Identifying the moral foundations of fictional characters

In a new study that appeared as part of EMNLP 2025 led by two students from our lab, Luca Mitran and Sophie Wu, we developed a new method for measuring moral values in stories. We call it the Moral Foundations Character Actions Questionnaire (MFCAQ), and we used it to probe the moral logic of nearly 3,000 folktales from 55 countries.

Most prior work measuring the moral foundations of stories has relied on dictionaries and keywords (e.g. if a story contains the word “father” or “king,” it gets tagged for “Authority.”). These kinds of studies are limited both because of the ambiguity of language and also because they disregard narrative structure. Our approach grounds the analysis of moral foundations within the actions of characters. Rather than look at the distribution of language across stories, it focuses on understanding the behavioural cues of characters.

To do so, we develop a questionnaire derived from Moral Foundation Theory and align it with characters: “In this story, are the character’s actions…showing care for others / driven by a sense of fairness / exhibiting loyalty to family [etc.].” We then prompt LLMs to read stories and then answer 15 questions related to 5 moral foundations.

We find that frontier models like GPT and open-source models like Gemma2 correlate reasonably well with human judgments (see Figure below). It is important to note that our human inter-annotator agreement was only moderate (alpha = 0.44 on 735 questions). This suggests there is a solid amount of interpretive ambiguity in this kind of task, which explains at least some of the model variance. But we also found that some categories exhibit higher accuracy than others. For example, all models struggled with the dimensions of “purity” and “authority.” And overall, models exhibited an over-prediction bias, generally predicting higher scores than their human counterparts.

What We Found in 2,697 Folktales

Some of our big takeaways:

1. Moral foundations appear in almost every folktale

Nearly every story contained at least one moral foundation, and most included multiple. Folktales don’t just teach one lesson.

2. Moral values are surprisingly consistent across cultures

While there are notable regional differences, the overall distribution of moral foundations is remarkably similar around the world. This lends support for the cross-cultural aspects of MFT.

3. The “positivity bias” in folktales is overstated

Past research claimed folktales primarily teach positive morality (heroes doing good deeds). But when we measure actions, not words:

• There is a tilt toward positive portrayal, but it’s modest.
• Negative content is just as common in certain foundations.
• Many characters are morally ambivalent, neither heroes nor villains.

What’s Next?

Future directions include:

• Applying MFCAQ to novels or film
• Studying character-level moral development over time
• Comparing moral positioning of models
• Testing multilingual models on non-English texts
• Evaluating how moral foundations interact with other literary indicators like gender, genre, and plot arcs