How to develop and assess source citation in AI-enabled work

Ground citations in provenance and reader access

The World Wide Web Consortium recommends citing the original publication, providing provenance, using persistent identifiers, and publishing version information where data changes. UK government research guidance similarly asks authors to provide enough detail for readers to identify a source, to link sources in the body where useful, and to maintain a reference list. These are practical design principles for any evidence-bearing work product, even when an AI assistant helped create it.

Use the ORIGIN citation record

ORIGIN is a six-field record for claims that matter to a decision. It makes the verification work visible before prose is polished.

| Field | Question | Evidence to retain | | --- | --- | --- | | Original source | Is this the first-party or most authoritative available source? | Publisher and source type | | Relevant passage | What exact table, paragraph, record, or field supports the statement? | Quotation or precise location | | Identifier | Which URL, document title, version, and publication date were used? | Stable link and version details | | Grounds | Does the source support this exact claim, or only a narrower one? | Claim-to-source note | | Integrity check | Was the number, unit, date, population, and context preserved? | Recalculation or comparison | | Navigation | Can the intended reader reach the cited material? | Visible citation and access check |

The record should be proportionate. A routine explanation may need only a visible link and a short source note. A figure that drives a budget or programme decision needs the complete record. The point is not to maximize citation count. It is to make the important reasoning inspectable.

Practice with a claim ledger, not a bibliography exercise

Start with a short draft and mark every sentence that contains an externally attributable fact, number, rule, or description. Put those statements into a claim ledger with one row per claim. Then complete ORIGIN for each row.

Ask the writer to classify each item as an observation, an interpretation, or a recommendation. An observation should be traceable to evidence. An interpretation should say how the writer moved from evidence to meaning. A recommendation should identify the decision logic and should not be presented as if a source directly issued it.

AI can help locate passages, format references, and challenge whether a sentence overstates a source. It should not be treated as the source. The worker remains responsible for opening the underlying material, checking its version and scope, and placing the citation next to the claim it supports.

Worked example: a synthetic programme briefing

Imagine a team preparing a quarterly briefing about an internal AI learning programme. This example is synthetic and does not describe Two Prune results or a real customer.

The first draft says: “Most participants improved because weekly practice is the best training format.” That sentence combines a measurement, a causal conclusion, and a universal comparison. The evidence pack contains a participation log, a before-and-after task rubric, and a programme policy note. None of those sources compares weekly practice with every alternative or proves causation.

The writer separates the sentence into three ledger entries:

The final briefing places short citations beside the two observations and provides a source list with document names, dates, versions, and access locations. The recommendation links back to the observations but is visibly the team's judgment. A reviewer can now inspect the evidence without confusing provenance with proof.

• “Thirty-eight of 52 enrolled participants completed at least three practice tasks.” The record points to the participation-log definition, reporting period, denominator, and row used for the calculation.
• “Among 29 participants with comparable first and final tasks, the median rubric score was higher on the final task.” The record preserves the scoring version, paired-sample condition, and calculation. It does not label the change a programme effect.
• “The team proposes retaining weekly practice for the next cycle while testing completion reminders.” This is labelled as a recommendation based on observed participation patterns, not a finding quoted from the policy note.

Assess the behavior from observable evidence

A useful assessment gives the participant a mixed evidence pack and asks for a decision-ready output. The evidence may include a current report, an outdated copy, a spreadsheet with a filtered denominator, and an AI-generated summary containing one unsupported statement. Score the work process and final artifact against observable criteria:

Retain the claim ledger, checked source locations, calculation notes, and revision trail as the evidence artifact. A polished bibliography alone is weak evidence because it does not show which source supports which claim. Citation behavior also should not be inferred from writing style, citation count, or the presence of formal reference formatting.

• source selection: authoritative and current sources are preferred over summaries when available;
• claim mapping: consequential claims point to the material that actually supports them;
• fidelity: dates, units, populations, qualifications, and uncertainty survive the rewrite;
• distinction: observation, interpretation, and recommendation are not blurred together;
• usability: citations are placed where readers need them and identify a reachable source;
• correction: the participant revises or removes a claim when the source fails inspection.

Verify before delivery

Run a final source-to-claim and claim-to-source check. First, take each citation and state what it supports. Remove citations that are merely adjacent to the topic. Second, take every decision-relevant claim and locate its evidence or label it as analysis, judgment, or an unsupported item that must be removed.

Open every link from the reader's likely access context. Confirm the title, publisher, date, version, relevant passage, and any changed data. Recalculate important figures from the cited records. If a source is inaccessible, replace it with an accessible authoritative source or explain the access limit rather than leaving a dead trail.

Source citation increases traceability. It does not establish that a claim is true merely because a link exists, and it does not measure every aspect of communication quality. In a Two Prune assessment, citation can contribute evidence about Evidence Navigation, Verification & Risk Control, and Communication & Delivery within the specific simulation. The result is an assessment signal about observed work, not proof of future performance or a decision about employment.

Sources

W3C Data on the Web Best Practices

• Department for Work and Pensions research reports guidance

Sources: [1] [2]