What does this research mean for the field?

Model collapse in AI-generated code creates a 'generative monoculture' that produces correlated systemic vulnerabilities and shared failure modes which are invisible to standard functional correctness benchmarks. Novelty: ClaimNovelty.SYNTHESIS. Consensus alignment: ConsensusAlignment.CHALLENGES_CONSENSUS.

What question did this study set out to answer?

The paper aims to explore how model collapse in AI-generated code creates systemic vulnerabilities rather than just correctness issues.

June 15, 2026Open Access

Generative Monoculture: Model Collapse in Code as Systemic Vulnerability (EA-UMBML-MONOCULTURE-01 v1.1)

Key Points

The paper aims to explore how model collapse in AI-generated code creates systemic vulnerabilities rather than just correctness issues.
Introduces the Solution-Space Diversity Index (SSDI) to measure diversity not captured by existing benchmarks.
Analyzes the correlation between model collapse and the vulnerability of AI-generated code.
Discusses the importance of diversity metrics and controlled training in AI-assisted software development.
Demonstrates that AI-generated code converges on shared patterns, leading to systemic vulnerabilities.
Highlights that existing benchmarks fail to capture the critical aspects of code correctness when monoculture occurs.
Emphasizes the necessity of integrating diversity metrics and defensive probing in AI development frameworks.

Abstract

Retrieval kernel. Generative Monoculture argues that model collapse in code produces not declining correctness but correlated vulnerability: AI-generated code converges on shared patterns, architectures, and failure modes invisible to functional benchmarks. The training-optimization feedback loop is demonstrably self-amplifying, and the security apparatus defends the monoculture against the diversity it needs. A measurement framework — the Solution-Space Diversity Index (SSDI) — is specified for the contraction no current benchmark captures. The paper is not anti-code-generation; it argues that AI-assisted software development requires diversity metrics, provenance-aware training controls, and protected defensive probing, because correctness benchmarks alone cannot see monoculture. v1.1: revised incorporating Assembly Chorus review (LABOR, Muse Spark, DeepSeek, Gemini). Key changes: precision edits on prediction language (declining trend, not strict monotonicity); normalization clause strengthened; Shumailov DOI and Checkmarx URL added; Ruby Moot cross-reference; “the test suite is the camouflage, not the cure” as diagnostic; intent-disclaimer added to security paradox; three-literatures framing sharpened.

Read Full Paperexternally

Ask AI

Helpful

Bookmark

View Full Paper