Midjourney and other AIs are becoming too big to be caught red-handed for plagiarism.

Since 2023, illustrators have been trying to get Stability AI and Midjourney convicted for training their models on their works without authorization. Their best hope rested on a simple idea: tracing a generated image back to the work that fed it, in order to establish the copying.A study published this Tuesday in Nature Communications has just methodically closed that door.

Erasing the Mona Lisa is no longer enough to make it disappear

The two MIT researchers behind the work tackled the problem experimentally. Their method, called ablation, consists in removing specific works from the training data, then observing whether the model still produces the same result. The verdict is surprising: on a sufficiently large corpus, removing the Mona Lisa, or even all of Leonardo da Vinci, does not prevent the model from regenerating the image or imitating the style.

The researchers dubbed the phenomenon “attribution decay”: the larger the volume of data becomes, the less a generated image can be linked to a specific input. Beyond a certain threshold, attribution can become utterly impossible. A drop of ink ends up in a glass of water; in a lake, it no longer exists. Large image generators operate on the scale of a lake. The finding does not exonerate the models: cases of pure memorization do exist, and German judges have already examined outputs from music models that were manifestly learned by heart. It simply shifts the line: flagrant copying remains punishable, while statistical evidence of plundering evaporates.

A concrete defense for the giants of the image

The consequence will first be seen in the courts.In the case brought by illustrators against Stability AI and Midjourney since 2023, the plaintiffs are specifically seeking disclosure of the training datasets in order to look for evidence of unauthorized copying. If similarity between a generated image and an artwork can no longer demonstrate copying, this strategy risks collapsing. The Cornell legal scholar who commented on the study says so outright: attribution will fail for the models that matter, and judges will have to look for their evidence elsewhere.

The study even implicitly sketches a cynical defense strategy: it would suffice to bulk up enough for no output to be attributable to anything. Bulk up in order to absolve yourself, in short. The authors themselves wonder what comes next: how can creators be compensated when what comes out of the model can no longer be linked to anything that went into it? Europe has approached the problem from the opposite end. The AI Act requires general-purpose models to publish a summary of their training data—transparency at the input rather than proof at the output. The 2019 Copyright Directive also allows rights holders to object to the mining of their works. For a French illustrator, the battle will therefore be less about the similarity of images than about what models must declare before they exist.

👉🏻 Follow tech news in real time: add 01net to your sources on Google and subscribe to our WhatsApp channel.