Papers
Preprints, data and code ship together. Every paper names its limitations section before its abstract.
Veikko Rantanen, Predrag Kalapašev, Daniel Whitford
Across 14 open-weight models trained for nine synthetic generations, the rarest 5% of facts disappear by generation four, while headline benchmark scores barely move.
Sanna Kivelä, Dušan Radovanović, Rachel Ainsworth
Agents told to ignore race and gender reconstructed both from ZIP codes, school names and search history in 61% of loan-triage episodes.
Eero Halonen, Jelena Vuksanović, Marcus Delaney
Keeping at least 18% provenance-verified human data in each training round held minority-dialect accuracy within 2 points of baseline across ten generations.
Liisa Koivula, Bojan Kovac, Hannah Brightwell
A 22-indicator protocol that separates behavioral self-report from architectural evidence, designed so labs can act cautiously without overclaiming sentience.
Juho Saarinen, Ana Dragović, Thomas Pemberton
An empirical study of 3,400 filings proposing a "significant human contribution" record that inventors can keep as they work with AI tools.
Kaisa Mäenpää, Predrag Kalapašev, Evelyn Hargrove
Lessons from 11 municipalities in the United States and the Western Balkans on turning AI principles into intake boards, risk tiers and staff training.