Llama 3.1's Vicarious Trick Beats Human Agreement on Hate Speech Moderation
A new study tests whether LLMs prompted with demographic identities can replicate how different human groups disagree on hate speech — and finds the answer is mostly no, with one surprising exception.
Takeaways
- Core finding: No LLM consistently replicates all three dimensions of human demographic disagreement in hate speech annotation across five demographic axes.
- Vicarious prompting wins: Asking Llama 3.1 to predict how another group would react outperforms asking it to act as that group, matching or exceeding human κ on 4 of 5 axes.
- Model matters enormously: Nemo best approximates human self-alignment patterns; Llama 3.1 self-alignment scores zero on Religion; Qwen 3 underperforms throughout.
- Richer prompts can backfire: Biographical persona descriptions cause models to over-commit to their identity, making vicarious perspective-taking worse, not better.
- Ideology is unsolved: Liberal vs. Conservative disagreement patterns remain the hardest to replicate for all models tested.
- Code available: Zero-shot experiment code and prompts are at the project repository; paper at arXiv:2606.06266.
Content moderation at scale has a dirty secret: what counts as hate speech depends enormously on who is reading it. A slur directed at one community lands very differently depending on whether the reader is a member of that group or not. Collecting diverse human annotations to capture this variation is expensive and slow, which is why researchers have turned to a shortcut , prompting LLMs to pretend to be members of different demographic groups and annotate from those perspectives. But a new paper from researchers at Universidade da Coruña asks the uncomfortable question: do these persona-conditioned models actually behave like the humans they are supposed to simulate?
The short answer is: not reliably, and it depends heavily on which model you use. But buried in those results is a genuinely surprising finding about a different prompting strategy that works much better.
Why this problem is hard to ignore
In practice, content moderation often collapses diverse viewpoints into a single ground-truth label, risking systematic under-enforcement or over-enforcement for certain communities and raising concerns about fairness, accountability, and the inclusivity of automated systems. The NLP research community has been pushing back on this for years under the banner of perspectivism , the idea that annotator disagreement is not noise to be averaged away, but a meaningful signal about how different people experience language.
LLMs are increasingly used as proxy annotators or social simulators, and persona conditioning , prompting models to adopt specific demographic identities , has been proposed to recover subjectivity by simulating different perspectives. Prior work has shown that conditioning GPT-3 on demographic backstories can produce response distributions resembling those of particular groups. But that work mostly checked whether LLM outputs matched human averages. This paper asks something harder: does the social structure of disagreement hold up?
Three tests, not one
The researchers broke the evaluation into three distinct research questions, each probing a different dimension of human social judgment:
- Inter-group disagreement (RQ1): Do LLM personas from different demographic groups disagree with each other in the same way that real humans from those groups disagree?
- In-group sensitivity (RQ2): Do annotators judge content more harshly when it targets their own identity group, and do LLM personas show the same shift?
- Vicarious prediction (RQ3):