AI Regulation Tracker  /  Judicial guidance

Taiwan lets AI draft criminal judgments while making judges verify every output

Taiwan's Judicial Yuan built an AI system to draft certain criminal judgments, but the government's generative-AI guidelines keep judges and staff personally responsible for checking every citation and fact. The result is a working guardrail against AI-fabricated authority.

By Anthony Guerriero  ·  Reviewed by The Leveraged Years Editorial Desk  ·  Published July 7, 2026  ·  Last updated July 7, 2026

Taiwan has done something few court systems have attempted. Its Judicial Yuan built an AI system to draft criminal judgments in specific, high-volume case types, and trialed it in selected courts from September 2023. At the same time, the government has made clear that the machine does not decide anything. The judge does, and the judge is answerable for every word.

What the courts built

The AI drafting tool was designed for narrow, repetitive categories: driving under the influence and fraud-assistance cases, where the factual patterns and sentencing structures recur. In those categories, the system can produce a draft judgment that a judge then reviews. The value is speed on routine matters. The limit is deliberate. The tool was not built to weigh contested evidence or to exercise discretion, and Taiwan has not claimed that it does.

That case selection is the whole point. DUI and fraud-assistance prosecutions arrive in volume, follow a predictable structure, and turn on facts that are usually established rather than disputed. Automating a first draft there frees judicial time without asking the model to do the work courts exist to do. It is a targeted efficiency measure, not a handover of adjudication, and the boundary between the two is where the verification duty does its work.

The verification duty that comes with it

The guardrail is the Executive Yuan's Reference Guidelines on the Use of Generative AI (行政院及所屬機關使用生成式AI參考指引), which the Judicial Yuan adopted as its vetting framework. The guidelines set a plain rule for anyone in government using generative AI. Users must objectively and professionally risk-evaluate the output. They must not fully trust it, and they must not directly use unverified generative-AI output as the sole basis for a decision or administrative act. Independent human judgment cannot be delegated to the model.

Applied to a courtroom, that rule is precise. A judge who accepts an AI-drafted judgment retains full authority over the facts, the law, and the sentence. The AI text is a starting point, not a source of authority. Every case citation, every statutory reference, and every factual assertion in the draft has to be checked against the record and the primary law before the judgment is issued.

The framework's reach extends past the bench. Because it governs the whole of the Executive Yuan and its subordinate agencies, it sets a duty pattern that court staff, and by practical extension the practitioners who appear before those courts, are expected to internalize. The lesson generalizes cleanly. Anyone in Taiwan producing legal or administrative text with generative AI works under the same expectation: evaluate the output, do not trust it on its face, and confirm it before it carries weight.

Why this matters against fabricated citations

The most concrete risk with generative AI in legal work is the fabricated citation, a case or authority that reads convincingly but does not exist. Taiwan's framework meets that risk head-on by design. Because the human user, not the tool, is responsible for the output, an AI-invented citation is the drafter's problem to catch, and the drafter has an explicit duty to catch it. The duty is a guardrail, not a decided sanction. No confirmed Taiwan-specific bar or court penalty over an AI-fabricated citation has been identified, so the point here is the standard of care the framework imposes, not any ruling that has tested it.

What it does not do

The guidelines are a non-binding reference framework, not a statute with its own penalties. They do not license AI to render decisions, and they do not remove any part of a judge's constitutional authority over facts, law, and sentencing. They also do not, on their own, create a fabricated-citation offense. They set an expectation of independent verification and place responsibility on the person who signs the output.

For a US reader, the contrast is instructive. In the United States, the fabricated-citation problem surfaced through sanctions, most visibly in Mata v. Avianca, where lawyers were penalized after filing AI-invented cases. Taiwan is unusual in the other direction: it is building the AI into the judiciary itself while mandating human verification up front. The practical standard converges. Whoever puts their name on the document owns its accuracy, and an unverified AI citation is the signer's liability regardless of jurisdiction.

Browse the full AI Regulation News tracker

Informational analysis for working professionals, not legal advice. Confirm how any rule applies to your situation with qualified counsel.