Source attribution in AI-enabled knowledge work

Separate origin, support, and responsibility

Treat three questions independently. Origin asks where an item came from and how it changed. Support asks whether the source actually justifies the statement being made. Responsibility asks who approved the final use. A provenance record may be intact while the content remains false or irrelevant. A correct source may be cited for a claim it does not support. Human approval may be recorded without evidence that the reviewer inspected the source.

Before drafting, label inputs as authoritative record, primary research, official guidance, licensed material, internal working document, generated suggestion, calculation, or synthetic example. Capture the authoritative URL or repository location, title, publisher, version, date, access condition, and owner. Preserve the exact source fragment for claims that are quantitative, consequential, or likely to be challenged.

The C2PA explainer describes Content Credentials as cryptographically bound provenance records and explicitly distinguishes verifiable provenance from a judgment that content is true or accurate. That boundary makes provenance useful evidence without turning it into a truth certificate.

Sources: [1]

Assemble the attribution packet

Create one entry per claim-source or asset-source relationship. Record the draft statement, source identity, relevant passage or data cells, version, transformation, generated contribution, calculation, reviewer, decision, corrected wording, and refresh trigger. Use dispositions such as direct support, partial support, contextual only, origin verified but claim unsupported, inaccessible, superseded, or remove.

For calculations, preserve input labels, units, periods, filters, formulas, and rounding. For summaries, retain the relevant passage and state whether the summary narrows or generalizes it. For generated text, identify the human-selected evidence and revisions rather than citing the model as a factual source. For an image or document with Content Credentials, validate the credential while still checking whether the depicted or stated content supports the intended use.

The NIST Generative AI Profile discusses information-integrity and confabulation risks and recommends practices around source verification and provenance. It is voluntary risk-management guidance and does not prove that a particular attribution workflow is complete.

Sources: [2]

A worked packet for a synthetic market briefing

A synthetic market briefing includes three elements: a regulator's definition, a chart calculated from an official dataset, and an AI-generated illustration. The first packet entry records the official page, publication date, exact definition, and a final sentence that preserves its scope. The second records the dataset table, download date, filters, numerator, denominator, formula, and chart file. The third records that the illustration is synthetic and not evidence of a real customer or event.

Review finds that the chart title says “all firms” while the dataset covers only surveyed firms above a stated size. The formula is correct, but the claim is too broad. The reviewer changes the title, adds the population and period, and preserves the rejected wording in the packet. The Content Credential on the illustration validates its editing history, but the team still labels the scene as synthetic because provenance cannot establish that the depicted event occurred.

The final packet allows another reviewer to reproduce the chart, locate the definition, identify the generated element, and see why the wording changed. This example is original and synthetic. It does not establish the reliability of any real briefing or credential implementation.

Test attribution at the final-output boundary

Give an independent reviewer only the finished output and the packet. Ask them to resolve each source, find the relevant evidence, reproduce selected calculations, identify generated or synthetic elements, and confirm that the final wording matches the support disposition. Sample the highest-consequence claims rather than treating every link as equally informative.

Check how the packet survives export, copying, format conversion, localization, and later revision. Reopen entries when a source changes, a dataset is refreshed, a claim is reused for a new audience, or provenance metadata is stripped. Track unresolved sources and unsupported claims separately from broken links.

Two Prune can observe how a participant attributes, checks, and narrows source-dependent claims within a bounded simulation. It does not certify provenance systems, determine legal compliance, establish factual truth across a domain, or make employment decisions.