AI Language Preservation: Lessons from Iceland for U.S. Services

A country of about 370,000 people built a plan to keep its national language from quietly fading out of daily digital life. That plan didn’t start with a big marketing campaign or a new dictionary app. It started with a blunt reality: if your language isn’t supported by the software people use every day—phones, chatbots, search, voice assistants—then your citizens will default to the languages that are.

Iceland’s collaboration with OpenAI to improve Icelandic performance in GPT‑4 is more than a feel-good cultural story. It’s a practical case study in AI in government and public sector digital services: how a government-adjacent coalition (public leadership, a nonprofit language tech center, and private firms) can strengthen access, equity, and service quality by investing in language infrastructure.

For U.S. agencies and vendors building AI-powered digital services, the Icelandic effort offers a clear message: language support isn’t a “nice-to-have” feature—it’s a service delivery requirement. If your chatbot, call center automation, forms assistant, or benefits navigator only works well in English, you’ve built a digital system that quietly excludes.

The real risk: “digital extinction” happens inside apps

Digital extinction isn’t about a language disappearing from homes first—it’s about it disappearing from screens. When the default interface language for banking, travel, healthcare portals, and government services is English, smaller languages (or minority languages within large countries) get pushed to the edges.

Iceland has been proactive for decades. Its language planning approach intentionally coins new Icelandic terms rather than importing “loanwords.” (A computer being called tölva—roughly “number prophetess”—is the kind of detail that makes the policy feel alive.) But vocabulary planning alone can’t keep pace with modern digitalization.

The pressure point is simple: people live inside products.

• If voice assistants don’t understand you, you switch languages.
• If customer support automation responds poorly, you rephrase in English.
• If school tools, productivity software, and AI writing assistants don’t work in your language, students adapt early—and that habit sticks.

Government digital services sit directly in the blast radius of this problem. Anything from unemployment insurance portals to state DMV chat support can become an unintentional “language filter.”

Why large language models struggle with smaller languages

Most large language models are trained on what the internet provides at scale, and that heavily favors English and other major languages. That reality isn’t political; it’s math. If there’s less high-quality Icelandic text online, the model has fewer examples to learn grammar, idioms, and cultural references.

The OpenAI/Iceland example makes the limitation concrete. Earlier models produced confident but wrong answers to basic cultural questions (like what Donald Duck is called in Icelandic). GPT‑4 improved accuracy, but still produced issues common to low-resource language output:

• Grammar errors (agreement, declensions, tense)
• “Translationese” (sentences that feel like English wearing Icelandic clothing)
• Cultural mismatches (answers that ignore local context)

Here’s the part U.S. public-sector teams should pay attention to: even when the language is “supported,” service quality can still be unacceptable. A benefits chatbot that answers in Spanish but uses awkward phrasing, incorrect formal register, or culturally wrong assumptions will still lose trust.

A quick myth-bust: “Translation fixes it”

A lot of digital transformation programs treat translation as the finish line. It’s not.

Service language quality means:

1. The answer is correct.
2. The language is natural and readable.
3. The system understands local context (institutions, forms, norms).
4. The system stays consistent across channels (web, voice, SMS).

Translation is only one component.

The Iceland approach: RLHF as language infrastructure

Iceland’s most useful technical lesson is how they improved output with Reinforcement Learning from Human Feedback (RLHF). Instead of trying to build an Icelandic model from scratch, or relying on huge new datasets, they used a workflow where human testers:

1. Provide a prompt
2. Review multiple candidate responses
3. Choose the best
4. Edit it into an “ideal” answer

Those preference and correction signals are then used to improve the model.

Two details matter for governments:

1) You can get meaningful improvement without massive datasets

The source story contrasts older fine-tuning efforts (hundreds of thousands of examples with disappointing results on GPT‑3) with the practicality of improving GPT‑4 behavior with as few as ~100 examples in an RLHF cycle.

That doesn’t mean “100 examples solves it.” It means the barrier to piloting is far lower than many agencies assume. You can start small, measure improvement, and then scale.

2) The “human feedback” should be institutional knowledge

Iceland’s volunteers corrected grammar and culture. In the U.S. public sector, the equivalent is often:

• program eligibility rules
• state-by-state terminology
• how an agency actually communicates (tone, formality)
• common misunderstanding patterns
• what you’re legally allowed to say (and what you must say)

RLHF isn’t just language polish. It’s policy accuracy training.

