An Open Ecosystem to Support AI Value Alignment & Human Feedback

Biases in AI systems are well-documented. As we move to the new world of Foundation Models (e.g. Large Language Models), due to the massive training datasets from disparate sources and the lack of transparency, biases are inevitable. Post-processes can remove some bias, but not all. The only option left for the users is to evaluate any specific areas of their concern with relevant benchmarks. If the system outputs do not meet the values of the users, there is really not much they can do to improve the status. This talk proposes a few practical approaches to address this situation: (i) Creation of an Open Trusted Data Initiative with a catalog that shows provenance of data sources, (ii) A flexible Testing & Evaluation framework that allows rapid changes to the choice of benchmarks & the evaluation criteria and, (iii) the ability to provide incremental knowledge and feedback to align the system for the intended use. The AI Alliance and IBM have already started this journey to build such an open ecosystem based on leading edge technology. We call on the AI community to join us in this adventure to deliver economic and social good.