Finnish prisoners are training AI models for $1.67 an hour

Inside a windowless room at Hämeenlinna prison in Finland, an inmate serving a six-year sentence clicks through real estate snippets on an HP laptop. She earns €1.54 ($1.67) an hour answering yes-or-no questions about construction permits, wondering if she is helping build a customer support bot.

She isn’t. She is labeling data to train a large language model for Metroc, a Finnish construction intelligence startup that recently raised €2 million to expand across the Nordics. And her workplace is one of three high-security prisons in Finland turning to data annotation.

The small-language bottleneck

The global AI boom relies heavily on clickworkers in Kenya, Uganda, or India to scrub toxic prompts and categorize data. But while that pipeline works wonders for English, it falls apart for languages like Finnish, which has only 5 million native speakers worldwide.

In Finland, high wages and robust social safety nets mean virtually no one signs up for micro-task platforms like Amazon Mechanical Turk. Automatic translation tools are still too clunky to parse complex local municipality filings, leaving Metroc founder Jussi Virnala with an acute shortage of Finnish-speaking annotators until he pitched the country’s prison system.

Finland’s criminal sanctions agency welcomed the initiative as part of its Smart Prison project, which aims to bring digital reality behind bars. Instead of sewing or making road signs for roughly €6 per six-hour shift, inmates can volunteer for three-hour laptop shifts paid by the hour with no output quotas.

Digital rehab or technological exploitation?

Prison officials argue that data work modernizes rehabilitation, exposing inmates to the kind of tech-driven jobs they might encounter after release. Prior to the pandemic, another Finnish startup named Vainu ran a similar trial before a founder dispute halted the effort.

Human rights advocates, however, see an unsettling pattern. Amos Toh, a senior AI researcher at Human Rights Watch, warns that the industry’s quest for dirt-cheap labor increasingly preys on captive and vulnerable populations, moving from economic crisis zones and refugees to incarcerated workers with virtually zero bargaining power.

Critics also point out the obvious skill gap: repeatedly clicking yes or no on paragraphs teaches virtually zero marketable tech skills, leaving little evidence that it actually improves post-release employment.

Why it matters

We think this Finnish experiment exposes an uncomfortable truth about generative AI: off-the-shelf automation still leans heavily on human drudgery, and the problem gets magnified outside global lingua francas.

For enterprise startups tackling localized, highly regulated niches like Nordic construction permits, cheap labor from the global south is not an option. Metroc’s pivot to prisons looks clever from a pure unit-economics perspective, but it relies on an institutional arbitrage that is ethically fragile and impossible to scale.

Packaging repetitive, binary text categorization as modern job training feels disingenuous. Calling it high-tech doesn't change the underlying reality: it is still digital sweatshop labor, just housed in spotless Nordic facilities.