Iceland enlists GPT-4 to save its language from digital extinction

It’s a strange kind of irony. Iceland, an island of roughly 370,000 people with a booming tech scene, is seeing its ancient mother tongue—so close to Old Norse it feels like a direct line to the Vikings—get crowded out by the sheer dominance of English online. The government isn’t just standing by. They’ve been coining hyper-specific Icelandic words for new tech for years, like tölva for computer. But that defensive crouch has now turned into an offensive play, spearheaded by the country’s president and a direct partnership with OpenAI. They’re betting that the very force threatening the language, artificial intelligence, might be its best shot at survival.

The initial results were comically bad. Asked for Donald Duck’s Icelandic name, GPT-3 coughed up a weirdly literal translation, while an earlier version of ChatGPT just insisted the name was the same in both languages. GPT-4 finally nailed it: Andrés Önd. That leap in cultural knowledge didn’t happen by accident. Vilhjálmur Þorsteinsson, CEO of Miðeind ehf, threw out the old playbook after a labor-intensive attempt to fine-tune GPT-3 with 300,000 examples flopped spectacularly. “The capability to generate grammatical or correct Icelandic just wasn’t there with GPT-3,” he said. His team of 40 volunteers pivoted to Reinforcement Learning from Human Feedback (RLHF), teaching the model proper grammar and local trivia with a shockingly small dataset.

Here’s the part that made my ears perk up: RLHF started producing real results with just 100 examples. Let that sink in. The brute-force method required a mountain of data and fell on its face. This new approach, where humans pick the best of four AI-generated answers and polish it into an ideal response, is lean and wildly effective. For language preservationists watching this from places like Wales or the Basque Country, that tiny data requirement is the whole story. It tears down the barrier that keeps massive language models from being useful for “long-tail” languages with limited digital footprints.

Jóhanna Vigdís Guðmundsdóttir, CEO of the language tech center Almannarómur, framed this not as a niche cultural project but as a way to avoid a looming “AI divide.” Her logic is sharp: it’s not enough to build a walled garden of Icelandic-language apps. If the tools everyone actually uses every day don’t understand you, your language slowly becomes a home-bound relic. The endgame isn’t just a chatbot that knows Icelandic folklore. It’s an economic argument—ensuring a language survives by making it a functional, default part of a digital life, not a museum piece.