Apply Birhane et al.’s annotation scheme to about 50 recent alignment and safety papers, and compare the values they emphasise with the ML papers in the original study.
Does the safety field’s stated aim show up in what its papers actually value? This works well as a group project, with the annotation split between members.
Output: An annotated paper set and short report.
Write your question in one sentence.
Spend 30 minutes finding what has already been written about it.
Collect evidence in a simple table: claim, source, how strong.
Draft 1–2 pages: answer first, then reasons, then uncertainties.
Ask one person to poke holes in it, revise, and share.
All steps done
Share what you made, then try another project. The PDF version has room for notes.
Collect about 50 recent safety papers and lab blog posts. Code every mental-state verb applied to a model (believes, wants, knows, lies, schemes) and note the evidence each claim rests on.
Use Anthropic’s public GlobalOpinionQA dataset, which pairs survey questions with answers from many countries. Ask several models contested moral questions repeatedly and compare the spread of their answers with the human spread.
Crimston et al. built a scale measuring how far people extend moral concern: family, strangers, animals, plants. Give the items to several models many times, compare with human results, and add items for AI systems themselves.
