Legal AI systems powered by retrieval-augmented generation continue to produce ungrounded responses at alarming rates, according to new research from the University of Luxembourg.

The study analyzed eight legal RAG systems across two legal corpora — the GDPR in English and French civil law — finding hallucination rates that ranged from less than 10% for the best-performing systems to nearly half of all responses in the worst cases.

Researchers Souvick Das, Sallam Abualhaija, and Domenico Bianculli conducted both claim-level and answer-level evaluations to measure how frequently these systems generate responses not supported by their source documents. The team validated their findings against 142 questions authored by legal experts.

False premises trigger highest error rates

The research revealed that questions containing incorrect assumptions — known as false-premise questions — produced the highest hallucination rates across all systems tested. These questions require AI systems to reject the underlying premise rather than attempt to answer based on flawed assumptions.

The study examined performance across different question categories and user personas, providing what the authors describe as a "fine-grained analysis" of hallucination behavior in legal AI applications.

Legal RAG systems face particular scrutiny because ungrounded answers can lead to serious professional and legal consequences. The domain requires exceptional accuracy given the high-stakes nature of legal advice and document analysis.

The persistence of hallucinations in legal AI systems raises questions about their readiness for widespread deployment in law firms and legal departments. Companies like Harvey and Supio are building specialized legal AI platforms, while broader legal tech firms like Eudia and Crosby focus on contract analysis and review.

The research comes as legal professionals increasingly experiment with AI tools for document review, legal research, and contract drafting. The findings suggest that human oversight remains critical for legal AI applications, particularly for complex queries or those involving nuanced legal interpretations.

The paper was submitted to arXiv on August 14, 2026, and represents one of the most comprehensive evaluations of hallucination behavior in domain-specific RAG systems to date.