⏱️ Reading time: 8 min
A New York lawyer was sanctioned in 2023 for submitting six cases to a court that ChatGPT had completely fabricated: citations with real judges’ names and case numbers that never existed. That story, already repeated in different countries, is the ghost that haunts any language model that wants to work with legal texts.
📑 En este artículo
Z.ai, the Chinese company behind the GLM model family, introduced GLM Legal, a version of its technology designed specifically to reduce that risk in legal tasks.
TL;DR
- GLM Legal is Z.ai’s model adapted to draft and review legal documents with fewer citation hallucinations.
- The Chinese company, formerly called Zhipu AI, builds the system on its open-weight GLM model family.
- The legal market already has established rivals like Harvey and CoCounsel, backed by OpenAI and Thomson Reuters.
- Real cases of lawyers sanctioned for citing AI-fabricated case law are driving this kind of product.
- Z.ai had not published independently verifiable accuracy figures at the time of this article.
What GLM Legal Is
GLM Legal is the variant of Z.ai’s GLM models aimed at legal professionals: contract drafting, case file summaries, and regulatory queries. The company presents it as an additional citation verification layer on top of its base model, designed for lawyers and firms that need answers traceable to a source.
Z.ai is the commercial name adopted by Zhipu AI, a lab that emerged from Tsinghua University and one of the most active Chinese groups in open-weight models, alongside DeepSeek, Alibaba, and Moonshot AI. The GLM (General Language Model) family is its flagship line, and its recent versions compete closely with Western models on public benchmarks. Taking that base and turning it into a vertical product for law is a common industry move: any general-purpose model performs worse in domains where an incorrect answer has real legal consequences.
What Happened
The announcement arrived through the company’s official blog, at z.ai/blog/glm-legal, with no parallel English-translated statement, which is typical when a Chinese lab launches a product designed first for its domestic market. The post describes an assistant capable of drafting contract clauses, summarizing long case files, and answering regulatory questions while citing the source it consulted.
The company did not accompany the launch with a third-party-verifiable accuracy table or an independent technical report. That absence of public benchmarks isn’t exclusive to Z.ai: most legal AI products, open or closed, avoid publishing error figures, because a single documented hallucination case is enough to damage an entire firm’s trust. What is verifiable is Z.ai’s positioning: competing not just with other Chinese general-purpose models, but with Western companies that have sold this kind of tool for years.
Context and History
The most cited precedent in the industry happened in 2023, when a federal judge in New York fined two lawyers 5,000 dollars for submitting ChatGPT-fabricated cases to the court in Mata v. Avianca. Similar episodes have since repeated in Canada, the United Kingdom, and Australia, each reinforcing the same conclusion: a general-purpose model, trained to sound convincing, cannot tell a real citation from a plausible one.
That problem opened up an entire market of specialized legal tools. Harvey, backed by OpenAI, has sold assistants to large firms since 2023. CoCounsel, originally developed by Casetext and acquired by Thomson Reuters, integrates verification against real case law databases. LexisNexis offers its own assistant, Lexis+ AI, built on its proprietary document base. Z.ai’s legal model enters that space to compete, but with a difference in origin: it comes from an open-weight family, something none of its direct rivals currently offer.
For firms in Latin America, where Harvey’s or Lexis+ AI’s enterprise contracts are often out of budget, that licensing difference is exactly what makes the announcement worth watching closely, even though regional availability hasn’t been confirmed yet.
Technical Details
The mechanism these systems typically use, and which Z.ai’s post describes for its product, combines two pieces: a retrieval-augmented generation (RAG) layer that looks up the exact text of the law, contract, or ruling before answering, and a subsequent verification step that confirms each generated citation corresponds to a real document in that database. The difference from a generic chatbot isn’t the model’s size, but the fact that the answer is never built solely from what the model remembers from training: it’s built from what it manages to retrieve and confirm at query time.
flowchart TD
A["Lawyer's query"] --> B["Document retrieval (RAG)"]
B --> C["Generation with GLM"]
C --> D["Citation verifier"]
D -->|"citation confirmed"| E["Answer with source"]
D -->|"citation not found"| F["Answer blocked"]The cost of this approach isn’t computational but operational: each additional verification means another query to an updated document database, and that database has to be kept current with every legal reform or new ruling. A legal model that doesn’t update its case law index becomes obsolete faster than a general-purpose one, because law changes by decree, not through the gradual consensus that shapes everyday language.
