Technology

The science behind the trust layer.

Detecting fakes now has to hold up outside the lab, every day, on real claims, real accounts and real calls, which is why it needs the scientific method more than ever. Label4.ai was built on that method, one deliberate choice at a time.

Contact us and challenge us
port_overview.jpgSuspect: manipulated
Aerial view of a container port, with two loaded container ships moored along the quay, a tugboat and the container yard
The same view, with the traces of editing in warm colours on the two ships, the tug and the quay
‹›
The imageThe traces

Scientific rigour, at industrial scale

Detection was born in research, and it still depends on it. What has changed is where it has to work: new generators keep coming out, and a detector that shines on a benchmark but not on your content is of no use to you. Label4.ai stands between the two worlds: our research team designs and trains its own detectors, in close contact with the CRIStAL laboratory at CNRS and the ARTISHAU team at Inria, and every model is tried on real cases before it reaches yours.

  1. With the labs

    A front-row watch

    The latest work on the traces that generation and editing leave in content, and on watermarking, discussed with the teams who lead it.

  2. At Label4.ai

    Our own detectors

    Designed and trained by our team, then tested on generators they have never seen.

  3. With your teams

    In production

    In the Anomaly Engine console and through the API, on the content your teams receive every day.

Patrick Bas, CNRS research director at CRIStAL, sits on our scientific committee, and Teddy Furon, Inria research director, is our scientific advisor. The committee, made of external researchers, reviews our methods.

Independent of generative models: we build none and depend on none, so we never grade our own work. InCyber Research Award 2026. Working with Sopra Steria, Capgemini and b<>com.

The choices that made us

There is no single breakthrough behind Label4.ai, only a series of choices that we hold to in every project.

1

No universal detector

A detector built for everything is tuned for nothing. Each project is set up for one use case: the content you receive, the fakes that target you, the action that follows.

2

A detection rate is not a result

A single percentage looks good on a slide, but nobody can act on it: it was measured on someone else’s content, at someone else’s threshold. We measure your false alarm and missed detection rates on your own content before you sign, so every result is one you can use.

3

False alarms and missed fakes, balanced for your case

Every detector trades one for the other. Flag too much and your team drowns in genuine content, flag too little and fakes go through. Where the balance sits depends on what a missed fake costs you, so we set it with you.

11fakes caught, out of 12
1fake missed
6genuine pieces your team checks for nothing

The same 400 pieces of content, 12 of them fakes, read at three thresholds. Illustration.

4

No detection of AI-generated text

Text keeps no physical trace of how it was made: nothing in the words themselves can reliably establish that a machine wrote them. A style can be spotted in a given context, but a person can imitate it or simply write that way, so false alerts cannot be kept under control, and writing with AI has become everyday practice anyway. It only makes sense where the human hand is the whole point, in literature for instance. Everywhere else, we focus on content that carries measurable traces: images, documents, voices and videos.

Every fake leaves a trace.

A fake is not a matter of taste, nor of what the content says. It is content that carries an anomaly against what you are entitled to expect from it.

A photo

What you can expect
The grain of one sensor and one lens, the same across the whole picture.
What we find
A region with regularities no camera produces.

A document

What you can expect
One export from one piece of software, with a single compression history.
What we find
One line that went through a second compression history.

A voice

What you can expect
A microphone, a pitch, a timbre.
What we find
A passage with frequencies no microphone produces.

We look for that gap in four places at once, with readings adapted to each medium, then settle on one conclusion.

  1. 1

    The metadata

    What the file records of the device and the software that touched it, read for what it can confirm and never stretched beyond it.

  2. 2

    The generator’s signature

    Content credentials and the invisible marks generators place in what they produce. When one is present and readable, it names the tool that made the content.

  3. 3

    The traces of generation

    The statistical regularities a model leaves in everything it produces, which the eye and the ear miss. They differ from one medium to the next: the grain and compression of a picture, the pitch and the timbre of a voice. Most answers come from here, and these traces are richest in the original file: every re-encoding, crop or screenshot wears some of them away, which is why we plug into your incoming flows.

