How much Icelandic news reads as AI-written?*
Outlet by outlet, day by day
What's up?
Table view
Most common topics within each category
How much Icelandic news reads as AI-written?
Articles scoring this or higher are marked as AI.
A lower number catches more AI, but risks flagging some human writing.
Kinds of AI text, from easiest to hardest for the detector to catch. Hover over one for details.Tap one for details.
Which detector gives the scores.
Only domestic articles on politics and government, business and the economy, and society and welfare that the outlet published itself: its own news, submitted articles and sponsored content. Foreign news and other topics, such as sport, culture and entertainment, are left out.
Why it matters: these articles shape how people see public affairs at home, and how they vote. AI-written text matters most here: an article on parliament or interest rates carries more weight than a football report.
Estimated % written by AIbased on your settings to the left
Based on your settings above. Click an outlet for its details.
1 Based on the training dataHow often the detector is right on texts whose origin we know
Flagged share of each known group
2 How these settings work on the training dataWhere each group's scores fall, and which ones your settings flag
Where each group's scores fall
How each group's scores spread. Solid bars reach your threshold (flagged as AI); faded rows are styles you didn't tick. ▲ = mean score.
Table view
3 Mean detection score by outletEach outlet's average score, compared with known human and AI text
Outlets with fewer than 10 scored articles in the period are faded, listed last and get no whiskers.
Table view
4 By article typeSubmitted articles and foreign news, apart from each outlet's own news
Each outlet's scored live articles in the period, split by type. Each cell: articles of that type, and the share of them at or above the threshold. The last two columns are the estimated % written by AI of all articles and of the outlet's own news only. Click a number to see those articles.
About GerviGrein
What is GerviGrein?
GerviGrein estimates how much of what Iceland's main news sites publish reads as if it was written by AI. It collects new articles from each outlet every day, scores them with two AI-text detectors and shows the estimated share for each outlet.
Why make this?
AI can now write news faster and cheaper than any newsroom, and it is increasingly used to churn out propaganda, push agendas and tell people what they already believe. The goal of this site is to lift up real journalism: work that people researched, wrote and put their name to. We don't judge people for wanting to voice their opinions, but we would prefer that they at least do it in their own words.
How does it work?
No detector is perfect, so after training, the site tests the detector on texts whose origin we know: Icelandic news articles published before ChatGPT existed (2018–2022), and AI-written articles that we generated with several AI models from different companies, in a variety of styles: from a headline alone, from the facts of a real story, as a write-up of foreign news, and as a human article reworded by AI. That shows how often it wrongly flags human writing and how often it catches AI.
Each outlet's flagged share is then corrected for both kinds of error. You can change the detector's settings yourself and see every step on the Settings page.
Which detectors does it use?
Two, built in very different ways. The main one is IceBERT, an Icelandic language model from Miðeind, which we then trained further here to tell human-written news from AI text, on thousands of articles from the Icelandic Gigaword Corpus (Árni Magnússon Institute, CC BY 4.0) and AI-written versions of them created by many of the most common and popular AI models. It runs on our own computer, so the text it reads isn't sent anywhere. The second is Claude: each article is sent through our own custom scoring harness to Anthropic's Opus model, which weighs the evidence and returns a structured AI-likelihood score.
Each article's page shows both scores on one shared scale, measured against human-written news: how many human articles score as high. With the default settings, an article is only said to have AI signals when it scores higher than every human-written news article it was compared with, and the strongest verdict needs both detectors to agree.
Is GerviGrein expensive to run?
It isn't cheap. Every new article is scored by a large AI model as well as by our own detector, and on top of that come the test texts and a computer running around the clock. With more funding we could add more detectors and make the scores even more reliable.
Does a high score mean an article was written by AI?
No. A score is the detector's guess, not proof. The numbers only mean something for an outlet's articles as a whole, where the errors even out. Please don't use this site to accuse a particular journalist or article.
What can't it see?
Text a person wrote and then had AI polish is the hardest to catch: the language detector finds most of it, the presentation detector little of it, so the real share of AI-assisted writing may be higher than shown. And the human comparison texts are a few years old, so newer writing styles could occasionally be mistaken for AI. Text written by AI isn't bad in itself, and it can help with things like making translations accessible. But readers should know what they are reading, so we would much prefer that outlets labelled their own articles to show when AI was used.
Are you against AI?
No. AI is a powerful tool, and this site couldn't have been built without its help. But with real power comes responsibility: readers should be able to trust that the news they read was written, checked and stood behind by people.
Was AI used to make this site?
Yes, in part, and we're aware of the irony.
Is there a browser extension?
Yes. The GerviGrein browser extension shows the estimated AI score of the article you're reading on the seven news sites, right in the toolbar: the robot turns redder the higher the score. It only sees page addresses on those sites.
Questions, corrections or feedback?
Feel free to contact me with questions or queries at [email protected].
How the site and the browser extension handle data: Privacy.
Latest articles
All outlets
All categories
GG-score
Viðskiptablaðið · 5. október 2026
FDA samþykkir aðra framleiðslueiningu Alvotech
Viðskiptablaðið · GG-skor 0.10: Nokkur óvenjuleg einkenni. Sjáðu niðurstöðuna á GerviGrein.
- Málfarsgreinir: 0.10
- Framsetningargreinir: 0.10
Viðskiptablaðið: 11% · Áætlað hlutfall greina skrifaðra af gervigreind, síðustu 30 daga
Lesa greinina á vef miðilsins · Um GerviGrein
Viðskiptablaðið · Greinar 5. október 2026
- 0.05Herjólfur óskar eftir 65 milljónum frá ríkinu
- 0.05Ef þær myndu bara sækja um …
- 0.10„Ekki einu sinni Baudenbacher gæti bjargað okkur“
- 0.04Leigabremsan ekki í samráð
- 0.04Evran ekki veikari í 17 mánuði
- 0.05Þrír dagar í uppsögn kjarasamninga
- 0.00Drangey SK 2 landar 147 tonnum á Sauðárkróki
- 0.16Smærri fyrirtæki bíða lengur eftir endurgreiðslum
- 0.06Stefán ráðinn til Fastus
- 0.09Fjarðabyggð takmarkar hækkanir við 2,5%
- 0.07Þörf á meiri liðleika í nýsköpun
- 0.17Íslenska áhættuálagið