AI-associated linguistic shifts in Swedish university theses: a corpus-level analysis, 2020–2025

Surveys document widespread student use of generative AI, yet corpus-level evidence of how post-ChatGPT thesis writing has changed remains scarce. We analysed approximately 17,000 Swedish bachelor's and master's theses from six universities, spanning the release of ChatGPT (2020–2025), scoring each on five language dimensions with OpenAI's gpt-4o-mini and grouping them into six topical clusters. Across the corpus, four writing-quality dimensions — grammar, vocabulary, academic quality, and wordiness — rose together after 2022, a modest but statistically clear change of trend that coincides with the release rather than continuing any earlier drift. The shift was present in every discipline, including fields that the published-research literature does not flag, and its disciplinary ordering was inverted relative to that literature: theses in Sustainability and Society changed most, those in Informatics and Engineering least. The rater's own AI-likelihood judgment, by contrast, proved unreliable — it rated more polished writing as less likely to be AI-generated, the opposite of what AI-text detection assumes, and was structurally miscalibrated against its own language-quality scores. Taken together, the results point to a linguistic shift consistent with AI-influenced writing across Swedish thesis prose, while cautioning that the direct AI-likelihood signal cannot be trusted and that analyses of this kind should rest on the language-quality dimensions instead.