Why AI Struggles When Two Systems See the World Differently

Imagine two people organizing a library. One sorts books by genre, the other by author's last name. Both systems work, but if they try to share a catalog, chaos ensues. That's an ontological mismatch: different ways of cutting up the world. In AI, this happens when two systems—say, a self-driving car and a smart traffic light—use different categories for 'pedestrian' or 'obstacle.' The result? Miscommunication and potentially dangerous errors.

A researcher is exploring how to measure this mismatch mathematically. They propose treating each ontology as a labeled partition or topology of a 'thingspace'—the set of all possible things. The difference between two ontologies would be a number (like a distance), but it might not be symmetric: one ontology could be finer-grained than the other. This is similar to how the Kullback-Leibler divergence, a common measure in statistics, is asymmetric.

But there's a catch: what if the mismatch never actually occurs in practice? If two ontologies disagree about something that never happens in the real world, does it matter? The researcher calls this an 'unwitnessed mismatch.' It might still matter if we expand the thingspace to include new possibilities, because then the mismatch could suddenly become relevant. So even hidden disagreements can have future consequences.

For everyday people, this research matters because AI is increasingly making decisions that affect us—from loan approvals to medical diagnoses. If different AI systems within a company or across companies use incompatible categories, errors can slip through. Understanding and measuring ontological mismatches could lead to better AI integration, fewer mistakes, and more reliable systems. It's a behind-the-scenes issue, but one that could save time, money, and maybe even lives.