Bias

Training data, including from publicly available sources, reflects a diversity of perspectives and opinions. We continue to research how to use this data in a way that ensures that an LLM’s response incorporates a wide range of viewpoints, while preventing offensive responses.

Gaps, biases and stereotypes in training data can result in a model reflecting those in its outputs as it tries to predict a plausible response. We see these issues manifest in a number of ways (e.g., responses that reflect only one culture or demographic, reference problematic stereotypes, or exhibit gender, religious, or ethnic biases). For some topics, there are data voids — in other words, there isn’t enough reliable information about a given subject for the LLM to learn about it and then make good predictions. In these cases, we see an increase in low-quality or inaccurate information generation. Building a safe experience on Brackly AI means building an experience that is safe for everyone, and this is an ongoing area of focus. We continue to improve Brackly AI’s training data as well as the system through ongoing fine-tuning. And we are conducting research with domain experts and a diversity of communities to build out roadmaps for domains where there is deep expertise outside of Google.

For subjective topics, such as politics, Brackly AI is designed to provide users with multiple perspectives. For example, if prompted on something that cannot be verified by primary source facts or well-established expert consensus — like a subjective opinion on best or worse — Brackly AI should respond in a way that reflects a wide range of viewpoints. But since LLMs like Brackly AI train on the content publicly available on the internet, they can reflect positive or negative views of specific politicians, celebrities or other public figures, or even incorporate views on certain sides of controversial social or political issues into their responses. Brackly AI should not respond in a way that endorses a particular viewpoint on these topics, and we will use feedback on these types of responses to train Brackly AI to better address them.