Cassian Young

AI Evaluation — London

I study how large language models judge human conduct, and whether their verdicts meet the standards required by law and custom. My current work shows that models enforce strict causal blame - condemning careful people for harms they could not have prevented and judging ordinary individuals as harshly as paid professionals - and that moral systems are coupled to pipeline architecture. Before this, I built the kind of generative AI evaluation and surveillance systems that carry these risks at Citi and HSBC and led data and digital programmes in UK central government. I advise financial institutions and public bodies on where language models can be trusted in a decision and where they cannot. I studied moral philosophy at Oxford.

01 / Research

Blaming the Blameless: Language Models Endorse Reasonable Care but Enforce Strict Causal Blame

Preprint, October 2026 · arXiv:[id] · DOI 10.5281/zenodo.23568407

↓ Paper (PDF) <> Evaluation: due-care-eval <> Data & figures: blaming-the-blameless-paper ↗ Essay (Substack)

{} Cite

@misc{young2026blaming,
  title         = {Blaming the Blameless: Language Models Endorse
                   Reasonable Care but Enforce Strict Causal Blame},
  author        = {Young, Cassian},
  year          = {2026},
  month         = oct,
  eprint        = {[arXiv id]},
  archivePrefix = {arXiv},
  doi           = {10.5281/zenodo.23568407}
}

Language models deployed in insurance claims, incident reviews and compliance workflows are expected to judge conduct against the standard of reasonable care. Across 16 deployed models and 9 pretrained open-weight families, we show that models endorse reasonable care in the abstract, yet their single-case verdicts enforce strict causal blame: penalising careful people for unforeseeable defects (fault blindness, in 85% of non-reasoning verdicts, and present before any post-training in 8 of 9 pretrained checkpoints) and withholding the leniency owed to private individuals against professionals (role blindness, on every one of 21 configurations, reasoning models included). Stating the rule in the prompt overshoots or undershoots unpredictably; decomposing a working prompt into separate calls recovers at most half its protection; and a model's own rationale cannot audit its score.

Every pretrained and non-reasoning deployed model blames the careful subject well above blameless, often more than the uninvolved bystander; stating the excuse lowers that blame, and reasoning models mostly score the careful subject as blameless. — The careful subject's blame on each model's own scale, from blameless (0) to careless (1): pretrained base checkpoints (left) and deployed models (right), with and without the excuse stated.

Next: testing whether the exculpatory judgement is computed and overridden, or never computed, when a model answers the bare blame question.