A civil ruling from Taiwan’s Taoyuan District Court has triggered debate in the legal community after the judge opened the decision with a rare warning: “Do not ask AI, do not use AI to write appellate briefs, and do not assume others cannot tell it was written by AI.” The plaintiff’s claim was ultimately dismissed.
The ruling, identified as Taoyuan District Court case 115 Su Zi No. 867, was widely discussed among lawyers. Some took it as proof that AI should not be used in legal drafting at all. Lawyer Lin Shang-Lun offered a different reading. He said the decision does not reject AI itself. Instead, it shows that AI is a tool for professionals, not a shortcut for amateurs trying to handle specialized legal work on the cheap.
The case, as Lin describes it, was about misuse rather than the technology itself
Lin said the core of the dispute has to be understood first. In his account, the plaintiff did not retain counsel and had no legal background. To save on legal fees, the person fed a case outline into a free or consumer-grade AI tool, then copied and pasted a complaint that looked complete on the surface but contained major problems.
That approach carries obvious risk in front of a professional judge and opposing counsel. For Lin, the court’s opening remark was a warning to litigants who want to cut costs by using AI as a substitute for professional legal work. It was not, he argued, a blanket instruction telling lawyers not to use AI.
Why consumer AI tools can fail in legal drafting
Lin broke the problem into two structural limits that large language models face in legal work.
Token limits and the “lost in the middle” issue
He said free or consumer-grade AI products available to the public have limited input and output resources. Legal practice does not deal in short prompts. A single matter can involve case files, contracts, or judgment analysis running into tens of thousands or even hundreds of thousands of words. Consumer tools often cannot process that volume of raw material in full.
They are also vulnerable to what Lin called “lost in the middle,” where the model pays more attention to the beginning and end of a long document while overlooking crucial clauses and facts buried in the middle. In legal work, that blind spot can alter the entire conclusion. The result is a higher chance of selective reading, broken reasoning, or outright error.
No specialist database support means a higher hallucination risk
Lin also stressed that law is a high-precision field with little room for error. A consumer chatbot is a general-purpose text generator, not a system built for legal retrieval. Without access to a proper legal database, such as a RAG vector database or newer MCP real-time retrieval tools, an AI model may generate text that sounds plausible but does not exist in the real record.
According to Lin, that can include invented statutes, fabricated judgments, and even made-up Supreme Court case numbers. If a litigant signs and files one of those hallucinated references in court, a judge can verify the problem quickly. That does not just weaken the credibility of the filing. If the submission invents facts out of thin air, it may create other legal risks as well.
AI output reflects the operator’s domain knowledge
Lin said some lawyers see an absurd AI-drafted filing and jump straight to the conclusion that AI is unusable. He understands the reaction, but says it misses how generative AI works in practice.
His rough formula is this: output quality equals base model capability, about 10%, plus workflow design, about 20%, plus the data and instructions fed into the system, about 70%. In other words, the old “Garbage In, Garbage Out” principle still applies.
He compared AI to an assistant with a strong memory but no background knowledge of its own. The clearer the material, the more precise the instructions, and the more complete the context, the better the output. What matters most, he said, is the user’s ability to handle context engineering and control the inputs.
That means AI output often mirrors the professional level of the person using it:
- A senior partner or managing lawyer with strong domain expertise, clear strategic instructions, and carefully selected precedents can get results closer to senior-level work.
- A trainee lawyer will usually get output closer to trainee-level work.
- A user with no legal training who treats AI like a wish-granting tool is likely to end up with a filing full of contradictions that cannot survive scrutiny.
For Lin, AI’s current value is not in replacing human judgment. It lies in cutting the time needed to build and refine a first draft. The tool is best used to lay down the initial structure, with an expert then revising it through knowledge and experience. If the operator starts from zero professional competence, AI cannot fill that gap. If the operator already has solid training, AI can raise productivity in a meaningful way.
Top law firms abroad are already treating AI as infrastructure
Lin placed the Taiwan debate in an international context. While some legal circles are still asking whether AI should be used at all, he said leading firms in the US and UK have already moved past that question. In those firms, AI is no longer viewed as an experimental tool. It is treated more like basic infrastructure, comparable to utilities or internet access, and used across the hierarchy from senior partners to junior trainees.
Even so, those firms are not simply relying on off-the-shelf consumer products. Lin said they build strict data-handling workflows, with dedicated teams organizing the most valuable precedents and internal know-how before feeding that material into AI systems. Over time, that process helps them build proprietary expert systems.
He also pointed to Singapore, where the government and judicial authorities have worked with legal AI company Harvey, and to top law firms in South Korea that have begun rolling out advanced legal AI systems. Inside some technology companies, he added, workflow rules have already shifted: first drafts are no longer expected to be written entirely from scratch by humans, but are generated in part by AI and then revised by staff.
A warning for Taiwan’s legal sector
Lin argued that if Taiwan’s legal profession stays overly cautious about AI, it could fall behind international practice over the long run. He framed that concern with a contrast: Taiwan can produce some of the world’s most advanced AI chips, yet there are still court clerks typing records by hand in the courtroom one word at a time.
In his view, that gap drains professional manpower from the judicial system and leaves Taiwan’s legal competitiveness lagging behind.
His reading of the Taoyuan ruling is narrow but clear. The court was warning against the use of consumer AI as a substitute for legal expertise by non-professionals. It was not, he said, a declaration that AI has no place in legal work.