Provenance

Provenance is the English word for origin. What is meant is a seamless record of where something comes from and what has happened to it. This has been known in the art world for centuries: an expensive painting comes with a list of all previous owners. If this list is missing, the painting is suspect, no matter how genuine it looks. In technology, this idea is applied to files and datasets. A photo, an audio recording, or a spreadsheet gets an accompanying documentation: created by whom, when, with which device or program, and afterwards edited by whom.

Why origin suddenly became a problem

Until a few years ago, a photo was a fairly reliable piece of evidence. A forgery cost time, skill, and often expensive software. Today, image generators produce shots that look real in seconds. This means the image itself loses its evidentiary value. The question shifts: no longer “Does this look real?” but “Where does this come from?”

This affects concrete areas. News editorial teams have to decide whether to air a video from a war zone. Insurance companies receive submitted damage photos. Courts assess digital evidence. Everywhere, it helps if a file brings its own history of creation with it.

A second area concerns the training data of AI models. Companies are being sued for allegedly having used copyrighted texts and images. Without clean provenance documentation, a company often cannot even prove what ended up in its training. The EU regulatory framework for AI now also requires information on this. Provenance is therefore not just a technical gimmick, but a legal obligation.

Signatures, metadata, and their weaknesses

The most widespread approach is called C2PA. Behind it stands a coalition of companies such as Adobe, Microsoft, Sony, and Nikon. The idea: a camera or program attaches a kind of digital information sheet to the file when saving. This contains the timestamp, device, and all subsequent editing steps.

So that nobody can secretly rewrite this sheet, it is cryptographically signed. A cryptographic signature is a computational procedure bound to the content. If someone changes even a single pixel, the signature no longer matches, and verification programs raise an alarm. Falsified information can be detected this way.

The major weakness lies elsewhere. Such additional information is easy to remove. A screenshot, an upload to a social network, or simply resaving a file is often enough. Provenance therefore only proves that something is genuine, not that something is fake. A file without provenance data is initially just unverified. As a complement, watermarks are used, which are embedded as tiny patterns within the image itself and survive resaving. Both methods together are more robust than either one alone.

From the phone camera to the model card

In everyday life, provenance is increasingly encountered inconspicuously. Newer cameras from Leica and Sony can sign shots directly. Image generators such as those from OpenAI or Adobe mark their outputs. Platforms like LinkedIn or TikTok display a notice on some content indicating that it was generated with AI.

In the AI industry itself, the term appears in so-called model cards and datasheets. These are profiles that accompany a model or dataset. They describe sources, languages, licenses, and known gaps. Anyone wanting to deploy a model within a company today often demands such documentation.

A common misconception: that provenance is a detector that recognizes AI content. It is not. Detectors guess based on anomalies and are regularly wrong. Provenance, on the other hand, relies on proof that is already supplied at the moment of creation. In the long run, this is the more reliable path, but only if enough devices, programs, and platforms participate.