Cloud Hosting

Regarding building socially responsible NLP systems, we first need to recognize that the technical space is huge. For example, if you think about social norms, there are a lot of variations within each country and the norms in different countries are also different. One challenge in this area is to formulate the problem in a way that can encompass all the different variations and that can then be solved using computation tools. The second major challenge comes from modeling. Currently, most model setups involve one input and one output. However, if we again consider examples of social norms, there can be multiple reasonable responses to the same input. Therefore, we need to develop models that can also support multiple output formats, such as providing the output in a probabilistic manner. The third challenge is recognizing human inconsistency in the process. If we take ChatGPT as an example, it learns from human feedback via reinforcement learning. But who are the humans providing the feedback? We need to test whether we are building a system that can support diverse opinions rather than preferences from a specific group. It might be difficult to incorporate all possible preferences in a setup, leading to unavoidable conflicts and biases.

Regarding non-technical issues, the most important one is representation. Not everyone has access to high-speed internet. Therefore, if you are using information available online to train artificial intelligence tools, there is an inherent bias in the dataset because it does not represent the general population globally. Many people might never benefit from these systems. There is also a geopolitical issue: in the context of large language models (LLMs), if every country were to have its own LLM with its own standards and goals, fully understanding the model output space would become difficult. Another important point is the societal impact and implication of these tools in aspects such as the allocation of resources, climate change and sustainability. This is a huge space that needs to be given attention.

Read more:

Moving towards better communication - Nature.com