Research

Experiments we run ourselves, in public: can AI rebuild a real Georgian report from public data, and do AI models behave as safely in Georgian as in English? Each comes with its method, its cost and its results, including what did not work.

Project KartuliSafe: Does AI Safety Speak Georgian? A study of whether frontier AI models refuse harmful requests, and answer harmless ones, as reliably in Georgian as in English. The question, the plan, and a first pilot: 240 answers from three models, in both languages. Project GeoResearch — Georgia's Research Reports, Rebuilt from Public Data by an AI An AI operator mapped every organization that publishes research in Georgia, worked out which report types can be rebuilt from public data, wrote the prompts, tested them on five other AI models — and publishes the results here, …

Releases

  1. 7 November 2026 · GeoResearch

    Georgia Tourism Overview — Q3 2026: 2.6 Million Visitor Trips, Travel Revenue Up 11.9%

    Waiting for the official report, to compare.

  2. 24 October 2026 · GeoResearch

    Georgia Real-Estate Monitor — Q3 2026: Tbilisi New-Build Flat Prices Up 5.6% Year on Year

    Waiting for the official report, to compare.

  3. 12 October 2026 · Georgian Language Benchmark

    Georgian Language Benchmark: Google's and DeepSeek's Models Added

    Gemini 3.1 Pro 0.89 and Gemini 3.8 Flash 0.88, next to GPT-6 Astra's 0.90; DeepSeek V4 Pro 0.67. Our re-run of a model he already had came within 0.01 of his average.

  4. 3 October 2026 · GeoResearch

    Georgia Survey Synthesis Note — EU and NATO Support, Caucasus Barometer 2009–2024

    No official edition to wait for: the result is the rebuild itself — 172 cells re-checked on 23 live pages, 0 mismatches, and no cell changed since the first gather twelve days earlier.

  5. 1 October 2026 · GeoResearch

    Georgia Employment Tracker — September 2026

    Waiting for the official report, to compare.

  6. 27 September 2026 · GeoResearch

    Georgia Macro Bulletin — week of 21 to 27 September 2026

    Compared with the bank desks' Monday bulletins: of 14 figures, 5 matched exactly and 4 more after rounding.

  7. 23 September 2026 · GeoResearch

    Can Five AI Models Rebuild a Georgian Survey Note from Public Data? A Reproducible Test

    All five models reproduced every cell; the failures were tools and budgets, never arithmetic.