GLM-Legal
A new proprietary legal-focused AI model from Chinese lab Zhipu AI has revived debate over how far large language models should be used in law, given their tendency to hallucinate cases and the high stakes of criminal, civil, and regulatory outcomes. Commenters weigh potential benefits — faster research, cheaper contract review, and improved access to legal help in overloaded systems — against concerns about attorney accountability, disclosure of AI use, privacy, data sovereignty, and the erosion of junior training pathways. Many see narrow, supervised use for routine tasks as plausible, but doubt that current systems can safely handle novel fact patterns or replace human legal judgment.
Use of AI by lawyers and disclosure
- Debate over whether lawyers should be required to disclose AI use.
- One side: disclosure is needed because AI increases error risk and can’t be sanctioned; clients deserve to know when AI is involved.
- Other side: lawyers are already fully liable for filings regardless of tools; courts should be vetting arguments anyway, so disclosure shouldn’t matter if the system is functioning properly.
- Concern about “vibe-practicing law” where lawyers rubber‑stamp AI outputs without understanding or checking them.
Quality, hallucinations, and benchmarks
- Frequent concern about LLMs hallucinating cases or misreading law; judges have sanctioned lawyers for this.
- Some see hallucinations as a serious, persistent problem in domains without clear right answers (e.g., legal reasoning).
- Others note that hallucination rates in coding have dropped and expect mitigation in law via research agents and verification loops.
- The touted ~20% pass rate on the Harvey benchmark is viewed by some attorneys as unimpressive for real litigation, but others note it’s currently near state of the art; benchmark quality itself is criticized as uneven and sometimes normatively wrong.
Impact on legal profession and access to justice
- Many expect AI to replace a lot of routine work now done by junior associates/paralegals (research, boilerplate contracts, wills, redlines).
- Some report strong practical results using general LLMs for contract review with human final checks.
- Worry that automating “grunt work” will hollow out the training pipeline for future experts.
- Others emphasize upside: cheaper access to legal help, especially for pro se litigants and overburdened systems (e.g., Indian court backlogs, settlements, arbitration‑like AI adjudication).
Model openness, jurisdiction, and data privacy
- Strong demand for open weights and transparent training data, especially in law/health; closed, hosted-only models are called “DOA” for serious legal use.
- Additional concern that GLM-Legal is run by a China‑linked company: fears about sensitive legal data being stored in a foreign, non‑extradition jurisdiction; counter‑view that foreign custody can sometimes act like “super‑privilege,” though others argue courts can still pressure the user.
- Prior reports that the vendor’s tools silently snapshot workspaces deepen mistrust; some say they wouldn’t trust these providers any more than US frontier labs.
Domain specificity and limits
- Questions about which legal systems are covered (US/UK vs others) and whether a single model can reliably handle many jurisdictions.
- Several lawyers stress that real litigation often involves novelty, human judgment, and messy power dynamics, which current LLMs can’t fully capture.
- Consensus in the thread: LLMs are promising assistants for standardized tasks, but uncritical reliance for high‑stakes legal decisions remains dangerous.