The new OSINT citation standard makes AI’s contribution part of the evidence record

An intelligence report may cite a commercial platform that collected material from elsewhere and used AI to classify or summarise it. The link can be correct while leaving the reader unable to tell which statement came from the original source and which was added later. As information passes through more services, its transformations become part of the evidence.

ODNI’s December 2, 2024 citation standard explicitly addresses public information, commercial information and AI-assisted products. The earlier March 2024 OSINT strategy sets out modernisation and professionalisation priorities. Open-source intelligence, or OSINT, derives from public or commercially available material used to address an intelligence need. These documents define practices and ambitions, not measured accuracy of an automated service.

A platform is not always the origin

A commercial interface may make discovery and access easier while remaining one step removed from the material being assessed. The user needs to know the original creator, relevant time and processing applied. Otherwise two platforms can appear to provide independent evidence when both obtained the same record from one source.

An AI-produced label is another transformation. The original document might mention an organisation while the model assigns a role or relationship. That derived judgement should not be cited as though the document explicitly stated it. A useful record preserves the source passage and identifies what the tool contributed so a reviewer can assess the relation.

This distinction is particularly important when the original material changes. A live webpage can be edited after a report is written; a social post can disappear. A citation pointing only to the latest version may no longer show what the analyst actually used. Preserving the relevant record connects the judgement to its evidentiary moment, rather than to a moving destination.

Recording an AI contribution does not require assuming the model’s explanation is a faithful account of its internal process. The practical requirement is to identify the tool and the output that affected the work, along with the source evidence. That allows a reviewer to separate collected information from automated interpretation without relying on an unverifiable reasoning narrative.

Sourcing can make automation easier to use

The counterargument is that detailed citations impose too much work and slow production. The burden is real, especially if staff must reconstruct provenance after the fact. Building those records into collection and processing can make them less expensive. The operational comparison should measure the total review effort, not only the time spent writing the first report.

A good automated workflow could preserve origins and transformations by default. That can let analysts focus on credibility and relevance rather than manually rebuilding a chain of links. The benefit is a design hypothesis, not an outcome established by the standard. It needs testing against actual sources and products, including those that change or disappear.

One proposed assessment would give a reviewer a finished product and ask them to recover the evidence for selected claims. Measure whether they can find the exact material, understand the tool’s contribution and recognise unresolved gaps. That evaluates the usability of the evidence record rather than merely checking whether the document contains footnotes.

Security constraints can limit which details are disclosed to a recipient. The answer is to preserve the appropriate accountable record and communicate the resulting limits, not imply a source is fully inspectable when it is not. The citation should help the consumer understand the basis of the assessment within the permitted access.

The standard therefore has a practical technology consequence. Vendors serving intelligence users need outputs that can carry origins, versions and transformations into downstream reporting. A polished analytic product without that structure may be harder to integrate than a less elegant one that a reviewer can reliably check.

A citation can expose where the judgement began

Consider an illustrative report that identifies a company as a supplier to a military organisation. The original source may show only that its equipment appeared in a photograph; a commercial service may add an entity label; an AI summary may infer a contractual relationship. One link to the service cannot establish all three steps. The reviewer needs to know which part is observation and which part is interpretation.

Recording those transformations makes correction less disruptive. If the entity label is wrong, the organisation can identify the dependent claims and revise them without treating every underlying record as unreliable. If the source itself changes, it can preserve the earlier version actually used. This is a practical reason for provenance: it provides a map of what an error affects.

The wider opportunity is for sourcing to become part of the data structure carried through an automated workflow. Analysts then receive a more usable basis for judgement instead of being asked to reconstruct it after generation. The standard’s significance is not a promise that citations eliminate error. It is a clearer expectation that the contribution of each source and tool remain visible enough for an accountable assessment.