Summary of “they Are Uncultured”: Unveiling Covert Harms and Social Threats in Llm Generated Conversations, by Preetam Prabhu Srikar Dammu et al.
“They are uncultured”: Unveiling Covert Harms and Social Threats in LLM Generated Conversations
by Preetam Prabhu Srikar Dammu, Hayoung Jung, Anjali Singh, Monojit Choudhury, Tanushree Mitra
First submitted to arxiv on: 8 May 2024
Categories
- Main: Computation and Language (cs.CL)
- Secondary: Artificial Intelligence (cs.AI); Computers and Society (cs.CY); Human-Computer Interaction (cs.HC); Machine Learning (cs.LG)
GrooveSquid.com Paper Summaries
GrooveSquid.com’s goal is to make artificial intelligence research accessible by summarizing AI papers in simpler terms. Each summary below covers the same AI paper, written at different levels of difficulty. The medium difficulty and low difficulty versions are original summaries written by GrooveSquid.com, while the high difficulty version is the paper’s original abstract. Feel free to learn from the version that suits you best!
| Summary difficulty | Written by | Summary |
|---|---|---|
| High | Paper authors | High Difficulty Summary Read the original abstract here |
| Medium | GrooveSquid.com (original content) | Medium Difficulty Summary Large language models (LLMs) have become essential for various applications, including personal assistants and recruitment tools. However, research shows that they perpetuate systemic biases. Previous studies mostly focused on Western concepts like race and gender, neglecting cultural concepts from other parts of the world. Moreover, these investigations typically examined “harm” as a single dimension, ignoring its varied forms. To address this gap, we introduce the Covert Harms and Social Threats (CHAST) metrics, grounded in social science literature. We utilize evaluation models aligned with human assessments to examine covert harms in LLM-generated conversations, particularly in recruitment contexts. Our experiments reveal that seven out of eight LLMs studied generated conversations riddled with CHAST, characterized by malign views expressed in seemingly neutral language unlikely to be detected by existing methods. Notably, these LLMs manifested more extreme views when dealing with non-Western concepts like caste compared to Western ones like race. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary Large language models can be helpful for many things, but they also have some bad habits. They often repeat the biases of the people who made them. Most research has focused on how this works in the West, where we’re used to thinking about things like racism and sexism. But what about other cultures? For example, there are many different kinds of social problems in India that aren’t exactly the same as racism or sexism. Researchers didn’t do a very good job of looking at these issues before. They mostly just looked at how language models talked about race and gender. So we decided to make a new way of measuring how bad language models can be. We call it CHAST, which stands for Covert Harms and Social Threats. It’s based on ideas from social science that help us understand when language is mean-spirited or unfair. We used this tool to look at how eight different language models talked about different topics, including some that are specific to India. We found that most of the language models had a lot of problems and said some pretty awful things. But it’s not just that they were saying bad words – it’s that they were saying them in a way that seemed okay on the surface but was actually really mean. |
