A Governance Gap Closes — Slowly
On June 24, 2026, Premier Cho Jung-tai chaired the inaugural meeting of Taiwan's Cabinet-level National Artificial Intelligence Strategy Committee, designating education, healthcare, finance, and justice as the four initial sectors for government AI demonstration. The timing is significant not only for what it authorises but for what it finally scrutinises: the Judicial Yuan's AI judgment-drafting tool, quietly in operation since September 2023, will for the first time undergo a formal risk assessment under the AI Basic Act's six-month deadline — now falling in mid-July 2026.
For three years, Taiwan's courts have used an AI system to produce draft rulings in driving-under-the-influence and fraud-assistance cases. That tool now moves from operating under informal executive guidelines to mandatory statutory accountability. The question is whether the review, once completed, will produce proportionate safeguards or become a pretext for over-regulation of a genuinely useful judicial efficiency instrument.
What the Judicial Yuan Built
The Judicial Yuan announced its AI drafting system in August 2023, following months of internal testing. Built on the MT5 large language model and fine-tuned on 25 years of Taiwan court records spanning 1996 to 2021, the system accepts inputs from judges — guilty or not-guilty determination, confession status, applicable statutes, evidence sufficiency — and generates a complete draft judgment including the verdict, statement of facts, legal reasoning, and citations. It was designed specifically for high-volume, structurally repetitive cases: DUI filings alone constituted roughly a quarter of criminal court caseloads at the time of launch, and monthly caseloads had grown from 56.1 to 59.67 per judge between 2013 and 2022.
Judges retain full constitutional authority over factual determination, legal application, and sentencing. The tool produces a draft; the judge writes the judgment. Under the Judicial Yuan's 2024 policy plan, AI judgment drafting was framed explicitly as an efficiency initiative — not an adjudicatory one.
Three Years Without a Statutory Framework
The steelman case for careful review is real and worth stating plainly. Judicial AI operates in a domain where errors carry exceptional stakes — liberty, not merely convenience. Critics who raised concerns in 2023 had substantive grounds: the training dataset's quality was opaque, accountability procedures between the Judicial Yuan and its contractor were not publicly disclosed, and legal advocates noted that attorneys, prosecutors, and defendants had no access to examine the training corpus for embedded bias. A Purdue University study cited at the time found that 40 percent of users cannot detect AI errors when system accuracy falls below 50 percent. These are not frivolous objections.
The deeper problem was structural. The Judicial Yuan launched the tool under the Executive Yuan's internal generative AI guidelines — a sensible interim measure in the absence of legislation — but Taiwan lacked any statutory AI framework until the AI Basic Act took effect on January 14, 2026. For roughly two and a half years, one of the highest-stakes government AI deployments in Taiwan operated without formal risk classification, an audit trail, or public accountability mechanisms. The June 24 committee meeting did not create that gap; it finally begins to close it.
What the July Deadline Actually Requires
Taiwan's AI Basic Act, which the Legislative Yuan passed on December 23, 2025, requires all government agencies to complete risk assessments for existing AI applications within six months of the Act's effective date — placing the deadline in mid-July 2026. The Ministry of Digital Affairs is responsible for developing the risk-classification framework and verification tools that agencies, including the Judicial Yuan, will use to meet that obligation.
For a tool that generates draft judgments in criminal proceedings, an adequate risk assessment should examine at minimum: the composition and bias profile of the training dataset, whether system outputs can be meaningfully contested by defendants or their counsel, how errors in AI-generated drafts propagate into final judgments, and whether judges are genuinely reviewing AI output or anchoring on it. The AI Basic Act's seven guiding principles — including transparency, fairness and non-discrimination, and accountability — give the Ministry a clear normative foundation. The challenge is converting those principles into specific, measurable requirements for judicial AI rather than generic compliance checklists.
Proportionate Governance, Not a Ban
The pro-innovation case here is not that judicial AI should avoid accountability — it is that the accountability mechanism now being applied is, in principle, the right one. A risk assessment that results in specific disclosure requirements, audit logging, and defined channels for defendants to raise concerns about AI-generated content would serve both efficiency and due process simultaneously. A response that effectively prohibits AI drafting assistance, citing precautionary reasoning without examining the actual error profile, would impose real costs: backlogged dockets, judicial burnout, and delayed justice for litigants in time-sensitive cases.
Taiwan's AI Basic Act is notably less prescriptive than the EU AI Act's approach to high-risk AI in the administration of justice, which mandates conformity assessments, registration in a public database, and CE-marking requirements before deployment. Taiwan's framework asks agencies to assess and mitigate — not necessarily to halt. That calibration deserves credit. The test is whether the Ministry of Digital Affairs' July framework distinguishes between genuine accountability requirements and procedural overhead that slows deployment without improving outcomes.
A Template Others Are Watching
Taiwan's combination of a lean statutory framework, a six-month implementation clock, and a Cabinet-level committee explicitly naming justice as a priority demonstration sector is worth watching closely. Premier Cho's framing of Taiwan as a democracy-grounded AI development leader positions this governance exercise as more than compliance formality — it is reputational differentiation in a region where judicial AI accountability is rarely examined in public. Delivering a credible, proportionate risk assessment of the court system's AI tool by that July deadline would do more to build international trust in Taiwan's AI governance than any policy document alone can.