

A retouched aerial view, a voice message attributed to an officer, footage of a strike that never happened, an operation order passed off as a leak: Label4.ai analyses every piece of content, whatever its format, and shows what was made up, and where.
Analyses the images, voices, videos and documents your teams collect, and shows for each one what AI generated or edited, and where.
Fabricated content can steer an assessment, a decision or public opinion, and genuine evidence can be dismissed as fake. Knowing what is synthetic has become the first step of intelligence work and of countering information manipulation.
OSINT cells, intelligence services, teams countering information manipulation, and military communications.
The problem
Geolocating an image, dating it, cross-checking accounts: open-source intelligence methods remain essential, but they assume the content shows something that existed. Generative AI now produces images, voices, videos and documents that pass those checks.
A genuine satellite image where jets, fuel tanks and a convoy were added. The base is real, the build-up is not.
An order to withdraw, a confession or a threat, spoken in the voice of a real officer, cloned from a few public appearances.
What you hear
What we detectOne voice message, a genuine speaker and a cloned one.
An explosion, a destroyed building, footage filmed “on the spot”: the clip spreads within minutes and calls for a response.
A real document template, real names, and a few lines changed to pass off a false plan.
It has already happened
A deepfake of the Ukrainian president telling his troops to lay down their arms, posted after a news website was hacked.
Source: NPR, March 2022A generated image of an explosion near the Pentagon, convincing enough to briefly move US markets.
Source: NPR, May 2023between the release of a fake audio recording of a party leader and Slovakia’s 2023 parliamentary election.
Source: International Press Institute, 2023Our answer
Label4.ai adds one question to the checks you already run, and leaves the rest of your method as it is.
Social networks, messaging apps, satellite imagery, intercepts: your usual sources.
Geolocation, dating, cross-checking. Unchanged.
Was this content made up, in whole or in part? Checked in seconds, with its reasons.
Use it, cross-check further, or report information manipulation, with the reason written down.
The useful answer is not true or false. The analyst sees what was generated, what was edited and where, in a few plain sentences like these.
The same piece of content can be fully generated, edited in places or genuine, and each case calls for a different response.
A scene, a voice or a document that never existed.
Genuine content where a part was added, removed or replaced.
Content with no trace of generation or editing.
We do not rely on six-fingered hands, inconsistent shadows or poor lip sync, which each new generation of models corrects. What we read is left by the way the content was produced, and it is still there when the imitation is perfect.
How we do it
Already in use by public institutions.
See the technology in detailYour content
You give us a batch of content whose nature you already know, genuine, generated or edited, in every format you handle. We return a result for each one, and a plain statement of what we did not see. On your own servers if you prefer.
We reply to every request within two working days, under NDA if you prefer.