Large Language Models Are Effective Human Annotation Assistants, But Not Good Independent Annotators