Event language
UI language
When a large language model (LLM) has been trained on text featuring social biases, those biases implicitly impact the outputs of the model. Training an LLM on censored content, i.e., those pieces of content which remain after being subjected to state censorship (including alterations, deletions, and self-imposed censorship), results in what we term censorship bias. Most Simplified Chinese content on the Internet and, as a result, in the common crawl, has been subject to state censorship. In recent published work, I outlined a novel methodology in which we analyze censorship bias through the framework of comparing responses to prompts made in Simplified Chinese and Traditional Chinese. We applied this method to evaluate a number of popular LLMs developed by Google, Meta, OpenAI, and Anthropic (developed in the west and blocked in China) and we found evidence of censorship bias in all of them<strong>. </strong>I will discuss our results and the implications for privacy and security. I will end with a discussion of our work understanding the security implications of using generative AI in the software supply chain and how we are working with open source projects to better understand this.
