VK asks contractors whether they train AI on scans of student ID cards

A student discount on a music service looks like a minor marketing mechanic. But it has an engineering underside: someone has to check, manually or automatically, a huge number of paper student ID cards. Now VK, while changing its partner for this verification, has asked candidates an awkward question: do you plan to train AI on this data?

What exactly happened

The VK holding is choosing a new IT contractor for bulk verification of paper student IDs. The verification is needed so that users can get access to VK Music at a reduced price. As part of the partner selection, the company asked potential contractors to say whether they plan to use the data they receive to train AI models. The agenda also includes confirmation of a ban on such actions.

In other words, VK is not just checking whether a contractor can recognize documents. It is establishing in advance that verification data must not become raw material for someone else's models.

Why student IDs are valuable material

The experts cited by CNews believe that the collected array of student IDs is indeed highly valuable for computer vision systems. The logic is clear. These are real documents, shot on different cameras, in different lighting and from different angles. Universities issue student IDs in different ways: the forms, stamps, fonts and photos all differ. For a model learning to find a document in an image, read its fields and tell a genuine form from a forgery, this diversity is the most expensive resource. It is hard to assemble artificially.

There is also a flip side. Such documents, as a rule, contain people's personal data, and the question of whether they can be used for training is not only about competition but also about user trust.

One separate nuance from the article: training the new IT service could also benefit VK itself, because it improves how the service works. A model that sees more real examples makes fewer mistakes on non-standard IDs. So the ban is not an obvious gain but a deliberate compromise between quality and control over data.

What this means in practice

The source does not disclose which conditions will ultimately make it into the contract, or which contractors are bidding for the project. But the approach itself is worth adopting.

• Write the ban explicitly. If a contractor processes your customers' documents, photos or correspondence, the phrase "used only to provide the service" should include a separate clause on model training. Contract silence is often interpreted in the contractor's favor.
• Ask before signing, not after. VK asked the question at the partner selection stage, when it still has leverage. After the project launches, negotiating is harder.
• Track the data lifecycle. Find out where the scans are stored, for how long and who has access to them. Ask the same question about fine-tuning the contractor's internal models.
• Treat data as an asset. If your process generates an array of labeled or verified images, that is already a dataset, even if you never intended to create one.

For a product manager the takeaway is simple: any "routine" function such as discount verification generates data that is valuable for AI. For a developer: if you are a contractor, be ready to answer honestly what the received files will be used for. Clients are starting to ask about this directly.

Limitations

It is unknown from the article how the partner selection will end, whether the ban will be technically verifiable and how VK intends to monitor compliance. The experts' assessment of the dataset's value is an opinion, not a measured quantity. But the signal to the market is clear: terms on using data for AI training are becoming a regular part of a tender, not fine print at the end of a contract.