Tools to generate realistic prompts help surprisingly little with Petri audit realism

A new study from the Anthropic Fellows Program reveals a significant limitation in current methods for auditing large language models. Researchers trained specialized prompt generator models on real WildChat data and gave them as tools to Petri, an automated auditing agent, to test whether more realistic synthetic prompts could make AI audits harder to distinguish from real deployment conversations. Despite creating single-turn prompts that were substantially harder to distinguish from human-written ones (outperforming simple Claude Haiku 4.5 baselines), these generators failed to improve the overall realism of full audit transcripts.

The research suggests that audit realism is bottlenecked by higher-level features rather than individual prompt quality. Analysis of the realism judge's reasoning indicates that the harmful nature of audit scenarios and the unnatural structure of multi-turn conversations remain the primary giveaways. This finding redirects focus from prompt-level improvements to structural solutions like grounding audits in real deployment data, planning more natural interactions, or designing better scenario seeds. The study concludes that current prompt generation methods aren't worth integrating into automated auditing agents like Petri in their present form.