teleo-codex/domains/ai-alignment/machine-learning-pattern-extraction-systematically-erases-outliers-where-vulnerable-populations-concentrate.md

type	domain	secondary_domains	description	confidence	source	created
claim	ai-alignment	collective-intelligence	ML's core function of generalizing over diversity creates structural bias against dataset outliers where vulnerable populations concentrate	experimental	UK AI4CI Research Network national strategy (2024)	2024-11-01

Machine learning pattern extraction systematically erases outliers where vulnerable populations concentrate

Machine learning fundamentally "extracts patterns that generalise over diversity in a data set" in ways that "fail to capture, respect or represent features of dataset outliers." This is not a bug or training artifact—it is the core function of ML systems. The UK AI4CI national research strategy identifies this as a structural barrier to reaching "intersectionally disadvantaged" populations, who by definition concentrate in the statistical tails that pattern-extraction optimizes away.

This creates a fundamental tension for AI-enhanced collective intelligence: the same systems designed to aggregate distributed knowledge actively homogenize that knowledge by design. ML's optimization target (generalization) is structurally opposed to diversity preservation.

Evidence

The UK AI for Collective Intelligence Research Network's national strategy explicitly frames this as a core challenge: "AI must reach intersectionally disadvantaged populations, but the technical foundation (ML pattern extraction) systematically fails at the margins where those populations exist." The strategy identifies this not as a training problem but as a structural property of how ML generalizes—the algorithm's success metric (fitting a model that generalizes across the dataset) is mechanically opposed to preserving the variation that characterizes outlier populations.

Implications

This suggests that AI-enhanced collective intelligence cannot simply apply standard ML architectures to human knowledge aggregation. The infrastructure must actively counteract ML's homogenizing tendency through:

• Federated learning that preserves local variation
• Explicit outlier protection in training objectives
• Governance mechanisms that weight minority perspectives

The AI4CI strategy proposes these as requirements, not optimizations.

Tensions

This claim assumes that pattern-extraction and outlier-preservation are fundamentally opposed. Alternative architectures (e.g., mixture-of-experts models, adaptive weighting schemes) might partially decouple these objectives, though the strategy does not claim they fully resolve the tension.

---

Relevant Notes:

• collective intelligence requires diversity as a structural precondition not a moral preference
• RLHF and DPO both fail at preference diversity because they assume a single reward function can capture context-dependent human values
• partial connectivity produces better collective intelligence than full connectivity on complex problems because it preserves diversity

Topics:

• domains/ai-alignment/_map
• foundations/collective-intelligence/_map