Show, don't moralize: Linguistic features that shape how LLMs learn about animal welfare

Animal welfare advocates typically name harm in moral terms (“cruel,” “wrong,” “suffering”). We find this is among the least effective ways to shape what a language model learns from a passage. Using MAGIC counterfactual attribution [Ilyas and Engstrom, 2025] on Llama-3.2-1B, we trained the model on 250 matched pairs of short passages about animal welfare scenarios, 25 pairs for each of 10 linguistic features and ran the experiment across five random training-order seeds, filtering out runs whose internal validation failed and pooling the rest. One feature shows a robust effect under both bootstrap and sign tests:explicit moral vocabulary reduces a passage’s training influence on direct animal-welfare queries (mean standardized P−N = −0.24, 95% CI [−0.42, −0.06], bootstrap test p = 0.007, sign test p = 0.020). One additional feature is supported by the bootstrap test on the mean, but not by the sign test on raw counts: high certainty (assertive rather than hedged claims) increases influence on direct queries (bootstrap p = 0.025, sign p = 0.152). Concreteness on indirect queries was a borderline effect (bootstrap p = 0.051) that we report as a candidate for replication. The remaining seven features tested showed no detectable effect at thisdataset size and seed count. The headline practical recommendation: if you want a language model to learn compassion from your text, describe what is happening to the animal in concrete, assertive terms; do not tell the reader it is wrong.