A weak OSINT workflow often starts too fast.
Someone sees a claim, opens a search engine, checks a social platform, runs a reverse image search, asks an AI tool for help, and begins collecting fragments before the real question has been defined.
That is how noise enters the investigation.
The first step should be slower.
Before asking whether a claim is true, ask what the claim is made of.
Take a sentence like this:
This video shows police using tear gas during a protest in Madrid on 10 June 2026.
It looks like one claim. It is not.
It contains several smaller claims:
the content is a video;
the actor is police;
the action is the use of tear gas;
the event is a protest;
the place is Madrid;
the date is 10 June 2026.
Each part needs a different kind of evidence.
A timestamp may help with publication timing, but not necessarily with recording time. A street sign may help with location, but not with the date. A local news report may confirm that a protest happened, but not that this specific video shows that protest.
This is where AI can be useful in OSINT.
Not as a truth machine. Not as a source. Not as a shortcut around verification.
AI is useful when it helps you slow down the claim.
You can ask it to separate entities, actions, time, and place. You can ask it to turn each part into verification questions. You can ask it to suggest source types: maps, archives, official pages, local media, social platforms, weather records, earlier uploads, or public documents.
Then the actual work begins.
You still have to search manually. You still have to check primary sources. You still have to decide whether a screenshot is enough, whether a report supports the specific claim, whether two sources are independent, and whether your conclusion deserves high confidence or only cautious wording.
The useful role of AI is not to close the investigation.
It is to expose the gaps before you pretend they are closed.
A simple prompt to use before you start
Break this claim into separate verifiable components. For each component, explain what type of evidence would support it, what would not be enough, and what source types should be checked first.
Claim: [paste the claim]
Then edit the result.
Remove weak suggestions. Add local sources. Add language variations. Add the source types that AI may have missed. Treat the output as a planning layer, not as evidence.
The difference matters.
An AI-generated investigation plan can help you ask better questions. It cannot tell you what happened.
The language of uncertainty
A good OSINT conclusion does not force certainty where the evidence does not support it.
Useful wording includes:
verified;
likely;
consistent with;
unverified;
false or misleading.
Those are not stylistic choices. They are analytical boundaries.
If a video matches a location but the date is still unclear, the finding is not fully verified. If a username appears across two platforms, that may be an indicator, not an identity confirmation. If a source has not been found, that does not automatically make the claim false.
Good OSINT protects the distance between evidence and conclusion.
AI can help maintain that distance if you use it as a reviewer:
Review this evidence log. Identify weak assumptions, missing source types, possible false positives, and claims that are not yet supported. Do not decide whether the original claim is true.
That last sentence matters.
Do not let the tool become the judge.
The full workflow
I published a practical guide on ProjectOSINT that turns this idea into a step-by-step workflow:
write the claim as one sentence;
separate entities, actions, time, and place;
turn each part into verification questions;
build a source map;
generate search queries;
test them manually;
keep an evidence log;
use AI to identify gaps;
assign confidence levels;
write the conclusion with limits.
Read the full guide here:
How to Turn a News Claim Into an OSINT Workflow With AI
The main point is simple: AI should not make OSINT faster at the cost of evidence.
It should make the workflow more explicit.
Break the claim. Map the sources. Log the evidence. Name the uncertainty.
That is where verification starts.
If this is useful, share it.
This is the weekly selection. But it’s not the only one.
If you’d like to read more: → full articles on the website
If you’d like to get the latest updates first: → Telegram