Generate-then-Ground
Google's approach of drafting an answer first, then searching for sources to attribute it — which we argue is more fragile than retrieving first.
When using artificial intelligence to write and research, the order of operations matters. There is a popular approach called generate-then-ground. In this method, an AI drafts an answer first, turns its own draft into search queries, and then searches for sources to fact-check and cite its work. Google’s RARR is a classic example of this.
But this approach gets the order wrong. By drafting first, you create a single point of failure. If the AI generates a bad search query based on its draft, it pulls the wrong evidence and makes the wrong edits. It also adds a massive lag time because the system has to make multiple passes. Worse yet, it often fixes factual errors simply by deleting them, which trims away useful nuance.
A much more robust alternative is to ground first and write second. This is known as retrieval-augmented generation. By gathering the evidence first, the facts sit in context before the model ever starts writing.
This weakness in drafting first also explains why adding citations after the fact is a poor way to fairly credit and pay content creators. To truly fix how we attribute information online, we have to start with the sources, not the draft.
Generate-then-ground is the approach of drafting an answer first and then searching for sources to attribute and revise it afterwards. Google's RARR (Retrofit Attribution using Research and Revision) is the classic example: the model drafts, turns its own draft into search queries, then retrieves passages to check facts and bolt on citations.
We think Google got this order wrong. It creates a single point of failure — one malformed auto-generated query cascades into wrong evidence and wrong edits — adds a latency tax of three passes, and often "fixes" facts by deleting them, trimming useful nuance.
The more robust alternative is to ground first and write second, as in retrieval-augmented generation, where evidence sits in context before the model speaks. This weakness also explains why post-hoc citation is a poor basis for fairly attributing and paying creators — the problem our CAPS proposal sets out to fix.