Courts, schools and police give AI different authority
A judicial trial in Pakistan, South Korea's marking proposal and Florida police purchases of smart glasses reveal three different forms of human oversight in government.
Artificial Intelligence··Evening
The judicial assistant went through a large trial
The study reported by IEEE Spectrum measures an AI assistant directly within a public workload. Sultan Mehmood of the New Economic School and colleagues connected OpenAI's GPT-4 to a knowledge base of nearly 130,000 Pakistani judicial opinions and statutes, creating a legal research and judgment-drafting tool. Beginning in 2024, the assistant was offered to 1,559 trial judges, roughly half of the country's justices. Judges who received the tool together with training resolved 6.3 per cent more cases; the study found no obvious decline in judgment quality and recorded fewer appeals. The trial sits within a severe backlog: Pakistan's judiciary has 2.26 million pending cases and fewer than 2 judges per 100,000 people, compared with 22 in the EU and 8 in Brazil. The scope of the result matters: the number belongs to a group that received training as well as the assistant. The 6.3 per cent increase cannot separate the tool's effect from the effect of training. Final authority remains with the judge; the trial measures how a research and drafting assistant relates to the number of cases completed.[1]
The model drafts the score; the teacher decides
South Korea's National Education Commission presented AI first-pass marking at an Incheon hearing on August 11 as an answer to fairness objections that have made essay questions difficult to use in large examinations. Lee Min-seok, head of the education panel of the national AI strategy committee, said the approach would begin in low-stakes school assessments and later expand to high-stakes examinations such as the CSAT university entrance test. In the described procedure, teachers and specialists write the questions and marking criteria. The model re-marks sample answers so the criteria can be calibrated; after the examination it produces a first draft of scores, which a teacher reviews and finalizes. Lee said final authority over questions and assessment would always remain with teachers, while time saved could go to individual feedback and lesson design. This is a plan presented to the public, rather than an implementation result: hearings continue in Gwangju on August 20 and Daegu on August 26, and a draft national education plan is expected at the end of October. Human oversight is built into the workflow as the step that makes a score final.[2]
The purpose of smart glasses differs by agency
According to Florida documents reported by Gizmodo via the Miami New Times, at least 2 South Florida sheriff's offices purchased Ray-Ban Meta AI glasses. The Broward County Sheriff's Office said it bought the glasses for its undercover Internet Crimes Against Children unit, but declined to provide details, citing its policy on surveillance methods. The Okeechobee County Sheriff's Office said its 6 pairs were used for remote troubleshooting in an IT role. Gizmodo places these purchases alongside earlier reports about smart glasses being developed for border surveillance with a possible link to a biometric database, and about ICE officers wearing Ray-Ban Meta AI glasses during raids. The distinction across the three public fields is clear: the Pakistan trial measures completed cases and appeals; the South Korean proposal explicitly leaves final scoring authority with teachers; and the Florida records give purposes ranging from technical support to undercover work for the same product category. In the last example, the publicly visible oversight consists less of an outcome measure than of procurement records opened through information requests and the agencies' limited explanations.[1], [2], [3]
