A new arXiv preprint introduces reward models designed to train language models to cite retrieved evidence when answering legal questions and to decline when that evidence is inadequate. The authors note that most existing reward models are not built for these high-stakes requirements. The work targets legal applications where unsupported or fabricated answers carry significant risk.
The New Mexico Supreme Court fined defense lawyer Stephen Aarons $5,000 and held him in contempt after he filed an appeal brief containing fake police testimony and nonexistent witnesses generated by ChatGPT. Aarons said he did not know the chatbot could invent facts. The court is also requiring him to notify his client and other parties about the fabricated content.
WHY IT MATTERS ↘Sanctions like this establish that liability for unverified model output falls on the professional who files it, not the vendor, which raises the cost of deploying general-purpose chatbots in regulated work and strengthens demand for retrieval-grounded tools that cite verifiable sources. It also gives courts a template for graduated penalties that other jurisdictions and regulators are likely to copy.
Researchers have proposed GANDR, an approach that audits legal question-answering systems at the level of individual claims instead of scoring whole answers. Each statement generated by the model is checked against the specific source it cites, so readers can verify a grounded answer piece by piece. The work targets a shortcoming of existing grounded-generation pipelines, which typically evaluate answers only in aggregate.
Researchers have introduced LexAgentHallu, a new benchmark for measuring how tool-augmented legal AI agents hallucinate. It uses a hierarchical structure to trace how errors in tool calls and reasoning cascade into fabricated case holdings and miscited legal authority. The benchmark aims to fill a gap left by existing legal evaluations that do not capture agentic workflows.
A new paper examines whether and how a defendant's courtroom statement shifts the verdicts of large language models simulating jurors in common-law trials. The authors focus on persuasion effects and potential bias in these simulated decisions, a setting where LLM-based modeling of human judgment had not previously been studied.
A revised arXiv paper investigates whether generative AI models can reproduce the legal judgments that humans make, a technique known as "silicon sampling." The authors situate this question within growing interest from scholars, lawyers, and judges in using AI to inform legal decisions. The work suggests that understanding how closely models track human legal reasoning matters as people turn to AI for personal guidance.
Australian law firm Gilbert + Tobin has detailed how it scaled OpenAI's ChatGPT Enterprise and Codex across its organisation. The firm credits executive sponsorship, formal governance controls and clear human oversight for the deployment. The case study highlights how professional services firms are adopting generative AI while keeping accountability with people.
WHY IT MATTERS ↘It gives a concrete template for regulated professional-services firms: executive sponsorship plus formal governance and human sign-off can unlock firm-wide generative AI deployment where pure productivity arguments stall. That raises the bar for vendors competing on enterprise controls, and shifts the differentiator from model capability to auditable accountability.