Project overview
Medieval narratives of othering targeting Muslims, Jews, and marginalised groups established foundational frameworks of white supremacy that persist in Western collective memory. These historical texts are now being actively recuperated by far-right movements to legitimise exclusionary ideologies, creating an urgent need to address biases embedded in medieval sources.
Large Language Models trained on open-source historical corpora risk amplifying these dangerous biases. Modern toxicity filters fail on historical vocabulary; this failure is compounded in pre-modern data, where bias is often coded in context-dependent discursive patterns. Unmitigated, this poses severe risks to ethical AI development.
This project aims to develops a methodology to detect and mitigate these biases in Anglo-Norman materials ready to be annotated. Our primary goals are to:
-Create the first human-annotated ground-truth dataset for medieval bias.
-Develop a context-aware historical toxicity classifier capable of identifying historical discursive patterns.
-Establish ethical annotation and moderation strategies to protect human reviewers from exposure to historical toxic content.
This pipeline establishes a new technical and ethical standard for safeguarding models trained on pre-modern open sources. This initial proof-of-concept secures groundwork for a larger grant focused on scaling the classifier across other European historical and cultural areas, ensuring AI contribute to social equity and safety.
Large Language Models trained on open-source historical corpora risk amplifying these dangerous biases. Modern toxicity filters fail on historical vocabulary; this failure is compounded in pre-modern data, where bias is often coded in context-dependent discursive patterns. Unmitigated, this poses severe risks to ethical AI development.
This project aims to develops a methodology to detect and mitigate these biases in Anglo-Norman materials ready to be annotated. Our primary goals are to:
-Create the first human-annotated ground-truth dataset for medieval bias.
-Develop a context-aware historical toxicity classifier capable of identifying historical discursive patterns.
-Establish ethical annotation and moderation strategies to protect human reviewers from exposure to historical toxic content.
This pipeline establishes a new technical and ethical standard for safeguarding models trained on pre-modern open sources. This initial proof-of-concept secures groundwork for a larger grant focused on scaling the classifier across other European historical and cultural areas, ensuring AI contribute to social equity and safety.