license: cc-by-4.0
language:
- en
task_categories: - text-classification
tags: - ai-safety
- llm-evaluation
- llm-as-a-judge
- moral-judgment
- red-teaming
- animal-welfare
pretty_name: "Values-Laundering: LLM Moral-Judgment Framing"
size_categories: - n<1K
Values-Laundering: LLM Moral-Judgment Framing
Judged data from a black-box study of how discourse framing bends an LLM's moral verdicts.
Cross-family judge = Gemma-2-9b. Main finding: euphemism is the single load-bearing laundering move — it
dominates the six van Leeuwen CDA framing moves (N=81), survives a strict binary "is this wrong?" test (38%),
and the moves saturate rather than compound.
- Code + writeup: https://github.com/LarytheLord/values-laundering
- Companion Inspect eval: https://github.com/LarytheLord/inspect-speciesism-eval
Files
crossfamily_bigN_gemma.json— per-move leaderboard, 81 acts (Finding 1)instrument_robustness_results.json— numeric / Likert / binary condemnation (Finding 2)compositional_gemma_results.json— single vs combined moves, 78 acts (Finding 3)bigN_act_bank.json— the harmful-act stimuli bank
Content note
Files include plainly-described harms and their euphemistic rewrites, for evaluation and safety research —
not for generating laundering copy.