February 26, 2024Open Access

Algorithmic Arbitrariness in Content Moderation

View Full Paper

Ask AI

Bookmark

Authors

JGJuan Felipe GomezHarvard University CMCaio Vieira Machado LPLucas Monteiro PaesGoogle (United States)

Discussion

Member takes

Implication

Experimental analysis reveals arbitrary toxicity classifications across language models, highlighting severe risks to freedom of expression and procedural justice.

Key Points

Identical text receives conflicting toxic classifications across models due to minor random seed selection, leading to disparate impacts across marginalized social groups.
Experimental analysis evaluates state-of-the-art language models against human classifications to quantify how parameter initialization drives inconsistent content moderation.
Highlights urgent calls for transparency in algorithmic content moderation, warning that unchecked model multiplicity threatens human rights and global regulatory compliance.

Cite This Study

Gomez et al. (2024) studied this question.

synapsesocial.com/papers/68e779ebb6db6435876eeaab https://doi.org/10.48550/arxiv.2402.16979

View Full Paper

Ask AI

Bookmark

Implication

Experimental analysis reveals arbitrary toxicity classifications across language models, highlighting severe risks to freedom of expression and procedural justice.

Algorithmic Arbitrariness in Content Moderation

Discussion

Member takes

Implication

Key Points

Cite This Study

Algorithmic Arbitrariness in Content Moderation

Discussion

Member takes

Implication

Key Points

Cite This Study

Also Consider

Also Consider