There’s no direct way to verify from the outside how often Z.ai updates its document index for this product: the company didn’t publish that information on its blog. It’s the kind of data that can only be audited by testing the tool with recent cases and checking whether it cites case law that postdates the base model’s training cutoff.
Impact and Analysis
Z.ai’s entry into this market matters for two reasons. The first is geographic: an open-weight legal model, or one with a more permissive license than its rivals, could be adopted by firms and universities that can’t currently afford Harvey’s or Lexis+ AI’s enterprise contracts. The second is regulatory: any system that advises on local regulations faces a problem no model solves with better engineering alone, which is jurisdiction. A correct answer about a contract in Mexico can be invalid in Argentina, and the company hasn’t yet detailed how many jurisdictions its document base covers.
⚠️ Heads up: no legal AI assistant replaces a human lawyer’s review before filing a brief in court, no matter how many layers of automated verification it has behind it.
| Tool | Company | Base Model | Approach |
|---|---|---|---|
| GLM Legal | Z.ai | GLM family (open weights) | Drafting and regulatory queries with citation verification |
| CoCounsel | Thomson Reuters | Proprietary models + Westlaw databases | Document review and litigation for large firms |
| Harvey | Harvey AI (backed by OpenAI) | Adapted GPT | Comprehensive assistant for corporate firms |
| Lexis+ AI | LexisNexis | Proprietary models + proprietary document base | Legal research with case law citations |
What’s Next
Z.ai didn’t announce an availability timeline outside its home market or licensing terms for the product, two details that will determine whether firms in other regions can adopt it. Following the pattern of previous GLM family releases, the most likely outcome is that a reduced version of the model ends up published with open weights, while the citation verification layer and access to the updated document base remain a paid service. That combination of an open model with closed infrastructure is already the dominant business model among Chinese labs competing with US big tech in artificial intelligence.
If you work at a firm and want to form your own opinion, the most direct way to test it today is to read the original announcement on Z.ai’s blog and ask the assistant to cite a recent ruling from your jurisdiction to see if the source actually exists.
Frequently Asked Questions
What exactly does GLM Legal do?
According to the company, it drafts contract clauses, summarizes lengthy case files, and answers regulatory questions while citing the source it used to generate the answer.
Is Z.ai’s legal model open-weight?
The GLM family it derives from does publish open-weight versions, but Z.ai hasn’t confirmed whether the full legal layer, with citation verification, will follow the same license.
What happens if Zhipu AI’s legal assistant cites a fake ruling?
The company hasn’t published a third-party-verifiable error rate. The industry recommendation, which applies to any tool of this kind, is that a human lawyer confirm every citation before filing it with a court.
How does Z.ai’s tool differ from Harvey or CoCounsel?
The main stated difference is the open origin of its base model. Harvey and CoCounsel build on proprietary models from OpenAI and Thomson Reuters, respectively.
Is Z.ai’s legal product available outside China?
At the time of this article, the company hadn’t announced international availability or public pricing.
References
- Z.ai Blog: official announcement of GLM Legal, the GLM family variant for legal tasks.
- Z.ai: the company’s official website, formerly known as Zhipu AI, developer of the GLM model family.
- Wikipedia: Zhipu AI: history and background of the company behind GLM.
- Wikipedia: Legal technology: a general overview of AI tools applied to law.
📱 Enjoy this content? Follow @programacion on Telegram for daily tech content in Spanish: quick summaries, fresh content every day.
Featured image: Foto de Tingey Injury Law Firm en Unsplash
Did it work for you? Got a different error? Say so below: questions get answered and help the next reader.
Leave a comment
0 Comments