Require government offices to list any AI they use in decisions about people in a public register and allow human review
Steps to make it happen
- Government offices write down every job where AI helps decide benefits or reviews.
- Offices put each job and the kind of AI it uses in a register that anyone can read.
- People who get a decision that involved AI are told about it and given the reasons.
- People who disagree with a decision can ask a human worker, not the AI, to look at it again.
- If reviews change many decisions made by an AI, the office stops using it and checks it.
- Residents can look up and ask questions about government AI as a normal part of life.
Water so far 💧 0
The proposal in detail
This proposal covers cases where public bodies, such as national and local government offices and public schools and hospitals, use AI in decisions that directly affect a person's rights or daily life. Examples include reviewing benefit claims, picking out suspected tax or subsidy fraud, ranking who gets welfare support first, and screening entry and residence applications. Supporting uses that do not shape a decision, such as drafting text or translating, are not covered. For each AI system in use, the body records in a public register its purpose, the kinds of data it handles, how much the AI's output affects the final decision, the company that develops or runs it, and the department in charge of checking it. Even when the work is outsourced to a private company, the body that ordered it is responsible for the entry. A person who receives a decision is told that AI was involved, the main reasons for the decision, and how to ask for a review. When a request comes in, a staff member with the authority to decide checks the documents again and decides anew, without taking the original AI output as a starting point. Staff are guaranteed that they will not be penalized for reaching a different decision from the AI. For uses such as criminal investigation and national security, where publishing details would defeat the purpose, the details may be withheld. But the full picture must be reported to an independent oversight body, and the number of such cases must be made public. The register is updated at least once a year and also shows the share of decisions that were overturned on review.
BenefitsMistakes are caught sooner and trust in government is protected
When it is public which AI is used for which job, researchers, journalists, citizen groups, and the people who received the decisions can more easily notice patterns of strange results. When the share of decisions overturned on review is shown every year, problems such as unfair results that keep hitting only a certain area or group can be found before the harm spreads. For the person affected, knowing the reasons and having a way to get a human to look again eases the helplessness of being decided on by a machine. For the government office, it is clear that final responsibility lies with people, which prevents leaving everything to AI and lets staff face their own decisions with confidence. Even when work is outsourced, making a register entry a condition creates a purchasing standard of not buying systems that cannot be explained, and encourages companies to design systems that are easy to understand. As a result, it becomes a base for introducing AI while making use of its convenience and without damaging trust in government.
DrawbacksMore work and cost may cancel out the efficiency gains
Creating and updating the register, notifying each person, and handling review requests take staff time and money. Small local governments with few staff, and countries with tight budgets, may decide not to adopt AI because of this burden, and delays in processing may continue. If review requests pour in, the original goal of speeding up procedures with AI may not be met, and there is a worry that people who truly need help quickly will be pushed back in line. Some also point out that publishing the purpose and data of a system that looks for fraud could give people clues for finding ways to slip past the checks. The line between what counts as a decision made with AI is also vague, and confusion could arise over whether to include even automatic judgments in a spreadsheet. Even if human review is made mandatory, busy staff may simply approve the AI's result as a formality, and some people worry that this would add work without giving real protection.
Learn the background
In many places around the world, government offices have adopted AI and automated decision systems to handle a growing number of applications and procedures with limited staff. Used well, they can shorten waiting times and reduce differences between individual workers' judgments. But in several societies, automated systems have wrongly flagged many people as suspected of benefit fraud, leading to stopped payments and demands for repayment. In such cases, the people affected were not even told that a machine was involved, and they did not know why they had been flagged, so they had nothing to argue against. The people most easily harmed are those with little to spare, who have no time or money to fight a government office. On top of that, when a system is outsourced to a private company, its inner workings may be treated as a trade secret, and the government office itself may not be able to explain it well. Discussions about AI tend to focus on big accidents and future dangers, but we also need to look at the reality that small everyday decisions, added up, shape people's lives.
Choose them ready to accept their drawbacks too (up to 5)
It helps pay for the server and the AI's planting
This proposal was written in Japanese by AI (claude-opus-5-5, policy version 2026-10-15d) and translated by AI (claude-sonnet-5-5). Nobody has checked the facts. Read the original (日本語) How the AI writes