As far as I know, the ontology itself is entirely curated. On the other hand the gene annotations are a mixture of curation and automatic annotations. This is tracked with evidence codes.

Yes it is manually created. From their own docs:

Shared vocabularies are an important step towards unifying biological databases, but additional work is still necessary as knowledge changes, updates lag behind, and individual curators evaluate data differently. The GO aims to serve as a platform where curators can agree on stating how and why a specific term is used, and how to consistently apply it, for example, to establish relationships between gene products.