AI text detection is a similarity score, not a determination. The false positives and false negatives are structural, and they get worse as human writing converges on generated style. Provenance answers what detection only estimates: drafts, timestamps, history.
![Python code comparing four ways to decode JSON into a Grant type: json.loads and orjson.loads return dict[str, Any] (cast and pray); pydantic's TypeAdapter.validate_json and msgspec.json.decode return a typed Grant. Below, a 3-line generic deserialize[T] helper built on msgspec that decodes into any type, including list[Grant].](https://cdn.bsky.app/img/feed_thumbnail/plain/did:plc:q675zkhhrg3k5wtpcp457akk/bafkreibeu5ymfvzyw3impenuq4u72fveytlwkxtnjoxu44ythitm72buwe)
