Japan built AI to draft answers for Parliament. Preserve the question-to-answer file

THE SIGNAL IN ONE SENTENCE

Parliamentary answers are not ordinary office documents. They become part of the public record. They explain what a government believes, what it knows, what it plans to do and which promises can be quoted back later. That makes them a tempting place for automation and an awkward place for mystery. Japan's Digital Agency launched a prototype on October 30 to help government officials prepare draft answers to questions in the National Diet. The application runs inside Government AI GENAI and is available to approximately 180,000 government employees during the 222nd Extraordinary Session. The prototype supports several pieces of the work: understanding a parliamentary question, finding previous answers and supporting documents, organizing information, generating a draft, anticipating follow-up questions and checking for contradictions. That is a fairly sensible list. Officials often work under short deadlines and need to find the government's earlier statements before writing a new one. Search is repetitive. Terminology drifts. A relevant answer may be buried in years of meeting minutes. Nobody wins a medal for opening the nineteenth PDF. The Digital Agency says 102 experienced officials from 16 ministries and agencies participated as evaluators. Their feedback on the interface, usability and needed functions has already shaped the prototype. The agency plans to test effectiveness, usability and output quality during the current Diet session, improve the system and move toward full-scale provision during the ordinary session beginning in January. The agency also puts a useful brake in the announcement. It says AI output should not generally be adopted exactly as provided. Officials must compare the output with supporting documents, verify factual accuracy, check consistency with previous answers and government policy and judge whether the wording is appropriate. Policy decisions, official government views and final answers remain human responsibilities. Ministers and other respondents remain responsible for what they say. Good. Keep that rule. Then make it inspectable. For every parliamentary question, preserve one question-to-answer file. It should connect the original question, retrieved records, generated passages, human edits, policy checks, reviewers, approvals, final answer and any later correction. The file does not need to publish every confidential internal note. It does need to make the chain of responsibility clear enough that an authorized reviewer can reconstruct how an answer came to exist. The starting point is the question itself. Save the exact text received, the member who submitted it, the chamber, committee or session, the date, the deadline, the responsible ministry and a stable identifier. If the question changes, preserve the earlier version and record who changed it and why. This matters because a system can answer a nearby question very fluently. A small shift in wording can move the legal or policy meaning. If the original question disappears behind a summary generated by the same model that drafts the response, reviewers may never notice the turn. Next comes retrieval. The Digital Agency has already described a related GENAI application that takes a natural-language request, extracts several search terms and looks through Diet proceedings for relevant questions and answers. That can make old records easier to find than repeated manual keyword searches. It can also create a hidden filter. The system chooses search terms, decides which records look relevant and presents some results before others. A reviewer may see a tidy bundle without knowing which documents were missed, excluded or ranked lower. The question-to-answer file should therefore keep the search trail. Record the query, extracted terms, filters, repositories, retrieval time, result identifiers, ranking scores where available and the documents actually opened. Preserve the version or snapshot of every cited source. Do not keep only the passages the model liked. If an earlier government answer is important, link to the complete official record and show the surrounding exchange. A sentence that looks clear in isolation may have been qualified by the next paragraph or corrected later in the session. Then record the generated draft. Save the model and version, system instructions, approved knowledge sources, generation time and exact output. If a second model checks the draft, record that model separately. If an employee asks for a rewrite, keep the instruction and the new version rather than silently replacing the first attempt. Version history is not clutter here. It is the evidence of judgment. Suppose the AI says a program began in April, an official changes it to May and the final answer says June after consultation with another ministry. The useful record is not merely the June sentence. It is the sequence showing which evidence changed the date, who made the correction and which source settled the disagreement. Human review needs more than a checkbox. The file should show which official checked facts, which legal or policy team checked consistency, which division coordinated with another ministry and who authorized the final text. Each reviewer should be able to mark a passage accepted, edited, rejected or unresolved and attach a short reason. That sounds bureaucratic because it is bureaucratic. Parliament is where bureaucracy is supposed to become accountable language. The interface can make the work lighter. Place the original question beside the draft. Make every factual sentence clickable to the supporting source. Flag generated claims with no attached evidence. Show where a passage differs from a previous answer. Mark the date