  4. 4

    Local manipulations

    A genuine piece of content changed in one place, by hand or by a model: a figure rewritten on a document, a dent added to a photo, a few words spliced into a recording. We look for the area that does not match the rest of the file: a different compression history, a break in the grain, a passage whose frequencies do not match the rest of the recording.

We look at how content was made and leave what it says to you: whether a document tells the truth or a total adds up stays with your own controls.

A trust layer, piece by piece

Label4.ai is a trust layer: one layer of defence, designed from the start to add to the ones you already have. Here is what it is made of.

  1. 1

    The content, and only the content

    We read what the file itself carries, not who sent it or from where. That makes us agnostic to your channels and tools, and complementary to every check that looks elsewhere.

  2. 2

    No generative AI judging generative AI

    Our detectors are deterministic models trained in-house, with no LLM involved: they measure instead of giving an opinion, and the same content, analysed twice, gets the same answer twice.

  3. 3

    On your premises when needed

    Banks, insurers, public bodies: many cannot let their content leave. So the whole stack also installs inside your infrastructure, with no outbound connection, and runs the same models as our cloud.

  4. 4

    Always up to date

    We watch every new generative model as it comes out, test ours on generators they have never seen, and update them continuously, because the next model will not wait for anyone.

  5. 5

    A brick, not a replacement

    Built from day one to plug into an existing stack, through the Anomaly Engine console or the API, next to your identity checks, your rules and your teams.

  6. 6

    Because only layers hold

    No single control stops every fake. A defence that works is made of several layers, each catching what the others miss, and ours is the one that reads the content itself.

Watermarking, our other expertise

Detection and watermarking are the two sides of the same trust layer: one finds what gives a fake away, the other places an invisible mark in content when it is produced, so anyone can later check where it comes from. Very few teams in the world master both. Our watermarking expertise also lets your organisation prove where its own content comes from.

1

Marking and detecting

A mark only proves where content comes from when the content carries it. For everything else, detection takes over. We do both, so a single analysis reads the mark when it is there and looks for traces of fabrication when it is not.

2

An expertise in content authenticity

Our scientific advisors have worked on the security of watermarks for more than twenty years, alongside the ARTISHAU team at Inria, our research partner. That expertise serves the whole trust layer: knowing how a mark holds and how it gets attacked is also knowing how to spot content that has been tampered with.

3

Where the rules are written

Article 50 of the AI Act requires providers of generative AI to mark what their models produce. As the only French company specialised in watermarking, we took part in drafting the European Code of Practice that sets out how, we signed it, and we work alongside the French authorities on content transparency.

Every mark is a balance between staying invisible, surviving what content goes through, resisting forgery and how much information it carries. There is no ideal mark in general, only the ideal mark for a given use: we design it with you, and train its robustness against the transformations specific to your distribution channels.

Every image or voice the model produces is marked as it is created, without a visible change, and the mark has to survive edits, crops and screenshots.

How a mark is made

  1. 1

    In the signal

    The mark is spread across the pixels or the sound, below what the eye and the ear perceive, so it holds up through compression and everyday edits.

  2. 2

    At generation

    The generator places the mark while it creates the content, so every output carries it from the moment it exists. Model providers use our solution in our cloud or on their own servers: it plugs into their existing pipeline when needed, and was designed from the start to adapt to it.

  3. 3

    A signature or a message

    A mark can simply signal its presence, “this comes from us”, which leaves all the room for invisibility and robustness. It can also carry a real message: encrypted metadata, an identifier, a source, a date, a version that leads back to the original. In between, its capacity is set by what you need to retrieve.

  4. 4

    Secured by a key

    Without the secret key, nobody can read the mark or forge one.

See what your content reveals

Send us content you already know the answer to, and see what Label4.ai finds in it.

Contact us and challenge us

The technology at work inInsurance Claim EvidenceDocument Fraud DetectionRefund & Delivery ClaimsIdentity Fraud DetectionDeepfake & ImpersonationMedia AuthenticityDefence & Intelligence