If an AI image generator can turn out something in the style of your work, can you prove it actually used your work to do it? A study published last month in Nature Communications suggests that, surprisingly often, the answer is no — and the reason is not that the evidence has been hidden. It is baked into the way these models learn.

Zheng Dai and David K. Gifford, at MIT's Computer Science and Artificial Intelligence Laboratory, call the effect "attribution decay." Their paper, "Outputs of generative diffusion models are often unattributable", was written up by MIT in August and drew fresh attention this week when The Art Newspaper tied it to the copyright fights now moving through the courts.

What they did

The obvious way to test whether a particular image mattered to a model's output is brutally expensive: pull that image out of the training data, retrain the whole model from scratch, and see whether the output changes. Do that for millions of images and you would never finish.

Dai and Gifford built a shortcut. They designed an architecture they call a "diffusion ensemble" — rather than one monolithic model, many smaller models, each trained on a different slice of the data. That structure let them cheaply simulate "what if this example had been left out?" across a huge number of scenarios without retraining anything from scratch.

What they found

When the training set is large, the researchers report, you can usually remove any single image — or any single creator — without meaningfully changing what the model produces. The influence that shapes a given output is spread so thinly across so many examples that no single one is load-bearing.

"When you train on large data sets, there is no piece of data that you can take out that significantly alters the image," Dai said, "and that [leads] to this idea of unattributability."

The uncomfortable corollary is that a model can reproduce the recognisable style of a specific artist even when that artist's work was not in the training data — because the style is reachable through countless other images that echo it. Resemblance, in other words, is not proof of use.

Why it matters

This lands squarely in the middle of the AI copyright wars, and it cuts both ways — which is worth being honest about.

For AI companies it is close to a gift. If a plaintiff cannot show that their specific image causally shaped a specific output, one of the central theories of infringement gets much harder to run in court. For artists it is bleak: the very diffusion that makes these models good is what makes individual authorship undetectable after the fact.

But it is worth being careful about what the result does not say. Attribution decay is about whether one image can be traced inside one output. It says nothing about the separate, prior question of whether it was lawful to ingest billions of copyrighted works to build the training set in the first place. A model can be trained on your work and, at the same time, be impossible to show as having used it in any particular picture. Those are different arguments, and this paper only settles the narrow one.

What it leaves behind is an odd new fact about machine creativity: these systems can be derivative and technically unattributable at the same time. The courts, and the lawmakers behind them, will have to decide what to do with that.