Multilingual and culturally adaptive tools for victim support.
Problem Statement
The broad reach and impact of any tool designed to support victims of terrorism depends on its
multicultural fluency in grief and trauma as well as its availability in multiple languages. The
under-representation of minority languages and communities in the Global South in the digital
corpus used to train Large Language Models (LLMs) has knock-on effects on AI-enabled tools’
ability to serve these populations safely and effectively. This is a problem larger than translation
alone, and points towards structural limitations in ways in which these tools capture cultural
context, social norms and other sensitive cues. Addressing these gaps requires investment in
the development and training of AI-enabled models that are better equipped to engage in with
linguistic diversity and cultural nuance, particularly in contexts involving trauma, vulnerability
and the specific needs of victims of terrorism from around the world.
The State of Multicultural AI-enabled Tools
As discussed in the Gap Analysis of Digital Tools to Support Victims of Terrorism, relevant work
is already underway to move from translation to cultural fluency. While models like Cohere’s
Aya have expanded support to 101 languages, initiatives like Google’s Amplify demonstrate that
technical capability must be paired with domain expertise to catch localized harms like
misinformation or culturally specific triggers. Benchmarks, such as WorldView-Bench and
CARE, allow developers to evaluate how models navigate conflicting global perspectives.
Furthermore, architectures like Cultural Mixture of Adapters (CuMA) are being developed to
prevent models from "averaging out" cultural values. National applications, such as India’s
Jugalbandi, show that government investment in specialized translation layers can be combined
with generative AI to deliver vital information via accessible messaging platforms like WhatsApp.
Core Questions for the Working Group
To translate these research advancements into functional tools for victims of terrorism, this
working group will address questions such as:
• How can we evaluate an AI-enabled system’s ability to recognize and respond to
"trauma cues" across different languages and cultural contexts?
• How can we fix pre-training, training, and reinforcement learning to ensure that diverse
cultural or social values are reflected in the final model?
• How can we collect multicultural training data on trauma and grief in a way that is
human rights-compliant, safe, trauma-informed and victim-centric?
• How can governments develop and support translation infrastructure for their national
and minority languages, and how can the private sector build services on top?
• How do we implement "red-teaming" protocols that use local domain experts to test the
safety and cultural relevance of AI-driven tools before they are deployed in high-stakes
environments?
• How can we ensure that a human rights-based approach informs the design,
development and use of relevant tools?


Comments
Post a Comment