AI Translation Fails in Crisis: LLMs Distort Urgency Across 32+ Languages
A new study reveals a critical flaw in using AI for emergency communication.
A new study warns against deploying general-purpose LLMs and translation models for crisis communication. Research across 32+ languages shows these systems suffer substantial performance degradation and instability when translating crisis-domain text. Crucially, even linguistically correct translations can distort the perceived urgency of messages, and urgency classifications vary widely based on the prompt language. This highlights significant risks for using current language technologies in high-stakes emergency scenarios where clear communication is vital.
Why It Matters
Relying on flawed AI translation in disasters could misdirect aid and worsen outcomes for vulnerable populations.