evidoria

← Back to browse

Good practice Imported

Preserving Icelandic Through GPT-4 — Iceland's Presidential AI Language Partnership

Iceland · Reykjavík · See the Iceland profile · See the Reykjavík profile

Evidence: Descriptive / self-reported Top 62% 53/100 · Ask Evidence Copilot about this practice

On President Jóhannesson's initiative, Iceland partnered with OpenAI so GPT-4 could learn proper Icelandic. Reykjavík firm Miðeind used 40 volunteers to fine-tune the model via human feedback, aiming to keep the language viable in AI-driven public services.

370,000 speakers
Icelandic speakers
40 volunteers
Volunteers applying RLHF

Details

Promoter
Miðeind / Government of Iceland
Period
2022-2023
Keywords
language technology, digital government, cultural policy

Context

Iceland, a small Nordic nation of roughly 370,000 speakers, risks seeing its language marginalised as digital public services increasingly rely on large language models trained mostly in English.

Objectives

The project aimed to improve GPT-4's Icelandic performance so AI-driven public services remain viable in Icelandic, and to create a template for other low-resource languages.

Activities

Following a spring 2022 meeting between President Guðni Th. Jóhannesson and OpenAI's Sam Altman, Miðeind assembled 40 volunteers who applied Reinforcement Learning from Human Feedback, selecting and refining the best of four model-generated answers per prompt; the work was announced in March 2023.

Results

The resulting model now underpins Embla, Miðeind's Icelandic voice assistant, and the collaboration was explicitly framed as a template for other low-resource languages.

Conclusions

No source publishes independent measures of accuracy gains, Embla adoption or citizen impact, and the effort relied on unpaid volunteer labour and a single private partner rather than a standing government programme.

Implementation
Indicative cost
Low (< €50k)
Time to results
Short (< 1 year)
Staffing & skills
President Guðni Th. Jóhannesson's office, Miðeind (led by Vilhjálmur Þorsteinsson), 40 unpaid volunteers, OpenAI

Conditions for success

  • a head-of-state-level relationship opening direct access to OpenAI leadership
  • a structured RLHF workflow (four candidate answers per prompt, volunteer selection/refinement)

Common failure modes

  • relied on unpaid volunteer labour and a single private partner rather than a standing government programme
  • no independent measures of accuracy gains, Embla adoption or citizen impact have been published

Commonly funded by

National / regional programmes Philanthropic / foundation funding

Indicative funding routes for practices of this type — always check each programme's current calls and eligibility rules.

Do you run this practice? Claim it — verified implementers get a public contact pathway and can propose corrections.

Data sources

Where this practice's information was retrieved from, and when.

Similar practices you may find useful