There's no distinction in the token-predicting systems between "instructions" and "information", no code-data separation.