Comment by in-silico

This assumes that the model's behavior and memories are faithful to their english/human language representation, and don't stray into (even subtle) "neuralese".