Using Analogy as Means to AI Explainability

AI Systems are becoming ubiquitous and assuming different roles in different domains: they can act as recommendation systems in multiple contexts, they can work as personal assistants, they can generate text, etc. Whilst their contributions are clear and present on our daily lives, the reasoning behind them are not so transparent and may need additional explanations and interactions. For example, researchers might use AI to predict whether a compound will be toxic or reactive under certain conditions. A model might flag a compound as hazardous, but a chemist won’t accept that result without an explanation. They’ll want to know which molecular features led to that conclusion – was it the presence of a nitro group? A particular ring structure? Experts need interpretability to validate predictions, guide experiments, and ensure safety.

The use of AI is also common in the chemistry domain, and the advances of generative models sped up adoption in multiple tasks. One such task is the identification/substitution/mitigation of PFAS. We propose using a human-in-the-loop approach with AI agents to identify possible data gaps in product descriptions and provide suggestions to fill these gaps. In each decision-making step, the user makes the decision, leveraging their tacit knowledge. The AI acts as a co-expert in the process, providing suggestions that must be validated by users, and thus need to be explained.

One of these collaboration opportunities is considering analogue products. Different approaches are used, from calculating the similarity between products from a knowledge base to using generative models given the known information. In both cases, users should be able to check the reasoning behind the suggestions to gain confidence in their decision. Especially since the concept of similarity is associated with context and task. To empower users, we developed a system illustrated in the attached figure.