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

Authority depends on the claim

UK government guidance for country-policy information warns that there is no universal hierarchy of sources. It recommends considering how information was obtained, the methodology, whether it is first-hand or secondary, possible false corroboration, and the source's limitations. The U.S. Government Accountability Office's 2024 Government Auditing Standards similarly treats validity and reliability as separate considerations within evidence quality.

That means authority is claim-specific. A regulator is authoritative for the rule it issued. A company may be authoritative for the current features it offers or the terms it publishes. Neither is automatically independent evidence that a policy achieved its intended effect. A primary source can accurately state what its author decided while remaining partial evidence for a broader impact claim.

Create a STANDING record

For every source carrying a consequential claim, create a STANDING record. It is a compact source profile, not a generic credibility score.

| Field | Question | Evidence to retain | | --- | --- | --- | | Source identity | Who is legally or professionally responsible for the material? | Named organization, author, or office | | Task mandate | What gives that source standing for this exact claim? | Issuing power, direct custody, or relevant expertise | | Acquisition | How was the information gathered or produced? | Method, sample, measurement, or record origin | | Named version | Which edition, date, jurisdiction, and status apply? | Version metadata and supersession check | | Dependencies | Is this independent evidence or a restatement of another source? | Upstream source and citation chain | | Interests and limits | What purpose, incentives, exclusions, or uncertainty affect its weight? | Declared purpose and material limitations | | Needed corroboration | What other source type would test the claim? | Independent record, direct observation, or primary text | | Grade for this claim | What can this source support here? | Direct, qualified, contextual, or unusable |

The last field must include "for this claim." A source should not receive one permanent grade. Its standing changes with the proposition. A product manual can directly support a feature description and still be only contextual support for a claim about user outcomes.

Trace sideways citations back to origin

Generated answers commonly present several links that all repeat the same upstream statement. Counting those links as independent corroboration creates false confidence. Open each source and follow its citations until the original record, study, standard, or announcement is found. Record repeated downstream summaries as one evidence lineage, not several confirmations.

When the origin cannot be reached, mark that limitation. Do not upgrade a statement because many pages copy it. Similarly, do not assume that "primary" always means "better." A first-hand anecdote may be primary but poorly suited to estimating prevalence. Authority must be matched to the claim type.

AI can assist by extracting version details, proposing the upstream source, or comparing two editions. The worker should verify those details on the source itself. Generated source descriptions, fabricated citations, and snippets without opened pages do not establish standing.

Worked example: an AI policy briefing

Imagine an operations team preparing a briefing on whether an internal workflow must change after a new external rule. This is a synthetic example and does not describe legal advice, a Two Prune customer, or a real regulatory requirement.

The evidence pack contains an official rule page, a regulator's explanatory note, a trade-association article, a software vendor blog, and an AI-generated digest. The digest says the vendor blog and trade article "confirm" the new requirement.

The analyst builds a STANDING record for the sentence describing what the rule requires. The official text has direct issuing authority, but its version and commencement status still need checking. The regulator's explanatory note can clarify intended administration, while the analyst verifies that it does not replace the controlling text. The trade article offers qualified interpretation for its members. The vendor blog explains how one product responds and may reveal operational options, but it cannot establish the governing requirement. The AI digest has no independent standing; its statements must be traced to the sources it summarizes.

The analyst then reviews dependencies. Both the trade article and vendor blog quote the same paragraph from the explanatory note. They are not two independent confirmations. The final briefing cites the official text for the bounded rule statement, uses the explanatory note for administrative context, and labels the proposed workflow change as the team's operational judgment. Because the issue could have legal consequences, the briefing also names where qualified review is needed instead of presenting the analysis as a legal conclusion.

The artifact is useful even if the recommendation later changes. It preserves why each source was given a particular role and prevents a polished secondary summary from silently replacing the source with actual mandate.

Practice by changing the proposition

Give a learner one source and ask what it can support across several propositions. A company release might directly support "the company announced feature X on this date," provide qualified support for "feature X is available under these published conditions," and provide no independent support for "feature X improves team performance."

Then reverse the exercise. Give the learner a claim and ask for the source roles needed to support it: the source with authority to define the rule, the record that shows implementation, and an independent source capable of testing effects. This develops the habit of assembling complementary authority rather than searching for one impressive link.

A version-control drill is also valuable. Place a current policy, a superseded copy, and a commentary article in the same pack. Ask the learner to determine which version controls the stated date, how the old copy should be treated, and what the commentary adds. The result should include a version note, not just a preferred URL.

Assess source-authority behavior

A realistic assessment can ask a participant to produce a decision note from a mixed evidence pack containing an official record, a secondary explanation, two derivative summaries, an outdated copy, and an unsupported AI statement. Require the final note and its STANDING records.

Look for observable evidence that the participant:

Do not infer this skill from citation style, domain names, or the number of first-party sources. A source profile with explicit limitations is stronger evidence than a long bibliography with unexplained authority. Preserve opened-source history, source profiles, version comparisons, dependency notes, and revisions as the assessment evidence.

• matches a source's mandate or competence to a particular claim;
• identifies origin, method, version, jurisdiction, and status where relevant;
• detects when several sources share one evidence lineage;
• separates official description from independent evidence of effects;
• calibrates claim language to direct, qualified, or contextual support;
• seeks corroboration suited to the remaining uncertainty;
• corrects or removes claims when the asserted authority does not survive inspection.

Verify before relying on authority

For each decision-relevant claim, ask: "What makes this source entitled to speak to this proposition?" Confirm the source identity on the page, inspect the current version and status, review the stated methodology or record origin, and follow any cited upstream material. Search independently for corrections, superseding documents, and material conflicts.

Run a role test on the finished work. Labels such as rule, observation, interpretation, and recommendation should align with the source's actual role. If a vendor's description is being used as independent proof of an outcome, or a commentary article is standing in for controlling text, downgrade or replace it.

Source-authority judgment helps assign defensible weight to evidence. It does not establish truth, remove the need for relevance and corroboration, or provide legal conclusions. In a Two Prune assessment, the STANDING records and verification trail can contribute evidence about Evidence Navigation and Verification & Risk Control within the specific simulation. The result is an evidence-backed assessment signal, not a hiring recommendation or automated employment decision.

Sources

UK Home Office, Purpose, methodology and use of country policy and information notes

• U.S. Government Accountability Office, Government Auditing Standards 2024 Revision

Sources: [1] [2]