Apsakaleidja — Estonia's AI Tool for Catching Legislative Drafting Errors Before They Become Law

In January 2026, a routine wording slip in an amendment to Estonia's Gambling Tax Act omitted the word for ‘games of chance’ (õnnemängud), leaving only ‘games of skill’ (osavusmängud) covered by a tax increase. The result: from 2026, online casinos in Estonia would owe no gambling tax at all — an error the state estimated would cost roughly €2 million a month in lost revenue if it went uncorrected before the law took effect.

Luukas Ilves, a former deputy secretary general for digital development at Estonia's Ministry of Economic Affairs and Communications and now an advisor on Ukraine's digital transformation, built a prototype in a matter of hours to see whether AI could have caught the mistake. His tool, Apsakaleidja (‘mistake-finder’), pulls every draft bill currently before the Riigikogu, feeds it to large language models, and flags inconsistent use of defined terms, broken cross-references, contradictions between a bill's explanatory memorandum and its legal text, arithmetic errors and impossible dates. When Ilves ran the gambling tax amendment through Claude and Gemini, both models immediately identified the missing word.

Applied to the 112 bills before parliament at the time, Apsakaleidja flagged 102 as high-risk for this class of error — a strikingly high hit rate that likely reflects a deliberately low threshold for flagging (favouring false positives over missed errors) rather than a claim that 91% of draft legislation is broken. Estonian Prime Minister Kristen Michal publicly urged the Riigikogu to start using AI tools like this to check bills, and Riigikogu officials told Ilves they had already used the prototype to catch and fix new drafting slips, inspired by the publicity.

Apsakaleidja is not (yet) an official parliamentary system: it is an independent civic-tech prototype built and run outside any formal procurement or oversight process, with no published accuracy evaluation, error-rate audit, or funding commitment behind it. It is included here as an early but well-documented example of AI catching a real, costly legislative drafting error — and as a governance case study in what it would take to move from a one-person prototype to an accountable, sustained parliamentary tool.

Read the full analysis: https://news.err.ee/1609910593/estonian-prime-minister-recommends-parliament-use-ai-to-check-bills