How legal AI is used
Prompt Engineering Impact
How prompting technique changes accuracy, completeness and relevance of legal answers.
Across eight prompting techniques in this illustrative scenario, accuracy ranges from 62 on zero-shot queries to a high of 91 for multi-step decomposition, with jurisdiction-scoped prompts scoring highest on relevance (92) and structured-output prompting leading on completeness (90). Chain-of-thought and role-based framing land in the middle of the pack. The spread suggests query structure, not just model choice, is a meaningful lever for output quality - pointing legal teams toward prompt design and jurisdiction-scoping as a lower-cost improvement path than model switching. These figures are illustrative, not measured benchmarks.
Die Daten
| Prompting technique | Accuracy | Completeness | Relevance |
|---|---|---|---|
| Zero-shot (basic query) | 62% | 55% | 68% |
| Few-shot (with examples) | 78% | 72% | 82% |
| Chain-of-thought | 84% | 80% | 86% |
| Role-based (act as lawyer) | 81% | 78% | 88% |
| Structured output (JSON/tables) | 88% | 90% | 84% |
| Multi-step decomposition | 91% | 88% | 90% |
| Jurisdiction-scoped | 86% | 82% | 92% |
| Citation-required | 82% | 76% | 80% |
Veranschaulichende Schätzung - ein Richtwert zur Szenariogestaltung, kein gemessener Benchmark. Diese sind nicht als gemessene Ergebnisse pro Anbieter zu verstehen.
Verwandte Forschung
- Prompt Type DistributionWhat lawyers ask AI to do, split across research, drafting, analysis and more.
- Risk & Hallucination RatesEstimated error rates across legal tasks, highest for citation and jurisdiction work.
- Top Legal Tasks by AI UsageThe legal tasks lawyers most often hand to AI, ranked by share of usage.