Limitations

Scores are a classifier's reading of the text, not a measurement of what anyone feels.

• A score is the model's confidence that the text is framed in terms of a moral foundation. It is not an intensity, and not a judgement of right and wrong.
• Moral Foundations Theory is one contested account of morality, and a classifier learns its annotators' readings of it.
• The labels say which foundation a text draws on, not which side it takes: praise for loyalty and a charge of betrayal both read as loyalty. Liberty is not measured.
• Every result names the model and version that produced it, because a different version can read the same text differently.

Known weaknesses

• Foundations with few examples in the training data are read less reliably than common ones.
• The annotators often disagreed with each other, and the models learn that disagreement. A middling score can mean the framing is arguable.
• The models read English only, and were trained on Reddit comments, so they may read other kinds of text, such as AI replies, less reliably.

Each model's card on Hugging Face reports its test results for every label. See Models.

Uses to avoid

• Decisions about a person: hiring, grading, discipline, credit, insurance, policing, immigration, or any medical or mental-health assessment.
• Reading a person: a score is not evidence of what a particular person believes, or whether they are a good person.
• Watching people: monitoring employees, students or anyone who hasn't agreed to it.
• Proof: a score is one model's reading. It doesn't settle what a text means, and it says nothing about whether an AI model has inner experience.