of each source. Warn when the source is a summary rather than the authoritative record. Let reviewers compare versions without opening five windows and one tragic spreadsheet. The system should separate facts from policy. A fact might be the number of applications received by a stated date. A policy judgment might explain whether the program is sufficient or what the government intends to do next. Retrieval can help with the first. It cannot make the second into an objective lookup. If the draft contains a policy position, the interface should identify the responsible human authority and require explicit approval. The model can suggest phrasing, but it should not smuggle a new commitment into a sentence because that wording resembles older material. Follow-up questions deserve their own caution. Predicting likely follow-ups can help officials prepare evidence and identify weak points. It can also encourage answers optimized to manage a debate rather than answer the question plainly. Keep predicted follow-ups labeled as preparation material. Do not let them replace the member's actual question or quietly reshape the answer around imagined political risk. Evaluation should test the whole chain. The agency says it will examine effectiveness, usability and output quality. Those categories need specific measures. For retrieval, measure whether the authoritative records were found, not merely whether the first results looked plausible. Test questions involving old terminology, policy changes, multiple ministries, corrected records and ambiguous names. For drafting, score factual accuracy, completeness, source support, consistency with current policy and clarity. Count unsupported claims, invented citations, omitted qualifications and contradictions with official records. For workflow, measure time saved separately for search, drafting, review and coordination. A model that writes quickly but doubles verification time is not a simple productivity gain. For accountability, ask whether a reviewer can reconstruct the final answer from the preserved file. If a sentence is challenged six months later, can the ministry identify its source, the generated version, the edits and the approving official? Break results down by ministry, question type and document age. A system may work well for recent administrative statistics and poorly for questions that require legal interpretation or coordination across agencies. Report human disagreement too. If experienced officials regularly reject a particular kind of output, that is valuable evidence about the boundary of the tool. Do not average the pattern away under one satisfaction score. There is also a public-record question. Parliamentary answers are eventually public, but internal drafting materials can include confidential information, deliberative notes or personal data. The question-to-answer file needs access controls and a retention policy that distinguish material required for oversight from material that should remain protected. At minimum, preserve a durable internal audit record. Publish a public-facing evidence note when practical, listing the official sources relied upon and the date they were checked. If an answer is corrected, link the correction to the original answer and explain what changed. Avoid a second, quieter archive owned only by the AI system. The official Diet record should remain authoritative. The AI index, embeddings and summaries are aids for discovery. They should not become the only place where relationships between questions, answers and documents are stored. Export the retrieval and review history in a documented format. Keep stable links to the underlying records. Test whether the ministry can reconstruct the file without the current model or vendor. This is particularly important because GENAI is a government-wide platform. A successful prototype could spread quickly across ministries and future sessions. Small interface decisions may become national habits. A source panel that hides dates can normalize stale evidence. A green check that means only that another model agreed can become mistaken for verification. A revision history that disappears after approval can erase the work officials did to make the answer accurate. The opposite is possible too. An evidence-first interface can improve work even when the AI draft is mediocre. Better search, clear source links, visible version history and structured review can help officials produce more consistent answers. The model does not need to be an oracle to be useful. It needs to sit inside a process that remembers who knew what and who decided what. Start the prototype in advisory mode. Compare it with existing practice. Keep samples of questions where it helped, where it wasted time and where it nearly introduced an error. Test correction and appeal paths. Test what happens when a source is unavailable, a record has been amended or two ministries disagree. Then publish enough evaluation detail for the public to understand the boundary. How many questions used the prototype? How often were generated drafts substantially rewritten? How many unsupported claims or missed records were found during review? Did preparation time fall without increasing errors? Which tasks remained unsuitable? Those results would tell readers far more than the number of employees who could open the application. The democratic test is pleasantly old-fashioned. When a minister answers a question, the words should still belong to accountable people. The AI may retrieve, organize, compare and draft. The file should show the rest: evidence, doubt, revision and responsibility. If the system makes that chain easier to follow, it can improve parliamentary work. If it makes the chain disappear behind a clean paragraph, it has automated the wrong part.