In justice and in science, we use correlation to suggest, but causation to prove.
If someone is suspected of a crime, a fair court wouldn’t say “he resembles most of the other criminals, therefore he’s guilty.” Correlation is not adequate, we have to show evidence that one has committed the crime.
A neural net ignores causation, it works purely on correlation. A neural net does not know whether A causes B nor does it care, it only knows whether seeing A raises expectations of seeing B. This is very good for getting quick answers, but not always good for getting correct answers.
The Spurious Correlations site is a rich source of accurate training datasets for creating anti-knowledge.
Correlation is when you tend to observe two things together.
It might be a correlation if global soybean sales tend to go up whenever the Red Sox are beating the Yankees, but that doesn’t mean that one causes the other.
Causation is when you understand why something happens. To say that A causes B is to have a model that explains the reason for B.
Neural net AIs work on the principal of correlation: in reinforcement learning the classification (output) is correlated with certain kinds of inputs.
Neural net AIs ignore causation. They don’t use causal explanations, and they don’t produce them.
Example: We know that the presence of beta-amyloid plaques are correlated with neural degeneration found in Alzheimer’s disease. As of recently, what we don’t know is whether the disease causes the plaques to form, or if the formation of these plaques cause the disease.
Example: Scurvy has been killing people painfully since ancient times, and the cure (eating citrus fruit) has been learned and forgotten many times across different societies. It wasn’t until the 20th century that it became widely established that it was caused by a lack of vitamin C in the diet.