Citations and Provenance

The Trust Problem

Most AI-generated content has a trust problem. The reader — and often the editor — can't verify where any specific claim came from. The text is generated from training data that's invisible, unauditable, and potentially inaccurate. When an AI writing tool says "studies show that thermal bridging accounts for 30% of heat loss," there's no way to check which studies, whether that number is right, or whether the AI made it up entirely.

Pilot solves this by building citations into the generation process, not bolting them on after the fact. Every substantive claim in every article traces to a specific source document in your knowledge base.

How Citations Work

When Pilot writes an article, it identifies which source documents support each claim. These appear in the article as inline citations — superscript numbers that link to a sources section at the bottom.

The reader sees something like this:

Continuous exterior insulation reduces thermal bridging at stud locations by breaking the direct conduction path through framing members.[1] Even one inch of rigid foam board applied over sheathing can improve whole-wall R-value by 15-25%, depending on climate zone and assembly configuration.[2]

At the bottom of the article:

Sources

1. Thermal Performance of Wall Assemblies with Continuous Insulation — Building Science Research Institute
2. Residential Energy Code Compliance: Field Study Results — National Lab for Building Energy Analysis

Each citation points to a real document that exists in your knowledge base. The reader can see which source backs which claim. The editor can verify accuracy by checking the cited document. The chain from published sentence to original source is explicit and traceable.

Provenance, Not Just Attribution

Citations tell you which document a claim comes from. Provenance tells the full story: which documents were considered, how they were weighted, when the article was generated, and which voice configuration shaped the output.

Pilot maintains a provenance chain for every piece of generated content. This chain records:

• Source documents — which specific documents from the knowledge base contributed to the article
• Topic connections — which topics in the knowledge base the article draws from
• Voice settings — which voice configuration was active when the article was generated
• Generation metadata — when the article was created, which model was used, and the cost of generation

This provenance data lives in the Pilot console. Editors can inspect it for any article. It's the editorial equivalent of a paper trail — complete transparency about how a piece of content came to exist.

Why This Matters

Editorial Trust

An editor reviewing a Pilot-generated article can click any citation and see the source. If the article claims that building codes in a particular jurisdiction require specific insulation values, the editor can check the cited document to confirm. This turns the editorial review from "does this sound right?" into "does this match the source?" — a much faster and more reliable process.

Reader Confidence

Readers encountering Pilot-generated content see that it's sourced. In a media environment where trust in content is declining, visible citations signal that the content is accountable. The reader doesn't have to wonder where the information came from — the answer is right there, linked and verifiable.

Regulatory and Compliance

Organizations in regulated industries — financial services, healthcare, legal, government — often can't publish claims without sourcing. Pilot's citation model means generated content arrives with its sourcing already attached, reducing the compliance review burden.

Institutional Accountability

When content carries your organization's name, you're accountable for what it says. Generic AI content is a liability because you can't verify it before publishing (not efficiently, at scale). Pilot's content is verifiable by design — every claim points back to material you uploaded and control.

What Happens When Sources Conflict

Your documents might contain contradictory information. A 2019 report might cite one figure while a 2023 update cites a different one. Two authors in your collection might disagree on a recommendation.

Pilot handles this by weighing source documents based on their position in the knowledge base and flagging potential conflicts for editorial review. The citations make conflicts visible — if an article cites two documents that disagree, the editor can see both sources and decide which position to present.

This is editorial judgment with a safety net. Pilot doesn't silently pick one source over another and hide the decision. It shows its work.

What Citations Don't Do

Citations link to source documents in your knowledge base. They don't link to external URLs, public websites, or documents outside your collection. When a reader clicks a citation number, they see the source document's title and author — not a hyperlink to an external resource.

This is intentional. Pilot's citations verify that content came from your material. External link rot, changed URLs, and third-party content changes are outside Pilot's scope. The citations guarantee internal traceability, not external web references.

The Contrast

Most AI content tools produce text without provenance. The output is generated from model weights trained on data you can't inspect, with no per-claim sourcing, no editorial trail, and no way to verify specific statements.

Pilot produces text where every claim has a receipt. Not "this article was reviewed by a human" (though it is) — but "this specific sentence comes from this specific document that your organization uploaded and controls."

That's the difference between AI-generated content and cited, verifiable content that happens to be generated by a machine. For how this fits into the broader editorial system, see How Pilot Works. For how Pilot compares to tools without this capability, see What Makes Pilot Different.