Answers are grounded in this paper and cite their sources.
Large language models may facilitate unlawful activities, yet this risk remains underexplored in studies assessing illicit-content recognition or narrow categories of illicit scenarios. We define complicit facilitation as model guidance that assists users with illicit instructions. To assess it, we develop a legally-informed, automated pipeline grounded in real-world cases and legal frameworks, and introduce EVIL, a benchmark covering 269 illicit scenarios and 50 illicit intents across Chinese and United States legal contexts. Evaluations indicate that complicit facilitation is widespread across models, with legal warnings and lawful alternatives often omitted. Risks vary across socio-legal contexts and are elevated for crimes against societal interests, non-violent offenses, and intents involving subjective motives or deceptive justifications. Models provide more complicit assistance when instructions describe users as older adults, racial minorities, non-religious individuals, or workers in lower-prestige occupations. Existing open-source safety benchmarks and alignment datasets offer limited coverage and mitigation of these risks.
No takes yet. Share an insight, caveat, or question.
Xie et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: