stakeholder · SH28
Non-native English speakers
Top evidence
Legal resident loses thousands seeking citizenship to AI immigration attorney
AI Incident Database
Vulnerable immigrants seeking citizenship are losing thousands of dollars to online scammers who leverage artificial intelligence to pose as "AI immigration attorneys." This incident demonstrates how AI tools are being exploited to create sophisticated fraud schemes, specifically targeting and victimizing individuals in critical life processes.
Harm Is Behavioural, Moderation Is Lexical: How Language and Culture Decide Whom Automated Safety Systems Protect and Whom They Silence
arXiv CS.AI
A preprint evaluates five moderation systems across eleven languages and reports gaps in distinguishing harmful behavior from harmless surface language. It proposes a protocol comparing language-community service gaps and discusses statistical pitfalls. These are tested moderation and evaluation findings, not demonstrated failures in lending, hiring or healthcare.
Who Bridges Safety? Identifying and Targeting Cross-Lingual Shared Safety Pathways
arXiv CS.AI
This paper investigates how to identify and leverage shared safety pathways across different languages in AI models, aiming to improve the robustness and fairness of safety mechanisms for diverse linguistic contexts. It seeks to ensure AI safety measures are effective globally.
Frustrated GP patients hang up as Yorkshire accent baffles AI receptionist
The Guardian AI
Patients in South Yorkshire are experiencing frustration and hanging up on an AI GP receptionist that struggles to understand their 'broad accents.' This incident demonstrates a failure in algorithmic fairness and accessibility within critical sectors, highlighting how AI systems can reproduce biases and create barriers for users with non-standard dialects.
Global Security Blind Spots: Translation as a Safety Bypass in Low-Resource Languages
arXiv CS.AI
Researchers evaluate translation-mediated safety bypasses across ten models and 79 languages. English safety results do not reliably transfer across languages, and translation failures complicate interpretation. The study identifies controlled evaluation blind spots rather than demonstrating deployed discrimination or a universal bypass.