They do actually constraint the behavior, to various degrees of success which depends on the model, the system prompt, the inference parameters, the current context length and a lot more. Add in the new `developer` role and you have another venue for constraining the assistant outputs. Finally, structured outputs can help in forbidding specific terms too.
