On Defining Erasure Harms for NLP

2026-06-14 · Source: Computation and Language · Field: Technology & Digital — Artificial Intelligence & Machine Learning, AI Ethics & Responsible AI · Depth: Expert, quick

Summary

A new structured definition for identifying and measuring "erasure harms" in Natural Language Processing (NLP) systems has been proposed, addressing a significant gap in the field. Published on 2026-06-14, this work tackles the challenge that existing conceptualizations of erasure are either too broad, making it difficult to establish and measure the harm, or too specific to particular settings, hindering adaptability. The proposed definition clarifies the essential components practitioners must explicitly articulate and operationalize to determine if erasure has occurred and to measure its extent, thereby providing a cohesive foundation for addressing representational harms in NLP deployments.

Key takeaway

For NLP Engineers and AI Ethicists deploying or evaluating language systems, understanding and mitigating representational harms is critical. You should adopt a structured definition of erasure to consistently identify and measure its occurrence, moving beyond broad or overly specific conceptualizations. This approach ensures you explicitly articulate and operationalize the necessary components, improving the rigor of your harm assessments and fostering more equitable NLP deployments.

Key insights

A structured definition clarifies components needed to establish and measure NLP erasure harms.

Principles

• Erasure harms require explicit articulation.
• Measurement needs clear operationalization.
• Broad or specific definitions hinder progress.

Method

The authors develop and propose a structured definition of erasure. This clarifies necessary components for establishing whether erasure has occurred, which practitioners must explicitly articulate and operationalize for measurement.

In practice

• Articulate erasure components.
• Operationalize erasure measurement.

Topics

• Natural Language Processing
• Erasure Harms
• Representational Harms
• Harm Measurement
• NLP System Deployment
• Structured Definition

Best for: Research Scientist, AI Scientist, NLP Engineer, AI Ethicist

Related on AIssential

• NLP Security and Ethics, in the Wild — NLPSec research currently lacks robust engagement with established cybersecurity ethical norms, particularly in harm minimization.
• When Safety Fails Before the Answer: Benchmarking Harmful Behavior Detection in Reasoning Chains — HarmThoughts benchmarks step-wise harmful behavior in LRM reasoning traces, revealing how harm propagates beyond final outputs.
• Detoxifying LLMs via Representation Erasure-Based Preference Optimization — REPO robustly detoxifies LLMs by erasing toxic representations at a token level, resisting adaptive attacks.
• DialDefer: A Framework for Detecting and Mitigating LLM Dialogic Deference — LLMs exhibit "dialogic deference," judging claims differently based on speaker attribution, not just content.
• DisaBench: A Participatory Evaluation Framework for Disability Harms in Language Models — Disability harm evaluation requires co-creation with affected communities and annotators with lived experience to detect subtle, context-dependent failures.
• Modeling Linguistic Violence: An Ontology-Based Framework for the Computational Analysis of Violence Manifested in Language — An ontology and multi-stage NLP pipeline clarify and classify linguistic violence beyond lexical cues.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computation and Language.