Scalable Kubernetes Infrastructure for AI Platforms

Chapter 5. Responsible AI

Up until now, we focused on building a model and preparing it for deployment. Before going live, though, there are additional important considerations to cover.

Large language models (LLMs) are not just powerful from a technical standpoint but also from a societal standpoint. They’re being used to amplify spam, aid scammers, and even generate plausible propaganda at scale. They are trained on massive corpuses of data, which include material we want to draw from as well as all kinds of bias, racism, sexism, and potentially illegal or harmful material (for example, early versions of ChatGPT cheerfully explained to users how to build bombs). Because of this, there has been increased emphasis on and scrutiny of ethical and responsible training and use of LLMs.

Data Safety and Transparency

Like any other machine learning model, training data is of utmost importance to LLMs. Unlike other machine learning models, however, LLMs generate text, which creates a new attack vector for bad actors. One way to combat this and help prevent unethical data usage is to publish clear information about how an LLM was trained, what data it used, and who developed it. This information makes it possible for users, researchers, and regulators to effectively scrutinize the model’s behavior and to report or manage instances of harm or misuse. There are several platforms and frameworks available to help achieve this, such as UNESCO’s Global AI Ethics and Governance Observatory