Open kelizhang opened 1 year ago
There are two different kinds of causal discovery capabilities. ChatGPT is using its knowledge to predict the causal relationship (knowledge-based), whereas standard algorithms use the observations (data-based). They are complementary and in a real-world problem, both can be useful.
In later experiments, we found that ChatGPT can also determine relationships in other domains like medicine and arctic atmospheric science. We do find evidence of memorization too. For details, see the paper: https://arxiv.org/abs/2305.00050
So yeah, ChatGPT is not really solving the traditional data-based discovery task, but solving another task that is arguably useful too.
Your test seems just to demonstrate that ChatGPT memorized these causal relations other than possessing the causal discovery ability from observations.