# How We Define Harm Impacts Data Annotations: Explaining How Annotators Distinguish Hateful, Offensive, and Toxic Comments

arXiv: https://arxiv.org/abs/2309.15827 · HTML: https://mlsanity.com/paper/2309.15827/

## Key facts

- Cited by 2 papers in the map (data through 14 Oct 2026).
- Placed in Safety alignment, refusal and guardrails, one of 12 directions in LLM safety and security.
- Builds on 10 of its 39 references that are in the map.
- First posted on arXiv 12 Sep 2023.

## Placed in

- LLM safety and security / [Safety alignment, refusal and guardrails](https://mlsanity.com/safety/safety-alignment-refusal-and-guardrails/)

## Builds on (most cited first, 10 of 10)

- [Automated Hate Speech Detection and the Problem of Offensive Language](https://mlsanity.com/paper/1703.04009/) (2017)
- [Understanding Abuse: A Typology of Abusive Language Detection Subtasks](https://mlsanity.com/paper/1705.09899/) (2017)
- [Hate Lingo: A Target-based Linguistic Analysis of Hate Speech in Social Media](https://mlsanity.com/paper/1804.04257/) (2018)
- [Cross-replication Reliability -- An Empirical Approach to Interpreting Inter-rater Reliability](https://mlsanity.com/paper/2106.07393/) (2021)
- [Peer to Peer Hate: Hate Speech Instigators and Their Targets](https://mlsanity.com/paper/1804.04649/) (2018)
- [KOLD: Korean Offensive Language Dataset](https://mlsanity.com/paper/2205.11315/) (2022)
- [Principled Evaluation with Human Labels: One Rater at a Time and Rater Equivalence](https://mlsanity.com/paper/2106.01254/) (2021)
- [Predicting Different Types of Subtle Toxicity in Unhealthy Online Conversations](https://mlsanity.com/paper/2106.03952/) (2021)
- [Conversational Networks for Automatic Online Moderation](https://mlsanity.com/paper/1901.11281/) (2019)
- [Impact Of Content Features For Automatic Online Abuse Detection](https://mlsanity.com/paper/1704.03289/) (2017)

## Built on by (most cited first, 2 of 2)

- [VARM-Bench: Benchmarking Verifiable Structured Reasoning in Chinese Abusive Speech Moderation](https://mlsanity.com/paper/2608.15600/) (2026)
- [Evaluating Criterion-Conditioned Behaviour of Large Language Models in Content Moderation](https://mlsanity.com/paper/2609.03814/) (2026)

## Abstract (arXiv)

Computational social science research has made advances in machine learning and natural language processing that support content moderators in detecting harmful content. These advances often rely on training datasets annotated by crowdworkers for harmful content. In designing instructions for annotation tasks to generate training data for these algorithms, researchers often treat the harm concepts that we train algorithms to detect - 'hateful', 'offensive', 'toxic', 'racist', 'sexist', etc. - as interchangeable. In this work, we studied whether the way that researchers define 'harm' affects annotation outcomes. Using Venn diagrams, information gain comparisons, and content analyses, we reveal that annotators do not use the concepts 'hateful', 'offensive', and 'toxic' interchangeably. We identify that features of harm definitions and annotators' individual characteristics explain much of how annotators use these terms differently. Our results offer empirical evidence discouraging the common practice of using harm concepts interchangeably in content moderation research. Instead, researchers should make specific choices about which harm concepts to analyze based on their research goals. Recognizing that researchers are often resource constrained, we also encourage researchers to provide information to bound their findings when their concepts of interest differ from concepts that off-the-shelf harmful content detection algorithms identify. Finally, we encourage algorithm providers to ensure their instruments can adapt to contextually-specific content detection goals (e.g., soliciting instrument users' feedback).