A research paper from Amrita Vishwa Vidyapeetham has received the Best Paper Award at the 4th Workshop on Cross-Cultural Considerations in NLP (C3NLP), held at the Annual Meeting of the Association for Computational Linguistics (ACL 2026), one of the world’s premier venues for research on language and artificial intelligence. The workshop is organised by researchers from Google Research, the University of Copenhagen, the National University of Singapore, and other leading institutions.
The paper, “Sorry, Can’t Help You”: How Large Language Models Judge Failures to Help Across Languages, was authored by Pavithra PM Nair, Gilad Gressel, and Dr. Krishnashree Achuthan of the Center for Cybersecurity Systems & Networks at Amrita’s Amritapuri campus. It was one of two papers by the team accepted at the workshop; the second, Lost in Translation? How Language Shapes Responsibility Attribution in Large Language Models, examines how language shapes the way AI models assign responsibility.
Millions of people now turn to AI chatbots for advice, in dozens of languages, on questions that have no easy answer: whether to help a relative, what they owe a friend, how to weigh their own needs against someone else’s. The study asked whether these systems give the same moral guidance to everyone, or whether the answer depends on the language you ask in.
To test this, the researchers adapted a landmark experiment from cross-cultural psychology, in which Indian and American participants were asked to judge a person’s refusal to help someone in need. They posed the same scenarios to four leading AI models in ten of the world’s most widely spoken languages, including Hindi, Bengali, Urdu, Arabic, Mandarin Chinese, and English.
Across more than 2,600 responses, the models agreed in every language that failing to help someone is undesirable. But they disagreed substantially on two deeper questions: does the model believe helping is an obligation, and does it believe a person should be sanctioned for failing to help? Depending on the language, the answers changed. In practice, this means a user asking for guidance in Hindi and a user asking in English may be told different things about what they owe the people around them, by the very same AI system, without either user ever knowing.
The award committee recognised the paper “for its rigorous and interdisciplinary investigation of cross-lingual variation in the normative judgments of large language models,” and credited it with “establishing cross-lingual normative auditing as a promising direction for culturally sensitive NLP.”
“This is exactly the research we aim to do at Amrita. Aligned to the University’s mission, we focus on science that serves everyone, not just a privileged few,” said Dr. Krishnashree Achuthan, Dean of Postgraduate Programmes and Director of the Center for Cybersecurity Systems & Networks, Amrita Vishwa Vidyapeetham. “Millions of people already talk to AI in their own languages. Someone has to ask whether it treats them all fairly, and I’m proud our researchers are the ones asking.”
“This is my first best paper award, and I take it as a huge encouragement in our effort to make AI safer for everyone,” said Pavithra PM Nair, Scientist at Amrita and the paper’s lead author. “Most AI systems are built and tested primarily in English. Our results show that the values a model expresses can shift with the language you ask in, and users deserve that transparency.”
The research was supported by Amrita Vishwa Vidyapeetham and the IndiaAI Mission, Ministry of Electronics and Information Technology, Government of India. The team has released its code and data publicly to support further research, and is continuing this line of work with larger studies of how AI systems handle moral questions across languages and cultures.
