Copilot agrees that “men are pigs”

But this LLM has no idea what that means

Proof that ‘men’ can be ‘pigs’. Photo by Diego San on Unsplash

What happens when these proposed “strong AI” or AGI systems deal with simple human concepts, like the difference between a word and its spelling and a word and its meaning? Children can play these games and because they understand the concepts, they can be tricked into a contradiction. Does Copilot notice?

Can we trick Microsoft Copilot? Let’s setup with a code to test the logic. It is going to use metalinguistics, our ability to refer to the language itself, like words, as well as its meaning — simple ambiguity to resolve.

Figure 1. Putting Microsoft Copilot through its paces to show it can find text sequences in context without understanding what it is talking about. Here Copilot agrees than ‘men’ equals ‘pigs’

Now that we have established the code, let’s confirm it understands it.

Figure 2. Now that we agree on the code, Copilot immediately ‘forgets’ and chooses a more probable interpretation here OF MY NEXT SENTENCE, because it doesn’t track the context of our conversation as a human would.

In Figure 2, the system is unable to continue applying the code just agreed to. The equality of the code should remain. The correct answer in context would be that using the agreed code, ‘men’ has 4 letters. Why? Because we agreed the mapping therefore ‘men’ IS ‘pigs’. But ignoring this the LLM just uses background…