How our services use AI
We are building SøkTilskudd, SøkStøtte and more services on the way. All of them use large language models to generate application drafts. This page sets out the principles that apply across them.
Model choice
Today we use Claude from Anthropic. We review the alternatives on an ongoing basis and switch if circumstances change.
What happens to the data
Anthropic is contractually committed not to train models on what we or our users submit, or on the model's responses. They delete inputs and outputs after 30 days. We have also turned off the option for user data to be piped back to Anthropic for model improvement.
What the AI does
It writes drafts. It structures the text according to what the administrator asks for, adjusts tone, and pulls facts from your organisation or company profile so you don't write the same thing over and over.
What the AI does not do
It makes no decisions. It does not submit applications without you reading and approving them. It does not judge who “deserves” funding. You sign, and that's how it should be.
Bias and limitations
Language models reflect the data they're trained on, and training data is rarely evenly distributed. In practice that means our services perform unevenly across user groups. It typically affects small organisations outside the major cities, less common subject areas, and users writing in Nynorsk (Norway's second written standard).
The concrete effects vary between services. See SøkTilskudd: AI-assisted application writing for an example.
How you can help
Let us know when a draft misses the mark. Thumbs-down in the service, or an email to marte@kveck.no with a concrete example. Patterns we hear about become test cases for the next round of improvements.