How to Build an AI-Assisted OSINT Briefing: From Monitoring Results to Analyst Review

An open-source intelligence (OSINT) briefing has to deal with large volumes of data, sources of uneven quality and the limits of verification. Generative AI can take on collation, translation and drafting, but verification and judgment remain the analyst's job. This guide covers the inputs to an AI-assisted OSINT briefing, how they are processed, the points that must be reviewed by people, the risks to watch when using AI, and a sample output.

How should work be split between AI and analysts?

Inputs: where does the briefing's data come from?

There are four common kinds of input, each with its own limits:

• Monitoring results: Alerts and article lists produced by a media and social monitoring platform from its keyword and source settings. Coverage depends on those settings; absence from the results does not mean there is no relevant discussion.
• Public content from social platforms: Public posts and comments obtained through a data service or your own collection. Content may later be deleted or made private, so record the URL and collection time when you collect it.
• Internal background material: Past reports, background on people and organizations, and earlier assessments. They add context, but old assessments may be out of date, so cite them with their dates.
• The analyst's question: What the briefing must answer, plus the time range, platforms and languages. State the question first and use it to check whether the AI summary addresses it.

Processing: from raw data to a reviewable draft

The Berkeley Protocol on Digital Open Source Investigations, published by the UN Human Rights Office (OHCHR) and the Human Rights Center at the University of California, Berkeley, School of Law, divides an open-source investigation into six phases: online inquiry, preliminary assessment, collection, preservation, verification and investigative analysis, and recommends documenting the work done in each. The Protocol concerns investigations of violations of international criminal, human rights and humanitarian law; in our view its phases and record-keeping practice are also a useful reference for everyday OSINT briefings. The five points below are this article's suggestions for AI-assisted briefings, not the Protocol's six phases or an AI workflow it prescribes:

1. Define the question and scope first

Analysts decide the question, time range, platforms and languages, and AI works only within that scope. A clearly written scope also makes it easier to explain later which data was not examined.

2. Collect and preserve

Keep the original content, URL and collection time; where needed, save a screenshot or file and compute a hash value so you can later confirm the file has not changed. Comparing hashes only shows whether a file has changed since its hash was recorded; it does not by itself establish that the content is true. An AI summary is no substitute for preserving the original material either.

3. Cluster, de-duplicate and build the timeline

AI can help group reposts of the same content, order them by time and flag the earliest version found online so far, but the grouping and timestamps still need checking. The Berkeley Protocol advises saying “the first copy found online” rather than “the first copy online”, because the original may have been removed or may have been shared first through closed channels such as email, private messages and closed groups.

4. Verify: assess the source and the content separately

The Berkeley Protocol divides verification into source analysis, technical analysis and content analysis. Source analysis covers provenance, credibility (for example, posting history), independence and impartiality, the specificity of claims, and the time between an event and its documentation; technical analysis examines the file itself and its metadata; content analysis examines the imagery and text themselves and compares them with other independent sources. Assess source and content separately: a credible account can repost false content, and an unknown account can post genuine footage. AI can list leads to check, but analysts should reach the verification conclusions.

5. Draft: write facts, judgments and open questions separately

When asking AI to draft, require a source for every sentence and present confirmed facts, analytic judgments and open questions in separate sections. Teams should agree in advance on a fixed set of likelihood terms (for example: almost certain, likely, possible, unlikely) and label confidence separately, so the same word does not mean different things in different briefings.

Which points must be reviewed by people?

• Scope: Analysts decide which platforms, languages and time range to cover.
• Verification conclusions: Whether an item is true or false, or an old image reposted, is decided by analysts based on the verification work.
• Citation checks: Open every source in the draft and confirm it actually supports the sentence, not just the topic.
• Judgments about specific actors: When attributing content to a specific account, organization or person, or assessing whether accounts are acting in coordination, state the supporting evidence, alternative explanations and unresolved limitations; if the evidence is insufficient, leave it as an open question.
• Before sending: Check that likelihood terms are consistent, keep only the personal data that is necessary, and state the data coverage and cut-off time.

What risks should you watch for when using AI?

Confabulated content and false citations

The Generative AI Profile (NIST AI 600-1), published by the US National Institute of Standards and Technology (NIST) in July 2024, calls this confabulation: a system confidently states erroneous or false content, and may also produce confabulated logic or citations that purport to justify its answer. The profile recommends reviewing and verifying sources and citations in outputs, both before deployment and in ongoing monitoring. When checking a citation, confirm not only that the link opens but that its content actually supports the claim.

Collected content can carry instructions

The OWASP Top 10 for LLM Applications (2025) ranks prompt injection first. Indirect prompt injection occurs when content in external inputs, such as websites or files, causes an LLM to behave or respond in unintended ways. When LLMs process OSINT material, they read large amounts of untrusted external content, and a single post or web page may contain text telling the model to ignore its instructions. OWASP's recommended measures include segregating external content and marking it as untrusted, giving the model only the least privilege it needs, and requiring human approval for high-risk actions.

Volume is not credibility

When reviewing a summary, check whether discussion volume has been mistaken for evidence. Repost counts show how far something has spread, not whether it is correct, and large amounts of similar content may come from coordinated accounts. When ranking items, show volume and verification status separately.

Translation and online slang

Machine translation can misread slang, sarcasm, wordplay and regional usage. Quote foreign-language content with the original text, and have key sentences checked by someone familiar with the language and its regional usage.

Over-reliance on automation

The same NIST profile also describes automation bias: excessive deference to automated systems. If reviewers read only the summary and never open the sources, human review exists in name only. Ask reviewers to mark which citations they have checked in the draft, so the review leaves a real record.

Sample output: one verification item

The following is a fictional example showing how facts, judgments and open questions are presented separately: