How massive data pools remove the “smoking gun” of AI training

The legal landscape surrounding AI-generated content is undergoing a significant shift, as a recent MIT research paper challenges the notion that AI models steal specific works during training. The study, titled "Outputs of Generative Diffusion Models are Often Unattributable," demonstrates that as the size of training datasets increases, the contribution of any single image or artist becomes increasingly negligible.

This finding suggests that visual similarity is not the same as causal theft, potentially offering a legal defense for AI firms like Midjourney and OpenAI. The paper shows that in large-scale models, an artist's work becomes one drop in a vast ocean, making it statistically insignificant. This research provides a powerful shield for these companies, allowing them to argue that even if they had never seen a specific artist's work, their model would have produced a nearly identical result.

However, the paper primarily focuses on diffusion models used for image generation, leaving text-based outputs, such as those created by LLMs, outside its scope. While the scale argument may provide AI labs with legal immunity, the real test will be in how they present this research in court. Judges may not consider statistical evidence as conclusive proof of non-infringement, especially in cases where the output is strikingly similar to a specific artist's work.

For text-based AI, the issue remains more complex, as the collective "uncompensated use" of literary works may be harder to defend against copyright claims. Despite these technical advantages for image generators, the industry remains in a holding pattern, awaiting further developments on the legal front.