From the source
Harvey published a benchmark evaluating hallucinations on legal tasks.
- Its Assistant model hallucinated at a rate of 0.2% (1 in 500 claims), lower than Claude (0.7%), ChatGPT (1.3%), and Gemini (1.9%), despite providing longer and more detailed answers.
- The company defines hallucinations as factual claims demonstrably disproven by a source of truth and uses a system of models to detect and measure them.
