Starting in 2028, a "judge AI assistant" will be introduced in criminal trials nationwide. The AI will draft a preliminary statement of reasons for sentencing in judgments and review whether the sentence was calculated in line with the sentencing guidelines. The aim is to cut judges' drafting and review time so they can focus more on hearings and persuading the parties.
According to legal sources on the 6th, the Supreme Court plans to invest a total of 10.3 billion won by 2028 to build an "AI-based sentencing support platform." To that end, it will also develop a large language model (LLM) specialized for sentencing. It plans to soon issue a bid notice and select a contractor. Development could begin as early as Sept. this year.
The core of the project is to implement a so-called "AI specialized for sentencing." First, the Supreme Court plans to use Generative AI to support drafting preliminary statements of reasons for sentencing. For example, when a judge selects the sentencing factors to apply to a case and inputs additional case details, the AI drafts a preliminary statement of reasons based on that. Frequently used combinations of sentencing factors can also be saved in advance by case type and retrieved.
The AI will also review draft judgments. It extracts the type of sentence, sentence length, offense name, and ancillary dispositions from a draft, then checks the applicable statutory provisions. It then verifies whether the sentence imposed falls within the ranges of the maximum allowable sentence and the recommended sentence. If an ancillary disposition such as an order to disclose personal information or an order to complete a sex offender treatment program is missing from the decree, it will flag the omission.
Searching for judgments with similar sentencing is also expected to become easier. For example, if a judge inputs case details conversationally, such as "caused an accident after drinking and fled," the AI will find relevant judgments. The search results will include summaries focused on aggravating and mitigating factors that influenced the sentence. It will also visualize, in a graph, what types of sentences were imposed in similar cases.
To that end, the Supreme Court also plans to build a separate sentencing-specialized LLM using internal court data. Based on internal data, it aims to minimize hallucinations and, using retrieval-augmented generation (RAG), present sources and grounds for its answers.
Behind the Supreme Court's development of an AI-based sentencing support platform is a surge in criminal cases. According to the Court Statistics Monthly, the number of first-instance criminal panel cases rose 72% in five years, from 15,050 in 2020 to 25,922 in 2025. During the same period, the average case processing time increased from 176.5 days to 206.5 days. On top of that, as the number of crimes covered by the sentencing guidelines expanded from seven to 48, the burden of entering data on sentenced cases also grew.
A Supreme Court official said, "Once the AI-based sentencing support platform is introduced, the burden on judges to draft judgments and review sentencing guidelines will decrease," adding, "We expect that will allow more focus on hearings."