Estonia’s latest AI fever broke because the government fucked up. Literally.
It started in December. The Riigikogu—Parliament for the uninitiated—passed a tweak to the Gambling Tax Act. They wanted to lower the tax on remote gambling. Simple enough.
The problem was the text.
The law referenced “skill games” for that specific year. It skipped games of chance. It ignored remote gambling entirely. Given that the sector is worth roughly €300 million, that’s not a typo. It’s a financial hemorrhage. Online casinos got left outside the net. The state lost €24 million. Just like that.
Someone spotted it. A lawyer for a gambling operator caught the glitch.
Luukas Ilves didn’t. Ilves was just checking the vibes.
He ran the bill through Claude. He ran it through Gemini. Both flagged the inconsistency immediately. No lag. No coffee breaks.
Within hours, Ilves built Apsakaleidja. Which translates to “Fuckup Finder.”
It’s crude. It works.
The prototype scrapes draft bills from the parliamentary site. It flags broken references. Arithmetic errors. Dates that don’t exist. It grades them. High, medium, or low risk.
“Of the 112 bills listed, 102 were rated high risk.”
He showed it on TV. The host looked shocked. The country looked embarrassed. But also curious.
Prime Minister Kristen Michal didn’t see a failure. He saw a hack.
“The situation demonstrated that AI can be incredibly useful,” Michal told WIRED. “We saw how agentic tools empower citizens.”
So they doubled down. Hard.
In January, Michal proposed using Apsakaleidja-like tools to draft laws before they leave the room. To catch loops before they bite. He launched Eesti.ai. The goal is doubling productivity by 2035. Bolt founder Markus Villig joined the advisory board. Ilves stayed on.
By April, the parliament was debating a new bill. This one allows state bodies to use AI to automate admin work.
In June, Michal went further.
He suggested Estonia might give official digital identities to AI agents. Not people. Code.
“Estonia will become the first in the world,” he said.
Why here?
Estonia already runs online. Ninety-nine percent of public services are digital. Digital IDs are standard. WIRED praised the setup ten years ago. It paved the road.
“Those investments let us move faster into the AI era,” Michal argues.
But not everyone is sold on the speed.
Catherine Flick at the University of Staffordshire points out a boring truth. Humans should have caught that error.
“Why are humans not doing the review process?” she asks.
She’s right. Someone has to read the whole thing. Someone has to understand context. A machine checks syntax. A human checks sense.
So what happens now?
The current draft bill draws a line in the sand. It splits decisions into two buckets.
- Rule-bound outcomes. Verifiable facts. You meet the criteria? You get the money.
- Discretionary judgment. Complex circumstances. Competing interests.
If the data says you qualify for benefits, you don’t fill a form. The agent files it. Tax declarations are already pre-filled in Estonia. Imagine an agent filing them for you.
But when things get messy? A human steps in.
Kirke Maar, leading Eesti.ai, explains the logic. Where judgment matters, “a human belongs in the loop from start.”
You can invoke the right to be heard at any time. The bot stops. A person takes over. If you dispute the decision? Human review is mandatory.
And there’s an audit trail.
Every automated decision leaves a footprint. What data was used? Which rule applied? When was it decided? How can you fight it?
“The purpose was never to remove the human,” Maar insists. “It was to make services less burdensome.”
But burden shifts. It doesn’t vanish.
Liina Vahtras runs e-residency. She sees the danger clearly. AI acting at scale is great. Until it goes wrong. And you can’t trace it back to anyone.
“The key risk is lack of accountability,” she warns.
Permissions blur. Misuse hides.
“The chain of responsibility must be visible,” she says.
When an AI agent talks to a bank, it needs to know who owns it. Who authorized it. What it can touch. And who gets the blame.
Michal agrees. He’s wary of handing the wheel to silicon.
“AI does not replace the constitution,” he says firmly.
It’s a tool. Like a highlighter.
If it finds a mistake in the law? Great. Fix it. Parliament does that. Courts do that. Not the code.
The $28 million mistake was costly.
But maybe it bought them time to build something that watches the watchers. Or just watches the spreadsheets.
