I never thought I’d be working as a detective, but as an AI Engineer at Fourthline, where I build the models responsible for catching fraudulent documents, that’s sort of what I’ve ended up doing. The scale of the problem is massive, and it's growing every day. The good news? So is our ability to fight it.  

The first question a lot of people ask about document fraud is how to spot the fake ones. Here, I’ll explore what document fraud actually looks like, how to detect it, and how human and AI-driven review work together to fight fraud effectively. 

Fake vs. cloned: how the terms are used, and what they actually mean 

A fake or counterfeit document refers to something built from scratch, where a fraudster creates a document that was never legitimately issued. Pretty self-explanatory. A cloned document implies something different: copying the digital data of a real, valid document that belongs to someone else, then reproducing a stolen identity's actual credentials. 

When people talk about a "cloned" document, they are typically referring to the digital layer: the data stored on the RFID/NFC chip embedded in modern passports and identity cards. That chip contains the same information printed on the document, such as your name, date of birth, photograph, and document number, albeit in a cryptographically secured digital format.  

Cloning, as we've established, extracts that data from a genuine document and copies it onto another chip. The assumption here is that if the digital data matches a real, valid identity, the document will pass verification. So, in a flow that validates only NFC chip data, without capturing video or an image of the physical document, cloned documents would become a meaningfully different and more serious threat category. 

Counterfeits and ID tampering: the more apt distinction 

At Fourthline, we validate the digital content of a document and its physical content simultaneously. The digital side includes the NFC chip data, while the physical side includes the MRZ, the Visual Inspection Zone, visible security features, and the photograph of the person. Because we cross-reference both, a cloned chip alone is not enough to fool the system. Even if a fraudster successfully copies the data from a real document's NFC chip, we can still assess whether the physical document it has been placed in is authentic.  

All of which this is to say: in the way our pipeline currently operates, cloning the digital data of a document does not constitute a distinct fraud type on its own. For the purposes of detection, the more operationally useful distinction is the one we use internally at Fourthline: we distinguish internally between counterfeits (created from scratch) and ID tampering (a genuine document whose content has been altered, rather than cloned). The detection challenge, and the red flags to look for, map onto those two categories. 

The scale of the problem 

Before getting into detection, it's worth addressing the scale of what we are dealing with. According to a 2025 threat assessment by Recorded Future, synthetic identity document fraud rose 300% in Q1 2025 compared to the same period the previous year.  

In our experience, a significant portion comes from organised criminal groups running high-volume, iterative attacks. These groups are submitting large numbers of documents across multiple attempts.  

Unfortunately, AI has made this significantly worse, particularly for static document images. Fraudsters can now generate highly convincing fake document images with relatively low technical skills. Deepfake attacks on selfies and biometric checks are already very common, but AI-generated document fraud is still growing.  

However, AI is also helping us catch criminals and stop them in their tracks.  

Spotting fraud: red flags across all document types 

Whether you’re reviewing an identity card, a driver's licence, or a passport, there are three key areas where fraud makes itself known: 

Font and character consistency 

Fonts are very often wrong on counterfeit documents, and they are often the easiest tell when reviewing. Every issuing country uses specific typefaces and character proportions in its documents, including in the MRZ. To someone who has reviewed the same document template hundreds of times, even subtle deviations quickly become obvious. 

Security feature continuity 

Genuine documents have security features that run across the entire surface: guilloché patterns, fine line work in the background, and sometimes holographic overlays that change appearance if you tilt the document. When someone alters a photograph or replaces a section of the document, those continuous features break, showing discontinuity, off alignment, or (on a digitally altered or AI-generated document) a compression artefact at the edit boundary. 

Dynamic behaviour under movement 

On a digitally altered or AI-generated document, text and other elements can become unstable under movement: characters blur, shimmer, or shift in ways that no physical document would produce. This is currently one of the clearest gaps in AI-generated document fraud. Fraudsters can now generate pretty convincing static images, but producing a realistic video of a manipulated document is still quite difficult to pull off.  

Counterfeits vs. ID tampering: which is harder to catch? 

When thinking about which type of document fraud is more difficult to detect, the answer depends on whether you are asking about human reviewers or automated models.  

Counterfeits are relatively easy for our models to flag. Structural errors, such as incorrect fonts, inconsistent security features, or weird character shapes are flagged with relative ease. But these same counterfeits can be harder for a human analyst to catch, especially if they’re under time pressure.  

ID tampering is the other way around; manual reviewers catch it much more effectively than our models do. A small, localised change (a single altered digit in a date of birth, for example) can slip past a model that is looking at a wider range of authenticity signals.  

Human and automated detection: why you need both 

At Fourthline, our analysts, AI and detection systems all work in symbiosis: they do their most effective work as a unit. The AI systems are trained on information fed to them by the analysts, and it is that human knowledge that keeps them improving. Their outputs are then fed back to the analyst when they detect concerning patterns that are difficult to detect with the human eye.   

The clearest example of where our AI systems do their best work is deepfakes. A well-made deepfake can be difficult to detect with the human eye. What our models look for are statistical patterns in the image data itself: inconsistencies in the way pixels are distributed, signals in the image that indicate it was generated rather than captured by a camera.  

Beyond deepfakes, we use a layered model architecture rather than a single classifier. Different models handle different signals: one reads and validates the character sequences in the MRZ, another assesses the visual authenticity of security features, a third looks for patterns associated with AI generation or digital manipulation. When these outputs are cross-checked for consistency, discrepancies that no individual model would flag in isolation become visible. 

We also have data points beyond the document itself that can contribute to a fraud signal when aggregated together. These are things a human reviewer looking at a document image simply would not have access to. 

Analysts only review the cases our models have already flagged as uncertain or suspicious. The straightforward cases are automated before they ever reach a human reviewer. At Fourthline, we never automate fraud rejection. If a model flags a document as fraudulent, it has to be reviewed by a human analyst before any decision is made. The model effectively gathers signals, maps the concerns, and sends it on the analyst for manual review. This is deliberate: we want to be completely certain something is fraud before rejecting it, and that certainty requires a human in the loop. 

Where fraud is heading next 

The threat landscape is expanding. Beyond individual fraud attempts, we are seeing organised networks selling complete identity packages (AI-generated photographs, manipulated document images, scripted liveness behaviour) through channels specifically designed to defeat KYC checks. These are purpose-built, iterative attacks, not opportunistic ones. 

AI has lowered the barrier to entry significantly, and I think the document side of this is still in its early stages. The biometric deepfake problem is already very large; document deepfakes are catching up. Synthetic identity fraud, where real and fabricated data are combined rather than a single real identity being copied, is a related and growing pattern that document-level checks alone cannot fully address. 

What keeps me engaged with this work, honestly, is that the technology available to us is the same technology available to the fraudsters. The gap between attack and detection shifts, closes, and opens again. My job is to make sure it keeps closing on our side. As AI makes it easier to fabricate identities, companies that verify and authenticate people's identities are becoming more important, not less. 

Davide Manco is a senior machine learning engineer at Fourthline. For a full overview of Fourthline's document and biometric verification capabilities, visit the document and biometric verification page.