Who wrote the abstract? Analysis of AI use in theses from Skemman

Since ChatGPT was released in late 2022, niversities have been asking what effect such tools have on thesis writing. Most of the evidence, however, comes from surveys or from detection tools built for English. This talk presents a study of the abstracts of final theses in Skemman, Iceland's open thesis repository. The data cover 22,240 theses from 2019 to 2026, with roughly 19,000 abstracts in Icelandic and 11,000 in English. General-purpose detection tools have not been developed specifically for Icelandic, so we trained our own detection models. They are based on Icelandic and multilingual language models that learn to distinguish between abstracts written before the ChatGPT era and nearly 2,000 abstracts composed by 36 language models from the text of the theses themselves. Thresholds were set so that only 1% of the pre-ChatGPT abstracts are flagged. At those thresholds, the proportion of flagged Icelandic abstracts rises from 2.4% in 2023 to 42.2% so far in 2026. For English, it rises from 7.6% to 62.1%. The rise appears across institutions and degree levels, and schools differ considerably. The commercial detector Pangram catches almost all generated English abstracts at its default setting, but only 12% of the Icelandic ones, which can be remedied by modifying the flagging threshold. The results point to a real shift in how students write. They say little about individual students since they are not a proof of AI usage.