Grounded Legal Generation
A standard for responsible behavior in legal AI systems, requiring the system to flexibly select response strategies—including direct answers, scope narrowing, clarifying questions, risk warnings, premise correction, or abstention—based on the grounding relationship between a legal claim and its authoritative sources.
Timeline (last 90 days)
The paper proposes a warranted legal generation standard, under which a system should flexibly choose among the following based on warrant relationships: answering directly, narrowing scope, asking clarifying questions, issuing risk warnings, correcting incorrect premises, or declining to answer when there is no basis for doing so.
The paper argues that in legal contexts, declining to answer or signaling uncertainty is often more responsible than forcing a seemingly complete answer, which contrasts with the current tendency of LLMs to always provide an answer.
The evaluation framework components proposed in the paper include benchmark records and claim boundaries, support labels, mixed response strategy scoring, risk weights, annotation reliability reports, and jurisdiction-specific authority ontologies.
All Facts (3)
The paper proposes a warranted legal generation standard, under which a system should flexibly choose among the following based on warrant relationships: answering directly, narrowing scope, asking clarifying questions, issuing risk warnings, correcting incorrect premises, or declining to answer when there is no basis for doing so.
50%UnverifiedThe paper argues that in legal contexts, declining to answer or signaling uncertainty is often more responsible than forcing a seemingly complete answer, which contrasts with the current tendency of LLMs to always provide an answer.
50%UnverifiedThe evaluation framework components proposed in the paper include benchmark records and claim boundaries, support labels, mixed response strategy scoring, risk weights, annotation reliability reports, and jurisdiction-specific authority ontologies.
50%