A Judge, Not a Regulator, Draws the First Line
On June 1, 2026, the Judicial Yuan's Judicial Weekly special-issue series (司法周刊文選別冊, Issue 2310) published a lengthy essay by Liao Jian-Yu, a presiding judge on the Taiwan High Court's criminal bench and a law PhD from National Cheng Kung University, titled "AI's Use and Limits in Courts." Liao's essay is not a rule. It carries no binding force. But it is the most detailed judicial argument yet for how Taiwan's courts — as opposed to executive agencies or private industry — should govern generative AI, and it lands at a moment when the rest of Taiwan's AI governance architecture is already moving without the judiciary in the room.
Liao's core proposal is a risk matrix: sort judicial AI use cases into low-, medium-, and high-risk zones based on how directly they touch fact-finding, sentencing, or case outcomes, then match each tier to a different level of human oversight. Alongside it, he calls for detailed AI usage logs to make AI's role in any given case traceable after the fact, and what the news hook describes as dual-track disclosure rules — separate obligations for court staff drafting internal materials and for external lawyers or litigants who submit AI-assisted filings. The essay explicitly names the three risks now showing up in courtrooms worldwide: hallucinated case citations, algorithmic bias in predictive tools, and deepfake evidence.
The Risks Are Not Hypothetical
Those warnings track what is already happening abroad. Damien Charlotin's widely cited AI Hallucination Cases Database — built specifically to track court rulings where judges found a party relied on fabricated AI output — has become a standard reference for the scale of the problem; a single U.S. court flagged 17 separate rulings noting suspected AI hallucinations in filings in one day, March 31, 2026, according to a Reason/Volokh Conspiracy analysis of the tracker. Deepfake evidence authentication is now a live issue for judges managing exhibits, not a speculative one. Liao is not inventing a problem — he is importing a documented one before it reaches Taiwanese dockets at scale.
Taiwan's judiciary has direct experience with the caution required. The Judicial Yuan disclosed on August 27, 2023 that it had built a generative AI tool to draft criminal judgments — but limited it to two narrow, high-volume categories: dangerous driving (DUI) and aiding-fraud cases, where facts are largely undisputed. The system runs on-premises, is fine-tuned on two decades of prior rulings, and — per the Judicial Yuan's own response to a 2023 civil-society press conference raising concerns about AI-drafted judgments — deliberately favors rule-based natural-language processing over pure generative output specifically to avoid ChatGPT-style hallucination. Judges retain full authority over guilt, sentencing, and legal application; the AI only assembles a draft from indictment text. That caseload pressure is real: criminal court judges handled an average of 59.67 cases a month as of 2022, up from 56.1 in 2013, roughly a quarter of them DUI matters — the exact bureaucratic strain AI drafting tools are meant to relieve.
The Governance Gap Liao Is Actually Filling
The more interesting story is structural. Taiwan's Artificial Intelligence Basic Act was promulgated January 14, 2026, and it already mandates a national risk-classification framework: Article 16 tasks the Ministry of Digital Affairs with building an internationally aligned AI risk taxonomy, and Article 19 requires government agencies to run risk assessments and set usage guidelines for their own AI use. But MODA sits under the Executive Yuan, and Taiwan's courts are a constitutionally separate branch. The AI Basic Act's risk framework was built for the executive branch and regulated industry — it does not, and under the separation-of-powers principle probably should not, bind judicial decision-making directly. That leaves exactly the gap Liao's essay steps into: without a judiciary-specific framework, Taiwan's courts would either improvise case-by-case or import an executive-branch taxonomy that was never designed with judicial independence in mind.
The Case for Getting There Carefully
The strongest argument for stringent AI rules in courts is not abstract. A hallucinated citation in a brief is embarrassing; a hallucinated citation a judge relies on in a ruling is a due-process failure with a defendant's liberty attached. Algorithmic bias in sentencing-adjacent tools is not a hypothetical either — it is a system faithfully reproducing skewed patterns already present in historical case data, which is precisely why regulators worldwide, including the EU under its AI Act's high-risk classification for judicial-administration systems, have singled out courts for the strictest tier of scrutiny. Liao's own essay takes that risk seriously rather than waving it away.
But a risk-tiered, disclosure-based model — rather than a blanket ban or a rigid import of executive-branch rules — is the right instrument, and Liao's framework is closer to that than to prohibition. Confining generative drafting to low-stakes, high-volume, largely undisputed case types, as the Judicial Yuan already does, while reserving heavier audit trails and disclosure duties for anything touching fact-finding or sentencing, lets courts capture real efficiency gains — a meaningful lever against a caseload that has grown steadily for a decade — without letting a chatbot anywhere near a verdict. The open question is whether the Judicial Yuan converts Liao's essay into an actual internal directive with teeth, rather than letting it sit as one judge's well-argued position in a weekly circular. Given that MODA is already moving on the executive side, a judiciary that governs itself on its own timeline, calibrated to its own constitutional role, is a far better outcome than a court system regulated by default because nobody defined the rules in time.