Estonia tests an AI to catch errors in draft laws
A beta prototype analyses legislative texts to flag broken references, contradictions and impossible dates before a bill is approved.
On January 14, 2026, Estonia unveiled a tool for a very specific application of artificial intelligence: reviewing draft laws before they are passed. What opened the debate was a real drafting error. According to public broadcaster ERR, a clerical mistake in the Gambling Tax Act inadvertently left online casinos exempt from tax in 2026 — an oversight no one in the review chain caught in time, which reopened a simple question: can a machine help find flaws a human chain did not see? The capability you take from this case is not "AI already writes laws," but how a tool that assists without deciding is designed, and why that matters as much as what the tool detects.
What the tool is — and is not
The answer being tested is not an AI that drafts or decides rules. Luukas Kristjan Ilves, formerly Estonia's chief of digital development, presented a beta-stage prototype called Apsakaleidja — literally, "mistake finder." Its job is to read the draft bills in progress on the Riigikogu (the Estonian parliament) website and flag the elements that deserve a second look. Ilves was sharp about what the tax episode reveals: in his view, the only conclusion possible is that "no one at the Ministry of Finance or in the Riigikogu bothered to conduct any sort of review."
Per the description reported by ERR, Apsakaleidja flags five classes of anomaly: inconsistent use of defined terms, broken references to other sections, contradictions between the legal text and its explanatory memorandum, arithmetic errors, and impossible dates. All share a trait that makes them tractable for a machine: they are problems of consistency, not of judgment. They can be framed as automatic checks over a public document — is this term used the same way throughout? does this reference point to a section that exists? does this sum add up? — without the system having to opine on the substance.
That delimits its role well, and Ilves himself frames it as an improvable beta. The system does not determine whether a policy is just, constitutional, or advisable; nor does it decide what economic consequences a country should accept. What it tries to do is draw attention to anomalies that could stay hidden among many changes, cross-references, and deadlines. Detecting an inconsistency does not amount to proving a legal error exists: an alert then needs review by people who understand the legislative context, the text's intent, and its relation to laws in force.
There is a reason this experiment is born in Estonia and not just anywhere, and it is instructive. The tool can read draft bills because they are published, in accessible format, on the parliament's website: the Estonian state has spent decades building a digital administration, and that base is what lets a program walk through a legal text the way one walks through a spreadsheet. It is not just that Estonia wants to use AI; it is that its legislative process is already machine-readable enough for an AI to have something to review. Where laws live in scanned PDFs or closed offices, the same idea does not start.
An extra check, not an authority
Estonia's public debate has carefully kept that boundary. In the project's coverage, Prime Minister Kristen Michal recommended parliament try the tool with a phrase that sums up the right attitude: "If something like this were put to use, we'd be better again. And there's something to say about learning from your mistakes." Justice Minister Liisa Pakosta backed the use of these tools in drafting and reviewing laws, and former Chancellor of Justice Allar Jõks supported it cautiously, as a support tool. No one of institutional weight proposed that the machine replace human judgment: if an AI identifies a flaw, the responsibility to fix it still belongs to parliament, the courts, or the administration.
That idea connects with another strand of the Estonian government's work worth reading alongside this one, because they share a principle. On June 17, 2026, at its second meeting, the Eesti.ai advisory board — launched on Michal's initiative — approved moving toward digital identities for AI agents, "AI ID codes" that would let a system act on behalf of a person, company, or organization within defined limits. The point of the proposal is that it does not force granting an assistant blanket access to all rights and data: specific permissions can be set — view only, prepare documents, authorize payments, or act within a financial cap — and every operation must be verifiable and auditable.
Michal placed it in the country's digital tradition: "Digital identities, the X-Road, digital signatures and footprints have made our country faster, simpler and more secure," and applied the same standard to AI: "it must be clear who is acting on whose behalf with what rights." Applied to legislative review, that principle translates into a useful rule: a tool can assist the process, but it must leave a trace of what it analyzed, what it flagged, and how each alert was resolved. Traceability lets you discuss the system's own errors without dissolving institutional responsibility between "the AI said so" and "I just looked at the screen." Estonia's two initiatives — the mistake finder and the identity codes for agents — are, at bottom, the same philosophy applied twice: define precisely what a system may do, keep every action logged and auditable, and keep a person accountable at the end of the chain. A country that demands that of an agent authorizing payments demands it too of an agent reviewing laws.
The challenge is not only technical
Estonia's parliament has acknowledged that some draft bills still have quality gaps, among them insufficient constitutionality analysis. A consistency tool could free time for those more complex reviews, but it does not replace them — and there lies the nuance that separates sensible use from an inflated promise. Apsakaleidja can also produce false positives, fail to understand a lawmaker's deliberate exception, or overlook a deep problem written with perfect coherence. A text can be flawless in form and terrible in policy; no consistency check catches that.
The interest of the Estonian case, then, is not in promising an automatic legislator, but in showing a very concrete administrative task where AI can be useful: finding sooner the details worth looking at again, within a flow with experts and a clear human decision. The lesson travels beyond Estonia and beyond law: faced with any tool that "reviews" — contracts, accounts, code, rules — the right questions are not whether it is right, but what kind of flaw it detects, which alerts are its own and which the human's, and who answers for the final decision. In lawmaking, detecting a signal is valuable; deciding what to do with it remains a democratic responsibility.
Sources
Sources for this piece
This piece draws on 4 primary source(s), gathered during reporting.
This article was produced with artificial intelligence under human editorial oversight.
