Attribution Crisis

When a photo, a text, or a voice message surfaces, almost always the same question arises: Who made this? The answer to this question is called attribution. For a long time, it was often achievable, because every recording, every handwriting, and every attack on a computer left typical traces. Programs that generate images, speech, and text themselves have since made such traces forgeable or entirely worthless. The attribution crisis describes exactly this state: there is more content than ever before, but fewer and fewer reliable statements about where it comes from. Courts, newsrooms, authorities, and ordinary users alike are affected by this.

What collapses without reliable authorship

A great many rules of our society presuppose that one can name an author. Copyright law protects whoever created a work. An insult is only punishable if it is clear who uttered it. School grades, too, rest on the assumption that the work comes from the person being examined. If attribution falls away, these rules lose their anchor point.

It becomes especially delicate in politics and in attacks on computer systems. When one country accuses another of a hacking attack, everything hinges on proof of origin. Attackers today can plant traces that point to someone else. A false attribution can then have political consequences that are barely reversible.

There is a second, less obvious harm. When forgeries become easy, any genuine piece of evidence can be dismissed as a forgery. Experts call this the liar’s dividend: whoever is caught simply claims the recording was fabricated. The attribution crisis thus undermines not only false content, but also trust in genuine content.

Why digital traces are so easily erased

In the past, files carried many involuntary clues with them. A camera stores the model, the time, and sometimes the location in what is called metadata, the additional information of a file. Every camera model also produces a tiny, characteristic noise pattern in the image. Such features functioned like a fingerprint. In an image that a program has completely recalculated from scratch, this fingerprint simply does not exist.

Efforts are therefore being made to actively build in provenance rather than search for it after the fact. One approach is the watermark: a pattern invisible to humans that a program embeds into an image or text while generating it. A second approach is cryptographically signed provenance data, as envisioned by the industry standard C2PA. Here, the camera itself already attaches a tamper-proof signature to the file, and every subsequent edit is logged.

Both approaches share a common weakness. A screenshot, a compression, or a crop often removes the watermark and signature completely. And tools that do not attach provenance data are freely available online. Therefore, the absence of a signature proves nothing, while its presence is at least a strong indication.

From homework to the news cycle

In everyday life, the term is encountered most often in schools and universities. Software meant to detect AI-written texts regularly gets it wrong and even suspects honest students. Many universities have therefore abolished such programs again. Instead, they rely on oral examinations and on assignments in which the process of creation is documented.

In the news, the topic comes up with fabricated videos of politicians, the so-called deepfakes. Before elections, audio recordings regularly circulate whose authenticity cannot be clarified quickly enough. By the time a verification is completed, the clip has usually already reached millions of people. It is precisely this time gap that makes the crisis practically effective.

A common misconception is that this is only about perfect forgeries. In fact, mere doubt is enough to devalue evidence. News agencies, banks, and insurance companies are responding to this with a change in strategy. They check less the file itself and more the chain of sources behind it: Who received the material when, from whom?