No--even a really tiny, underpowered model like GPT-2 with no system prompt produces coherent (though not necessarily desirable or correct) responses.