Aerial imagery, air baseOpen source
Aerial view of an air base with fighter jets, helicopters, a fuel depot and a convoy of trucks
12/11/2026, 06:40Capture data: inconsistent
AI tracesManipulation traces
Aerial imagery, air baseOpen source
12/11/2026, 06:40Capture data: inconsistent
What we readWhat the eye sees
Logged to the intelligence fileAnalysis kept, reason written
ConclusionSuspect
Justifications
Manipulation traces, four areas
AI tracesHigh
Capture data inconsistent with the scene
No generator signature found

Image, voice, video, document: with generative AI, anything can be made.

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.

What it does

Analyses the images, voices, videos and documents your teams collect, and shows for each one what AI generated or edited, and where.

Why it matters

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.

Who it’s for

OSINT cells, intelligence services, teams countering information manipulation, and military communications.

The problem

We used to check where a source came from. Now we check what it contains

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 retouched aerial view

A genuine satellite image where jets, fuel tanks and a convoy were added. The base is real, the build-up is not.

A voice message attributed to an officer

An order to withdraw, a confession or a threat, spoken in the voice of a real officer, cloned from a few public appearances.

Waveform of a voice message with two speakersWhat you hear
What we detect

One voice message, a genuine speaker and a cloned one.

Footage of a strike that never happened

An explosion, a destroyed building, footage filmed “on the spot”: the clip spreads within minutes and calls for a response.

An operation order passed off as a leak

A real document template, real names, and a few lines changed to pass off a false plan.

It has already happened

Fakes that have already spread

2022

A deepfake of the Ukrainian president telling his troops to lay down their arms, posted after a news website was hacked.

Source: NPR, March 2022
2023

A generated image of an explosion near the Pentagon, convincing enough to briefly move US markets.

Source: NPR, May 2023
2 days

between the release of a fake audio recording of a party leader and Slovakia’s 2023 parliamentary election.

Source: International Press Institute, 2023

Our answer

Where Label4.ai fits in your analysis

Label4.ai adds one question to the checks you already run, and leaves the rest of your method as it is.

Collection

Social networks, messaging apps, satellite imagery, intercepts: your usual sources.

Your existing checks

Geolocation, dating, cross-checking. Unchanged.

Label4.ai analysis

Was this content made up, in whole or in part? Checked in seconds, with its reasons.

Decision

Use it, cross-check further, or report information manipulation, with the reason written down.

What was made up, and where

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 image is genuine, but four areas were added: two groups of aircraft, a fuel depot and a convoy.
  • The voice in this message carries the signature of a speech synthesis model from its first sentence.
  • The document uses a genuine template, but three lines were rewritten after it was created.

Knowing what is synthetic changes how content is read

The same piece of content can be fully generated, edited in places or genuine, and each case calls for a different response.

Fully generated

A scene, a voice or a document that never existed.

What to doSet it aside, and trace those who spread it.

Edited in places

Genuine content where a part was added, removed or replaced.

What to doKeep what is real, isolate what was changed, and ask why.

Genuine

Content with no trace of generation or editing.

What to doRely on it, including against those who call it fake.

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

Sovereign analysis, across every format

  • One tool for every format Images, videos, voices and documents go through the same analysis and return results in the same form, so a whole file can be cross-checked.
  • Installed on your premises, offline Label4.ai runs on your own servers, including on an isolated network. No data leaves, and our models depend on no generative AI provider.
  • A conclusion that holds up The same content, analysed twice, gives the same report twice, with its reasons written down, so it can go into a file.

Already in use by public institutions.

See the technology in detail

Your content

See what your own sources reveal

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.