Even specialized retrieval-augmented professional tools remain unreliable enough to make human verification indispensable: Magesh et al. 2025 found leading legal research tools "each hallucinate between 17% and 33% of the time," informing "the responsibilities of legal professionals in supervising and verifying AI outputs." Tsamados et al. 2024 add that for some deployment contexts "our risk tolerance for a particular task or environment is too low" for AI integration to be a net benefit at all.
The Magesh quotes (17%–33% hallucination rates; informing legal professionals' responsibilities in supervising and verifying AI outputs) are verbatim, and E38 contains the Tsamados risk-tolerance quote with the conclusion that integration would not become a net benefit even with human control.
Written by Kimi K3 via Ollama Cloud · checked by GLM-5.3 via Ollama Cloud · 3 Oct, 04:46
Source chain
AEvery quote below was checked, without a model, to appear verbatim in its source.
- 01
“we find that the AI research tools made by LexisNexis (Lexis+ AI) and Thomson Reuters (Westlaw AI‐Assisted Research and Ask Practical Law AI) each hallucinate between 17% and 33% of the time”
Abstract · no page number · evidence E27
- 02
“it provides evidence to inform the responsibilities of legal professionals in supervising and verifying AI outputs”
Abstract · no page number · evidence E27
- 03
“it may be that our risk tolerance for a particular task or environment is too low and incompatible with the unpredictable behaviour that is inherent in AI systems”
Full-text passage · no page number · evidence E38
Indeed, rather than just seeking to combine human abilities and AI capabilities to complete a task more efficiently, the HMT literature can benefit from expanding its research into how this combination also creates a more robust entity vis-à-vis unforeseen risks of all types, including unprecedented socio-legal risks. Preliminary to the development of any approach to control AI systems, whether based on foundation models or not, is the identification of acceptable risk thresholds. To this end, it is crucial to conduct a systematic risk analysis of the deployment context and consider the extent to which AI can cause damage and harm and the risk appetite linked to specific uses of AI technologies. Indeed, it may be that our risk tolerance for a particular task or environment is too low and incompatible with the unpredictable behaviour that is inherent in AI systems (thus far) and that even with human control, the integration of AI would not become a net benefit from an operational perspective.
Passage read from www.ebi.ac.uk, which may be a preprint rather than the published version.