Unless we're talking about something performance critical up front like an embedded system, odds are by the time an organization is going to care about logging performance overhead, they probably already have a few engineers dedicated to just optimizing logging.
Of course some developers may just go ahead and create a JSON/CSV string from scratch, which is about as safe as string interpolation instead of parameter bindings for SQL statements /s. But provided they're using a proper JSON/CSV/proto object builder library that uses String references, no data is going to get copied until a logging component needs to write it to a file or a network buffer. Therefore the only overhead are the pointers that form the structure.
Of course some developers may just go ahead and create a JSON/CSV string from scratch, which is about as safe as string interpolation instead of parameter bindings for SQL statements /s. But provided they're using a proper JSON/CSV/proto object builder library that uses String references, no data is going to get copied until a logging component needs to write it to a file or a network buffer. Therefore the only overhead are the pointers that form the structure.