You don't label the default

On August 2nd, Europe's new AI transparency rules kicked in. Deepfakes must be labeled, and AI-generated content will become more transparent. Sure, having to stick an AI logo on your work is annoying and doesn't look good. Still, it's a good law because it's desperately needed. It's aimed at deception, and AI enables deception on an industrial scale. So far, so reasonable.

Here's what makes it more interesting to me: In blind tests, people often rate AI-generated content perfectly well. Sometimes better than human work. Then you tell them where it came from, and the rating collapses. Which is the famous Pepsi Challenge, more or less. Maybe someone still remembers. In the seventies, Pepsi kept beating Coca-Cola in blind tests. When the brands were revealed, the preference shifted. That's what branding does.

Now the other side: A study published in PNAS last year showed that language models seem to have a similar reflex (link in the comments). Asked to choose between descriptions written by people and descriptions written by other models, they leaned consistently toward the machine-written ones. With one difference: no one told them which was which. The signal was in the text.

Which raises a question: A large and growing share of what sits on the web is machine-generated, machine-edited or machine-replicated. An even larger share of what gets retrieved, ranked and summarized is handled by a machine before a person sees anything at all. But where does this lead if the machines that retrieve content prefer content generated by other machines?

A lot of people in our industry are still counting on the human side of this. "Authenticity wins in the end. People will keep preferring the human thing." My guess: the human bias doesn't survive. Put an AI label on everything and eventually the label becomes wallpaper, because habituation eats the signal. And then we are back in the blind test, where AI already has the edge. The machine bias, on the other hand, doesn't depend on anyone getting used to anything. It can happen on every retrieval.

Human work doesn't get rejected because of this. It just gets found less often. And people only get what comes their way, which is, in fact, less and less human-made. Still, it won't disappear. But it may become what vinyl became: not the standard, but a smaller, more expensive category, and one rather proud of itself.

At that point, labeling AI content may start to feel backwards. Because you don't label the default. You label the rare stuff. Human-made.