reaches a self-reinforcing equilibrium farther from reality than where it started; past a threshold the drift increases monotonically, and recovery requires signals from outside the selection environment.

Validation proceeds on two fronts. Computationally, six mechanistically distinct domains are tested, and 500 random Monte Carlo configurations yield zero counterexamples to sufficiency (75 of 75 configurations where all conditions hold showed benefit). Cross-domain evidence maps C1-C5 onto published work: Vaughan's 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 the conditions, and full formal verification within each domain is future work.

Key Findings

• Endogenous lossiness: Any channel with preference divergence between sender and receiver is lossy; lossiness is structural, not a design flaw
• Generative residual: Reconstruction under a conformity constraint produces structured divergence from the source, not random noise
• Dual valence (Proposition 2.10): The same residual produces pathological drift 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 drift crosses a self-reinforcing threshold, recovery requires external intervention
• Five sufficient conditions: Zero counterexamples to sufficiency across 500 Monte Carlo configurations; the conditions are sufficient but not necessary

Key References