The $28 Million Mistake That Inspired Estonia’s AI ‘Fuckup Finder’

Estonia’s AI embarrassment began with a single wrong phrase.

In December, the Riigikogu, Estonia’s parliament, passed changes to the country’s Gambling Tax Act meant to lower the tax rate on remote gambling. But the wording of the law referred only to “skill games” for that year, not games of chance or remote gambling. Estonia’s entire gambling industry is worth around €300 million ($343 million), and its online gambling market is one of the fastest growing in the EU.

That single blunder meant online casinos were accidentally left outside the tax net for an entire year, losing the government €24 million ($27.4 million) a year in gambling revenues.

The error was spotted by a legal counsel for a gambling operator. But the embarrassment deepened when Luukas Ilves, former undersecretary for digital transformation, ran the legislation through Claude and Gemini. Both AI systems, Ilves said, immediately identified the inconsistency.

Within hours, Ilves had built a prototype tool—called Apsakaleidja, or “Fuckup Finder”—that could pull draft bills from the Riigikogu website and flag problems such as broken references, contradictory wording, arithmetic errors, and impossible dates. It categorizes the problems as high, medium, or low risk—of the 112 bills currently listed, 102 are rated high risk. One example highlighted on the Fuckup Finder suggests contradictory wording in draft text. Ilves even demo’d it on national TV, to the host’s amazement.

The blunder was embarrassing—but also triggered a revelation within the government. “The situation demonstrated that AI can be an incredibly useful assistant,” Kristen Michal, Estonia’s prime minister, told WIRED. “And—in the form of a vibe-coded platform to check draft legislation created in response to the incident—we saw an example of how agentic tools can empower civil society and individual citizens.”

So Estonia doubled down on the use of AI in government. In January, Michal suggested the country may well use tools like the Apsakaleidja to draft legislation to preemptively find and fix loopholes. He launched the Eesti.ai program designed to skill up Estonians in AI use, with the goal of doubling productivity within the country by 2035. Among the advisers to the Eesti.ai initiative are Bolt founder Markus Villig and Ilves, creator of the Fuckup Finder.

Then, in April, the country’s parliament received from the government a bill that gave state and local government the right to use digital solutions, including AI, to automate administrative processes. That bill is going through parliament, which is debating the changes it could introduce, with the intention of becoming law. In June, Michal told an Eesti.ai meeting that, if things go to plan, "Estonia will become the first country in the world to create official digital identities for AI agents.”

“This is a new environment for the public sector,” Michal told WIRED. “It demands agility and the ability to adapt as technology changes.” Estonia is better placed than many countries to adapt to those changes: It has led the way on integrating digital identity thanks to a digital-first state, while 99 percent of public services are already online, Michal says. Estonia is held up as an example of how to run a modern digital state—as WIRED first covered a decade ago. That laid the groundwork for easier AI adoption. “Those investments now allow us to move faster and more confidently into the AI era,” he explains.

Catherine Flick, who researches technology ethics at the University of Staffordshire, says the gambling tax error raises a more basic question: “Why are humans not doing this review process as part of the legislation drafting procedure?” she says. “At some point someone has to sit down and read through the whole thing, with the understanding of the context and all that sort of stuff, in order to make sure that this is a decent law.”

It’s that final element—the human in the loop—that is is being debated by representatives in the Riigikogu. The bill is deliberately drawn widely, explains Kirke Maar, team lead of Eesti.ai.

Estonia’s current thinking is that choices can be broadly divided into two buckets. “The natural dividing line is between rule-bound decisions and discretionary ones,” she says. “Where the law determines the outcome from verifiable facts—you meet the criteria, you get the result—automation is appropriate.” If the state already has the data needed to establish that someone qualifies for a benefit, Maar and Ilves argue, the person should not necessarily have to fill in a form at all: They could simply be told they qualify. Tax declarations, which in Estonia are already prefilled, could move from citizens checking and confirming a form to an agent preparing and filing more complex declarations end to end, with the citizen confirming or intervening where needed.

When things are more complicated, and “a decision requires genuinely weighing competing interests or judgment about a person's specific circumstances, a human belongs in the loop from the start,” says Maar.

Maar and Ilves say at any point during an AI decisionmaking process, a person would be able to invoke their right to be heard, at which point the automated procedure would end and a human official would take over. Automated decisions would also be ruled out where a citizen disputes a decision. And every automated administrative decision would also have to leave an audit trail of what data was used, which rule was applied, when the decision was made, and how the citizen could challenge or correct it.

“The purpose of Estonia’s digital state has never been to remove the human from government,” Marr says. “It has been to make services more accessible, faster and less burdensome.”

Accountability is also a central part of the AI process, and an area that those who work with the government on other digital interactions with citizens know well. “The key risk is AI systems acting at scale without accountability—where actions cannot be traced back to a responsible party, permissions are unclear, or misuse goes undetected,” says Liina Vahtras, managing director of e-residency, Estonia’s program that gives nonresidents a government-issued digital identity to access Estonian business services online. “This is what we’re working to prevent.”

“As AI agents begin to interact with public services, banks, registers, and other digital systems on behalf of people and companies, it must be clear who the agent belongs to, under whose authorization it acts, what it is allowed to do, and who remains responsible for its actions,” Vahtras says. “The idea behind the agent code is to make the chain of responsibility visible.”

It’s also something the prime minister knows and worries about, too. He’s careful to draw a line between AI as an assistant in Estonia and AI as an authority. “AI does not replace democratic institutions, the constitution, or the will of voters,” he says. “If AI identifies a mistake in legislation, it is no different from a human spotting one. The responsibility to correct it remains with parliament, the courts, or the public administration.”

This reporting was supported by a grant from Tarbell Center for AI Journalism.