MWS AI Tests Multimodal Models on Detecting Forged Documents

MWS AI, a division of MTS Web Services, tested how well multimodal models (Visual Language Models, VLMs) can tell forged documents from genuine ones. The report appeared on CNews on 9 October 2026. The topic is timely: banks, insurers, marketplaces and HR departments deal with forgeries every day, and generative models have made them cheap. The available excerpt of the material contains no list of tested models, no sample size and no final percentages, so no ranking can be drawn up yet. What we can do is look at what this kind of testing measures and where its pitfalls lie.

What was tested and why

VLMs take an image together with a text prompt as input. For documents, this means the model receives a scan of a passport, certificate or contract and answers a question such as "are there signs of editing?" Previously, this task was handled by a chain of several components: OCR to recognize text, a set of rules to validate fields, and separate forensic models to analyze image artifacts. VLMs promise to fold all of this into a single call.

Forgeries come in many forms. Some simply redraw digits in an amount or a date. Others swap the photograph, and still others stitch together a background from several sources. The newest variant is a document generated entirely by a model, with plausible fonts, a stamp and a signature. Each type calls for its own signals, and VLM quality on them can vary widely.

Why independent tests are rare here

Public benchmarks for multimodal models mostly test document understanding: find a field, read a table, answer a question about the text. Authenticity checking works differently. The model has to notice that something is wrong with a document even though its content looks flawless. This is a different skill, and a high score on DocVQA-style datasets does not guarantee it.

What the publication tells us:

• Who tested: MWS AI (part of MTS Web Services).
• Model class: multimodal VLMs.
• Task: detecting forged documents.
• Date and source: CNews, 9 October 2026.
• Not disclosed in the available excerpt: the list of models, dataset size and composition, accuracy, false positive rate.

What to look for in the full report

First, the ratio of the two error types. A missed forgery costs money; a false alarm costs a customer and operators' time. An average accuracy figure that does not break out these two numbers says almost nothing.

Second, the mix of forgeries. If the sample contains only crude manual edits, the result will be optimistic. If it includes documents made with modern image generators, the picture will look harsher.

Third, language and format. Russian passports, 2-NDFL certificates and bank statements have their own layouts, and models trained mostly on English-language data may behave differently on them.

Fourth, robustness to prompt wording and scan quality. A phone photo taken in poor lighting breaks many systems that work perfectly on clean images.

It is also worth remembering that MWS AI develops its own AI-based products. That does not negate the value of the test, but the methodology deserves a careful read and, where possible, a cross-check against your own data.

What teams building anti-fraud systems should do

Do not hand the final decision to a VLM. A sensible role for it today is a first-pass filter that ranks documents by suspiciousness and leaves the most contested ones to a human. Build your own test set from real cases at your company, add synthetic forgeries from current generators, and run every new model version against it. Calculate the cost of each error type in money and set the threshold accordingly. And keep the classic checks: verification against government registries, file metadata analysis and validation of details. A multimodal model adds one more signal to these, but it cannot replace them yet.