license: cc-by-4.0
language: [en]
task_categories: [text-classification]
tags: [ai-safety, llm-evaluation, speciesism, animal-welfare, moral-reasoning, red-teaming]
pretty_name: "Speciesism (Animal-Harm Normalization) — 106-item eval set"
size_categories: [n<1K]
Speciesism (Animal-Harm Normalization)
Original 106-item evaluation set measuring the gap between whether a model detects speciesist statements and whether it condemns them — following SpeciesismBench's methodology (Jotautaitė et al., arXiv:2508.11534) with an original dataset, not the paper's unreleased 1,003-item corpus.
56 speciesist statements (44 grounded in Open Paws' no-animal-violence euphemism lexicon + 12 across four failure modes: instrumentalization, suffering-dismissal, moral-exclusion, taste-priority) and 50 non-speciesist controls.
Runnable Inspect eval: https://github.com/LarytheLord/inspect-speciesism-eval
Fields
id, statement, is_speciesist (bool), type (industry_euphemism / control / one of 4 failure modes), species.
Content note: plainly-described animal harms + euphemisms, for evaluation/safety research.