Compare tone of voice compliance when generating Alan Turing's hypothetical final conversation with Arnold Murray, using a simple text prompt versus a structured JSON output.
The result is a clearly dulled and more analytical response in the structured example, due to technical prompt elements distracting from tone of voice. A possible contributing factor is LLM endpoint forcibly turning temperature down to 0.0 when JSON output is requested.
Surprising twist: The text based response did not comply with the requested number of turns. It's as if the model decided the requested qualities demanded more turns and gave that more importance than the requested structure. For most use cases in AI writing this is an advantage.