How a Tax Blunder Pushed Estonia to Build an AI ‘Fuckup Finder’

Estonia’s government turned a costly legislative mistake into a national experiment in AI governance after a tax-law drafting error left online gambling operators outside the tax net for a year. The incident, which cost the state an estimated €24 million annually, helped inspire a new tool designed to catch errors in draft laws before they reach parliament.

That tool, called Apsakaleidja — translated loosely as “Fuckup Finder” — is now at the center of Estonia’s broader push to use artificial intelligence in public administration while keeping humans responsible for final decisions.

How one wrong word exposed a bigger problem

Estonia’s AI moment did not begin with a breakthrough product or a sweeping reform. It began with a legal typo.

In December, lawmakers approved changes to the country’s Gambling Tax Act with the intention of reducing the tax burden on remote gambling. But the wording that emerged referred only to “skill games” for that year, not games of chance or remote gambling. The result was unintended but serious: online casinos were effectively left out of the tax system for an entire year.

The stakes were not trivial. Estonia’s gambling sector is worth roughly €300 million, and its online gaming market is among the fastest-growing in the European Union. By allowing the omission to stand, the government lost about €24 million a year in tax revenue, or about $27.4 million at recent exchange rates.

The error was eventually identified by a legal adviser working for a gambling operator. But by then, the damage was already done — and the embarrassment had become public.

What is Estonia’s AI ‘Fuckup Finder’?

The AI tool is a prototype designed to scan draft bills published on the Riigikogu website and flag potential drafting problems. It looks for broken references, contradictions, arithmetic mistakes, and impossible dates, then assigns each issue a risk level.

Within hours of seeing the gambling tax blunder, former undersecretary for digital transformation Luukas Ilves tested the legislation in large language models, including Claude and Gemini. He said both systems quickly identified the inconsistency, which helped convince him that a more systematic tool could be useful.

From there, he built Apsakaleidja, a web-based checker that can pull draft legislation and highlight likely problems for human review.

What the tool catches

Apsakaleidja is aimed at errors that often slip through conventional drafting and review processes. According to Ilves, it can identify:

• contradictory wording in the same bill
• broken or incorrect cross-references
• mathematical mistakes
• dates that do not make legal sense
• other inconsistencies that may create loopholes or ambiguities

The system does not replace the legislative process. Instead, it serves as an early-warning layer, prompting drafters and lawmakers to inspect suspicious passages before mistakes become law.

How risky are the bills it scanned?

Very few draft bills appear to pass the first AI check cleanly. Of the 112 bills currently listed in the tool, 102 are labeled high risk, suggesting that the system is intentionally tuned to be cautious rather than permissive.

That high volume of flagged issues does not necessarily mean lawmakers are routinely passing broken laws. It does indicate that legal drafting is messy enough that automated review can produce a large number of useful alerts — alongside false positives that still need human judgment.

Why did the mistake become a political opportunity?

Because Estonia saw in the blunder not just a failure, but proof that AI could help public administration work better.

Prime Minister Kristen Michal described the incident as a demonstration of how useful AI can be when it is used as an assistant rather than a substitute for democratic institutions. Speaking to WIRED, he argued that a vibe-coded platform built in response to the tax mistake showed the promise of agentic tools for both citizens and civil society.

Michal said the episode showed that AI can be a powerful helper, and that the prototype built after the tax error illustrated how agentic systems can empower people outside government as well as within it.

The government’s response has been unusually swift. Rather than treating the incident as an isolated embarrassment, officials appear to have used it as momentum for a wider modernization agenda.

Estonia’s broader AI push

In January, Michal signaled that the country could use tools like Apsakaleidja not just to catch mistakes, but to help draft laws and identify loopholes before they are enacted. Soon after, he launched Eesti.ai, a national program intended to build AI skills across the population and raise productivity over the coming decade.

The government’s ambition is broad: double national productivity by 2035, in part by making AI a routine part of public and private work. Among the advisers to the initiative are Markus Villig, founder of Bolt, and Luukas Ilves, the architect of the Fuckup Finder prototype.

In April, the government sent parliament a bill that would give national and local authorities the right to use digital systems, including AI, to automate administrative processes. That legislation is still moving through the Riigikogu, where lawmakers are debating how far automation should go and where human review must remain mandatory.

By June, Michal was taking the idea a step further. At an Eesti.ai meeting, he said that if the plan advances as intended, Estonia could become the first country to create official digital identities for AI agents.

What does an AI digital identity mean?

In Estonia’s vision, an AI agent would not merely be a chatbot or assistant sitting on the sidelines. It would be a digital actor with a defined identity, authorization, and scope of responsibility when dealing with public services, banks, registries, and other systems.

That matters because digital systems need to know who or what is making a request, under whose authority, and with what permissions. Without that clarity, the use of AI in official processes can blur accountability and create risks that are hard to trace.

How Estonia plans to keep humans in control

Estonian officials insist that automation will not be used where judgment, discretion, or competing human interests are involved. The government’s current approach separates routine, rule-based decisions from cases requiring evaluation and context.

