The original GPT-3 also did bad translations, just from scraps of foreign languages in the dataset. They seem to be worse though, so perhaps you're right that C

The original GPT-3 also did bad translations, just from scraps of foreign languages in the dataset. They seem to be worse though, so perhaps you're right that Chat-GPT is using another model the way it uses Internet searches