A useful way to frame RLHF for government: it turns your best frontline staff and policy experts into training signal, not just QA reviewers.

Cultural context is a product requirement, not a bonus

Language and context are inseparable. The Iceland example shows GPT‑4 responding differently depending on whether the prompt is in Icelandic or English—because the model infers different implied context.

That same effect shows up in U.S. digital services every day:

• “How many representatives are there?” means different things depending on whether someone is asking about Congress, a state legislature, or a city council.
• “Who is the president?” could mean the U.S. President, a university president, or a union president—context decides.
• “How do I renew my ID?” depends on whether the user means a driver’s license, state ID, professional license, or immigration document.

For public-sector AI assistants, the right design stance is strict:

If context is ambiguous, your system should ask a clarifying question—not guess.

That’s one reason language preservation and public-sector AI overlap so naturally. Both require building systems that respect local institutions, not just words.

What “good” looks like for AI in government digital services

If you’re deploying multilingual AI chatbots or voice assistants for public services, aim for outcomes that users can feel:

• Higher first-contact resolution (fewer handoffs to live agents)
• Lower abandonment on forms and portals
• Consistent terminology across translated pages, chat, and call center scripts
• Measurable accuracy on program rules, not just “fluency”

What U.S. agencies and vendors can copy from Iceland (without copying Iceland)

The transferable lesson is governance, not geography. Iceland is small, but the operating model scales.

1) Treat language as core infrastructure

If your agency serves multilingual communities, language support should be budgeted like uptime and cybersecurity. That means:

• dedicated owners (not “someone in comms”)
• a maintained terminology bank (glossaries that match legal phrasing)
• continuous evaluation, not one-time localization

2) Build a “human feedback bench” early

I’ve found that AI service quality rises fast when you stop depending on ad hoc feedback. Create a repeatable panel of reviewers:

• bilingual frontline staff
• community liaisons
• policy SMEs
• plain-language editors

Then give them a tight workflow: short weekly batches of prompts, review rubrics, and clear escalation paths.

3) Measure what matters: error rate by category

Public-sector AI errors aren’t all equal. Track them like an incident program:

• Policy accuracy errors (wrong eligibility, wrong deadline)
• Procedural errors (wrong steps, missing required documents)
• Language quality errors (unnatural phrasing, wrong formality)
• Safety errors (medical/legal guidance beyond scope)

When you can say “policy accuracy errors dropped from 8% to 2% on our top 200 intents,” you’re managing a service—not running a demo.

4) Plan for voice, not just chat

Iceland’s goal includes an Icelandic voice assistant. That’s smart because voice is where language exclusion gets brutal: accents, code-switching, and dialect differences are common.

For U.S. public services, voice matters in:

• 311 systems
• Medicaid/benefits call centers
• disaster response hotlines
• appointment scheduling for clinics and social services

If you only optimize for web chat, you’ll miss the populations who rely most on public services.

People also ask: practical questions about AI language preservation

Is language preservation really a government AI priority?

Yes—because it directly impacts access to services. When residents can’t use public digital services in their strongest language, you get higher call volume, lower compliance, and worse outcomes.

Can RLHF work for U.S. “low-resource” variants like dialects?

It can help, especially when the goal is service quality and local phrasing, not perfect linguistic coverage. The biggest wins come from focusing on high-volume intents (applications, renewals, status checks).

What’s the first step if we don’t have a dataset?

Start with a “top intents” pack: 50–200 real user questions (de-identified), plus ideal answers reviewed by SMEs and bilingual staff. Then evaluate the model against that pack every release.

Where this goes next: building an AI public service that doesn’t force English

Iceland’s effort is a reminder that AI systems don’t automatically respect linguistic diversity. They reflect the data and incentives we give them. When a government, nonprofits, and private companies treat language as digital infrastructure—and commit to ongoing human feedback—the results can move from “mostly understandable” to “trusted enough for daily use.”

That’s the bar U.S. public-sector AI should aim for, especially as agencies expand AI chatbots, voice assistants, and multilingual self-service portals in 2026 planning cycles. Better language support reduces operational burden, improves equity, and boosts public trust.

If you’re building AI-powered digital services right now, ask a hard question: Are residents switching to English because they prefer it—or because your systems leave them no choice?