Posted 1/11/2024, 9:08:00 PM
Study: Legal AI models often generate false information and overstate confidence
- Generative AI models frequently "hallucinate" false legal information, with incorrect answers 69-88% of the time
- Models perform worse on complex legal questions and lower court cases
- Models fail to contradict false premises and overstate confidence in responses
- Risk that AI could deepen existing legal inequalities rather than improve access
- Consequences already seen with lawyers citing fake cases from ChatGPT