AI model quality improved in 2025, but hallucinations still happen often enough to impact legal, medical, technical, and financial workflows. Research across retrieval, long-context, and citation tasks shows that hallucination rates vary significantly by prompt type and guardrails.
Graph: Approximate hallucination rates from selected 2025 evaluations
(Lower is better. These ranges represent consolidated public benchmark reporting and task-specific studies, not a single benchmark.)
| Task Type | Rate (Approx.) | Visual |
|---|---|---|
| Grounded RAG Q&A | 6% - 12% |
▇
▇
|
| Citation-Constrained Summaries | 8% - 15% |
▇
▇
▇
|
| General Open-Ended Answers | 12% - 22% |
▇
▇
▇
▇
|
| Long-Context Synthesis | 18% - 30% |
▇
▇
▇
▇
▇
|
Takeaway
AI is strongest when grounded to trusted sources. The best production workflows in 2026 combine retrieval, source citation, and human review before publishing.