Where Does AI Get Local Government Information? Opening Up the Sources of 128 Answers

What We Asked

We put 8 questions citizens would realistically ask to the 8 metropolitan governments and asked AI twice each — 128 answers in total, all successful. The questions were six in Korean — move-in report, bulky waste disposal, youth rent subsidy, childbirth support, a 2-day trip, transit fares — and two in English from a foreigner's viewpoint (residence registration, attractions). Every question included the region name. The engine was Gemini 2.5 Flash with Google Search grounding, measured October 20–21, 2026. The design follows the same census approach as the previous post.

Headline Results

Fewer than half the answers used the official domain as evidence, and by source count it was less than one in ten.

What the Evidence Was Made Of

Most evidence came from private sites, blogs, and video. Even where a government holds the information, AI relied elsewhere. That ties back to our earlier point that it may be less about the amount of information than about structure AI can read and carry over.

Differences by Topic and Language

Official citation was higher where governments must provide the guidance themselves, such as benefits and transport, and lower for civil services like move-in reports, where the central government portal and private explainers appeared together. For English questions it was less than half the Korean rate. We had expected tourism to have lower official citation than civil services and were wrong: tourism (37.5%) was slightly above civil services (34.4%). But counting any public domain, tourism was only 40.6% versus 96.9%, meaning public domains were rarely used there at all.

Regional Confusion — Gwangju and a Similarly Named City

The most striking case was Gwangju. Whether we wrote "Gwangju" or the full metropolitan name, 50% of answers cited the city hall of a similarly named city in Gyeonggi Province (gjcity.go.kr)^[1] as evidence. The text described the metropolitan city while the evidence came from a different government. Separately, a move-in report guide from Gunpo city in Gyeonggi was used as evidence in 10 answers across six regions' questions^[2]. This suggests AI search evidence can mix when place names are similar or the guidance is generic. Institutions with homonymous names should make their formal name and regional identifier explicit on their pages.

A Clue for the Next Installments: District Sites

57 of the 128 answers used district or lower-level government sites (.go.kr) rather than the metropolitan government's site as evidence, because the administrative services citizens actually use are guided at the district level. That is the starting point for the series' next topics: Seoul's 25 districts and Gyeonggi's 31 cities and counties.

We Did Not Judge Accuracy

We tried to check the deposit, rent, and income criteria in the Seoul and Daejeon youth rent answers against official notices, but could not obtain the latest official notice, and secondary sources disagreed on the deposit limit (KRW 50 million, 80 million, 100 million). Programs change every year. So this article does not say whether AI answers are right or wrong. It says only which sources AI relies on, and that you must check the latest notice before applying.

Six Improvements

Limits of This Study

This is a one-day snapshot of a single engine, Gemini 2.5 Flash with Google Search grounding. It does not include other channels such as ChatGPT or Naver AI Briefing, and it is a small sample of 8 questions, 8 regions, and 2 repeats. The 8 are the whole of their category but we do not generalize statistically.

Closing Thoughts

Governments hold plenty of information. The question is whether it is structured so that AI uses it as evidence. You can diagnose your institution's structure with AEGIS Insight, and the fastest start is to fix 10 core questions, ask AI directly, and check what it cites. The series continues with the 9 provinces, Seoul's 25 districts, and Gyeonggi's 31 cities and counties.