According to Kirke Maar, who leads the Eesti.ai team, automation is appropriate when the law already defines the outcome and the state already holds the necessary facts. In those situations, a person may not even need to file a form. If the data shows they qualify, the system could simply notify them.

That logic already shapes parts of Estonia’s digital state. Tax declarations, for example, are prefilled. The next step could be more ambitious: an AI agent preparing and filing more complex returns, while the citizen reviews, confirms, or intervenes if needed.

When must a human decide?

A human must stay involved from the beginning when a decision requires balancing interests or making a judgment about a specific person’s circumstances. Estonia’s officials say that is the dividing line between administrative automation and decisions that require discretion.

They also say that if a person invokes their right to be heard during an automated process, the AI procedure must stop and a human official must take over. The same applies if the citizen challenges the decision.

Every automated administrative action would also need to leave a clear audit trail. That record should show what information the system used, which rule it applied, when it made the decision, and how the citizen can challenge or correct the outcome.

Maar said Estonia’s aim is not to eliminate people from government, but to make public services more accessible, quicker, and less burdensome.

Why Estonia believes it can move faster than other countries

Estonia is unusually well positioned to adopt AI because it already runs as a digital-first state. The country says 99 percent of public services are online, and decades of investment in digital identity and state infrastructure have made online interaction with government normal rather than exceptional.

That digital foundation means new AI tools do not have to replace an analogue bureaucracy from scratch. They can plug into a system that already expects citizens to interact with the state through secure digital channels.

Michal argues those earlier investments now create a practical advantage. Estonia can move more quickly into the AI era because the state has already built the basic infrastructure required for identity, authentication, and online service delivery.

What are the risks of AI in government?

The main risk is not that AI will make decisions, but that it will do so without clear responsibility.

Liina Vahtras, managing director of Estonia’s e-residency program, said the danger appears when systems begin acting at scale without accountability, especially if no one can easily identify the responsible party or determine whether permissions were properly granted.

Vahtras said Estonia is trying to prevent situations in which AI acts on behalf of people or companies while the chain of responsibility becomes unclear.

That concern grows as AI agents increasingly connect to public services, banks, registries, and other digital systems. Estonia’s planned “agent code” concept is meant to make those relationships visible, so it is always clear who owns the agent, what it is allowed to do, and who remains liable for the result.

The worry extends beyond technical errors. If AI can file forms, move data, or initiate processes, then mistakes, misuse, or unauthorized actions could spread quickly unless the system is designed with firm boundaries and traceability.

What experts say about the human review problem

Not everyone is convinced that AI should be the answer to legislative quality control, even if the technology is useful.

Catherine Flick, a technology ethics researcher at the University of Staffordshire, says the gambling tax error raises a more basic question about legislative process: why humans are not already required to perform thorough review before laws are passed.

Flick argued that someone still has to read the full text with the right contextual understanding to make sure the law actually works as intended.

Her point highlights the limits of automation. AI can catch internal inconsistencies, but it does not absolve lawmakers of the duty to read, understand, and test legislation. The machine may spot a typo, but it cannot alone determine whether a policy is wise, fair, or politically coherent.

How the Estonia case fits the global AI government debate

Estonia’s experiment arrives at a moment when governments around the world are struggling to define how much authority AI should be given.

Public-sector AI offers obvious efficiencies: faster processing, fewer manual forms, improved consistency, and lower administrative burdens. But the same systems raise hard questions about transparency, due process, appeals, bias, and legal responsibility.

Estonia’s approach is notable because it frames AI not as a replacement for government, but as an operational layer inside a digital state. That makes the country a useful test case for a broader idea: AI as a bureaucratic assistant that can speed up routine work without taking away democratic oversight.

The gambling tax error also illustrates a lesson many governments learn too late. Administrative systems are only as robust as the language that governs them. A single imprecise clause can have major fiscal consequences.

In that sense, Estonia’s “Fuckup Finder” may be less a novelty than a warning system for modern lawmaking. As legislation becomes more complex and digital services become more automated, governments may need machine assistance simply to keep up with their own procedural complexity.

Timeline: how a tax error turned into an AI policy shift

What happens next?

For now, Estonia’s AI “Fuckup Finder” is still a prototype, but the political direction is clear. The country is testing how far it can go in automating public administration while preserving accountability, appeal rights, and democratic control.

The legislation now before parliament will help determine whether AI becomes a limited back-office aid or a more central part of Estonia’s administrative machinery. The coming debate will likely focus on the same tension that has defined the whole episode: how to use AI to improve state efficiency without allowing it to make decisions beyond its mandate.

What makes Estonia’s case unusual is that the experiment was not launched by a tech company looking for a market. It emerged from a public-sector failure — one costly enough to sting, but also small and specific enough to inspire a fix.

That may be why the idea resonates. In Estonia, the route to AI adoption did not begin with hype. It began with a law that said the wrong thing.

Michal emphasized that AI does not override parliament, courts, or the will of voters. If it finds a mistake, he said, it is no different from a human spotting one — the responsibility to fix it still belongs to democratic institutions.