Selective Control under Noisy Perception: Governance Failures Hidden by Aggregate Metrics in Modular Networks

Hugging Face Daily Papers Papers

Summary

This paper demonstrates that content moderation systems can cause disproportionate harm to bridge users connecting separate communities, even when aggregate accuracy metrics appear satisfactory, with governance loss increasing under false-positive-heavy conditions.

A content-moderation system can score well on every standard accuracy metric and still cause real harm, if its mistakes fall on the few users who connect otherwise separate communities. We show this in an agent-based model where N=240 learning agents on a community-structured network each post harmless, productive, or dangerous content, and a regulator removes or penalizes whatever a noisy classifier flags. Overall usefulness barely moves as the noise changes (one-way ANOVA, p=0.96): by aggregate measures, nothing looks wrong. The damage instead concentrates on these bridge users, whose useful posts are wrongly suppressed and whose dangerous posts are wrongly spared. A governance loss (L_gov) that prices these two mistakes separately from the cost of enforcement more than doubles under false-positive-heavy noise. Aggregate accuracy hides who is harmed, and the cheap quantity to audit is how many connections a user has (degree), a near-perfect proxy for the betweenness that defines a bridge (r=0.96).
Original Article
View Cached Full Text

Cached at: 06/16/26, 11:32 AM

Paper page - Selective Control under Noisy Perception: Governance Failures Hidden by Aggregate Metrics in Modular Networks

Source: https://huggingface.co/papers/2606.14819

Abstract

Content moderation systems can cause disproportionate harm to bridge users connecting separate communities, even when overall accuracy metrics appear satisfactory, with governance loss increasing significantly under false-positive-heavy conditions.

A content-moderation system can score well on every standard accuracy metric and still cause real harm, if its mistakes fall on the few users who connect otherwise separate communities. We show this in anagent-based modelwhere N=240 learning agents on acommunity-structured networkeach post harmless, productive, or dangerous content, and aregulatorremoves or penalizes whatever anoisy classifierflags. Overall usefulness barely moves as the noise changes (one-way ANOVA, p=0.96): by aggregate measures, nothing looks wrong. The damage instead concentrates on thesebridge users, whose useful posts are wrongly suppressed and whose dangerous posts are wrongly spared. Agovernance loss(L_gov) that prices these two mistakes separately from the cost of enforcement more than doubles underfalse-positive-heavy noise. Aggregate accuracy hides who is harmed, and the cheap quantity to audit is how many connections a user has (degree), a near-perfect proxy for thebetweennessthat defines a bridge (r=0.96).

View arXiv pageView PDFGitHub0Add to collection

Get this paper in your agent:

hf papers read 2606\.14819

Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash

Models citing this paper0

No model linking this paper

Cite arxiv.org/abs/2606.14819 in a model README.md to link it from this page.

Datasets citing this paper0

No dataset linking this paper

Cite arxiv.org/abs/2606.14819 in a dataset README.md to link it from this page.

Spaces citing this paper0

No Space linking this paper

Cite arxiv.org/abs/2606.14819 in a Space README.md to link it from this page.

Collections including this paper0

No Collection including this paper

Add this paper to acollectionto link it from this page.

Similar Articles

Social Pressure Breaks Majority Voting in LLM Safety Panels

arXiv cs.CL

This arXiv paper studies how shared social cues from simulated peers break the majority-voting protection in LLM safety panels. It shows that when all reviewers receive the same incorrect 'unsafe' label, panel false-alarm rates jump to 100%, revealing a failure mode and offering a pre-deployment diagnostic.

Toxicity on Social Media – The Noisy Room

Hacker News Top

A Stanford study analyzing billions of social media posts reveals that only ~3% of users generate severely toxic content, but engagement-driven algorithms disproportionately amplify this minority, distorting public perception and driving self-censorship among the majority.