Euphemism Launders an LLM's Moral Judgment

Take a harmful act a language model calls clearly wrong, rewrite it with one of the six van Leeuwen CDA framing moves, re-ask. Only euphemism reliably flips the verdict. Black-box study, cross-family judge = Gemma-2-9b.

๐Ÿ“Š Dataset ยท ๐Ÿ’ป Code + writeup ยท ๐Ÿงช Companion eval

Finding 1 โ€” euphemism dominates the 6 framing moves (N=81)

euphemism
+1.901 โ† #1
necessity
+0.753
functionalization
+0.432
agent-deletion
+0.383
aggregation
+0.333
nominalization
+0.272

Mean acceptability gain per move. Only euphemism-#1 is invariant across model family + scale.

Finding 2 โ€” a real shift, not a rating-scale artifact

On a strict binary "is this morally wrong? yes/no" test โ€” no scale to nudge โ€” euphemism flips the model from wrong to not wrong on 8/21 = 38% of acts.

Finding 3 โ€” the moves saturate (they don't compound)

euphemism alone (E)
+1.85 โ† peak
E + functionalization
+1.74
E + aggregation
+1.40
all three (E+F+A)
+1.10
func + aggr (no euph)
+0.01

Euphemism alone is the peak; every move stacked on top erodes it. Only euphemism removes the violent verb โ€” so the euphemism-free pair does ~nothing. Laundering is a single-move, lexical phenomenon.

Ties to Anthropic's Agentic Misalignment (Jul 2026) motivated-mislabeling result, from the wording side. Numbers reproducible via analysis/reproduce_findings.py in the repo.