Traditional AI/ML does well at prediction, but only if the future looks like the past. It struggles with prediction under any unseen scenarios since it is designed to recognize patterns from historical data. This is a problem in changing environments.
Causal AI/ML goes further. It understands the processes that generate the data, enabling it to piece together new distributions and discover innovative solutions that go beyond the historical data it was trained on.
This has two implications. First, causal AI deals better with unforeseen situations. Second, it can optimize to go beyond human ability rather than being restricted to mimic it.
Think of making a cake. Standard AI/ML recognizes that eggs and flour often appear together in recipes but doesn’t understand why. Causal AI would learn how the ingredients interact to form the final cake: that flour gives structure, eggs bind, and baking powder makes the cake rise. This means that if an ingredient changes, causal AI knows how to adjust the rest of the ingredients in response. From there it can create new cakes that are better than any of the existing ones by combining the best elements from its training recipes and cakes. Standard AI is doomed to perform, at best, as well as its training data.
Dealing with unforeseen situations makes it valuable in applications where safety and adaptability to new situations are critical. In self-driving, for example, it could better handle rare unseen scenarios involving complex interactions such as the simultaneous appearance of a deer and slippery road conditions, because it understands the interplay between them. It might have been trained on slippery roads on its own, and deer its own, and can put them together.
Optimization beyond its training data means that it can provide improvements on the status quo. For example, if causal AI is trained on historical production line data - based on historical human settings - then it can assess what the optimal actions are that should be taken to maximize yield. This is because it understands how the different parts of the production line impact each other - it has a holistic view of the system. The figure below shows a side-by-side comparison of a neural network versus causal ML. Standard AI (here a neural network) fails because it does not understand how tweaking one part of the system impacts how you need to tweak another part of the system. In this example it just ends up blaming the last machine in the process and so does worse than the training data. Causal AI goes beyond the training data to perform better than before, increasing yield and giving the actual actions to take in the real world, rather than just insights.