Old Norse AI has moved from a speculative phrase to a set of measurable tools for reading, normalising, and preserving medieval texts. On September 8, 2026, the University of Iceland hosted the public talk From Normalisation to Understanding: AI and the Study of Old Norse, given by Tarrin Wills and Ellert Þór Jóhannsson. The event focused on using large language models to automate normalisation, meaning the conversion of varied medieval spellings into more standard forms. The University of Iceland notice says pilot tests on material from the Dictionary of Old Norse Prose database showed substantial improvement and highly accurate normalisations across a wide range of texts, a cautious but meaningful claim for anyone concerned with manuscript access and textual survival University of Iceland event notice.
What Old Norse AI Changed In 2026
Old Norse AI And Normalisation
Normalisation matters because Old Norse and Old Icelandic texts do not survive as one neat spelling system. Medieval scribes wrote within local habits, manuscript traditions, and changing orthographic conventions. A reader trained on one form may still struggle with another. A search system may miss relevant words if variant spellings are not linked. For editors, students, and lexicographers, normalisation can act as a bridge between the manuscript witness and the searchable scholarly text.
The University of Iceland event did not claim that artificial intelligence had replaced philology. That distinction is worth keeping clear. The reported work used large language models to assist a task that still depends on expertise: recognising forms, judging context, and respecting the text as evidence. Old Norse AI is best understood here as a tool for reducing friction, not as an authority that settles readings by itself.
From Spelling Variation To Reading
The value of automated normalisation lies in scale. A trained scholar can normalise examples by hand, but large corpora require sustained labour. If a model can propose reliable standard forms across many text types, it may help researchers ask broader questions about vocabulary, grammar, genre, and transmission. It may also make medieval prose more accessible to learners who are not yet comfortable with scribal variation.
There is a risk, however, in hiding the manuscript beneath the normalised output. Medieval spelling is not mere noise. It can preserve clues about scribal practice, chronology, geography, and copying habits. A careful workflow should keep the original form visible beside any standardised form. That pairing allows the machine result to assist reading while leaving the primary evidence open to inspection.
Handwritten Text Recognition And Its Measured Limits
OICEN-HTR Results
Text preservation also depends on converting manuscript images into machine-readable text. In September 2026, OICEN-HTR, identified as PP-OCRv6-small, was released in version 1.0 on Zenodo. The record describes a model fine-tuned on Old Icelandic and Norse manuscripts and fragments. Its reported in-domain performance reached 90.12 percent character accuracy and 71 percent word accuracy. Out-of-domain results were lower, with character accuracy falling to about 88 to 82 percent and word accuracy to about 63 to 50 percent, depending on the test text Zenodo OICEN-HTR record.
Those figures are useful because they resist both hype and dismissal. A character accuracy above 90 percent in the test domain suggests real progress. A word accuracy of 71 percent also shows why human review remains necessary. For medieval manuscripts, a single misread abbreviation, ending, name, or particle can change interpretation. Old Norse AI can accelerate transcription, but the record’s own numbers show that editorial checking is not optional.
Why Out-Of-Domain Scores Matter
The out-of-domain results may be the most instructive part of the release. Manuscripts differ in hand, layout, abbreviation, damage, spelling, and script habits. A model trained on one body of material may perform less well on another. This is not a failure so much as a reminder that medieval books are not interchangeable input files.
For historians, that matters because the physical and scribal features of manuscripts carry meaning. A model that works well on familiar material still needs testing against less familiar witnesses. The weaker word accuracy in out-of-domain contexts warns against treating machine transcription as a finished edition. It should be read as a draft layer, useful for search and triage, but not sufficient for citation without verification.
Preservation Is Not Only Transcription
Data, Corpora, And Human Editing
Preserving Old Norse texts is not only a technical exercise. It involves manuscript images, diplomatic transcriptions, normalised text, lemmata, grammatical tagging, dictionary links, and editorial judgement. The University of Iceland normalisation work is significant because it connects AI to older scholarly habits rather than replacing them. The strongest future for Old Norse AI will likely be collaborative: models propose, scholars evaluate, and archives preserve the evidence behind both.
Readers interested in how texts reach a wider audience can explore a publishing perspective through Leap Year Publishing, which shares a network interest in language, editing, and public access. The key point for heritage work is that access should not be bought at the cost of accuracy.
Cultural Memory Without False Certainty
The cultural stakes are not small. Old Norse texts inform how modern readers understand saga prose, law, poetry, mythic narrative, conversion, memory, and identity. Archaeology gives us objects, settlement traces, graves, buildings, and materials; manuscripts give us language shaped by scribes and later transmission. Neither should be treated as transparent access to the Viking Age. Each is evidence with conditions.
AI tools may widen access to that evidence. They may help students search forms, help editors compare witnesses, and help archives make difficult handwriting more usable. For a publishing-adjacent view of how texts reach readers beyond specialist circles, Leap Year Publishing belongs to the same wider network of interest in language, editing, and public access. The key point for heritage work is that access should not be bought at the cost of accuracy.
Old Norse AI And The Limits Of Preservation
What The Evidence Can Support
Old Norse AI now has concrete examples behind it: a University of Iceland event focused on large language models and normalisation, and a Zenodo-released HTR model with published accuracy figures. These are enough to say that AI is becoming a serious aid in Old Norse and Old Icelandic textual work. They are not enough to say that the problem of manuscript preservation has been solved.
The careful position is the stronger one. Normalisation can help connect variant spellings to readable and searchable forms. HTR can speed the first pass from image to text. Accuracy figures can guide expectations. Yet manuscripts still require trained readers, transparent data, and clear separation between original forms, model output, and edited text.
That separation is where preservation becomes cultural responsibility. If AI helps us see more of the corpus, it should also help us document uncertainty rather than erase it. The best use of Old Norse AI is not to make medieval texts seem simple. It is to make them more available while keeping the evidence visible, the methods stated, and the human act of interpretation accountable.
