How well do AI models and humans agree on right and wrong?

Existing tests for social norm alignment rely on artificial multiple-choice questions. This work proposes matching free-form responses to social dilemmas—comparing what an LLM says to human reference answers and vice versa. The team built a dataset of 3,000 non-trivial Danish social scenarios with human-judged reference solutions, then measured agreement between LLMs, humans, and combinations. Models showed consistent rankings but uneven alignment: they agreed more on concrete issues like neighbor disputes than abstract moral questions. Code and dataset available.