The researchers also tested broader removals: all works by one artist and all photographs of a particular person. Their results indicate that these removals can likewise produce little or no measurable change in generated outputs when the model and dataset are large enough. The experiments covered 24 diffusion ensembles and datasets ranging from 256 to more than 160,000 images.

This does not mean that an image played no role in the model’s overall training. It means that, for a particular output and under the study’s removal test, its individual causal contribution may be too small to detect.

How the diffusion ensemble worked

Testing this question directly is difficult. A conventional approach would retrain a model after removing a source image, then compare the new output with the original. Repeating that process for individual images, artists, or subjects would be expensive, and approximate methods may not remove every downstream effect of the data.

Dai and Gifford instead designed a diffusion ensemble: multiple smaller diffusion components trained on different, overlapping data subsets. To test a source’s influence, the researchers could disable the components that had seen that source and generate the corresponding counterfactual output without retraining the entire system.

They compared the ensemble method with 24 conventional diffusion models trained on the same data, reporting broadly comparable output quality by standard measures.

The approach gave the researchers an exact removal test within their architecture: if the output remained unchanged after every component exposed to a source was switched off, the experiment provided no measurable evidence that source had caused that output.

Why the finding matters for copyright lawsuits

Attribution decay could make a specific type of claim more difficult: the argument that a particular generated image was caused by, or copied from, one plaintiff’s image or an entire artistic corpus. If removing that material does not change the output in an exact counterfactual test, the test offers little support for assigning that output to the removed source.

That issue could arise in disputes involving image-generation services or model developers, including companies such as Midjourney, OpenAI, and Microsoft. But the study itself does not decide any company’s legal liability. It provides evidence about the difficulty of establishing causal links between individual training data and individual outputs.

Several separate questions remain open:

• How was the training data obtained? The study does not determine whether collecting or using a copyrighted work for training was lawful.
• Is the output substantially similar to a protected work? Visual similarity may be relevant even when causal attribution to one training file is difficult.
• Did the system memorize or reproduce an example? A model may sometimes reproduce training material, which is different from proving that every output is attributable to one source.
• Was the system configured to imitate an artist? Targeted prompts, fine-tuning, or other model behavior may provide evidence not captured by a single-image removal test.

The practical implication is that courts and investigators may need a broader evidentiary record: data-acquisition records, provenance information, direct similarity analysis, targeted-prompt testing, model-behavior evidence, and expert analysis. That conclusion follows from the study’s finding that one-image-to-one-output attribution becomes unreliable at scale; it is not a legal rule created by the research.

Why this is not a free pass for AI developers

Attribution decay should not be read as proof that diffusion models never copy. The study’s result is that an individual source’s measurable influence can become negligible as training datasets grow. Models can still memorize or reproduce training examples in some settings.

The scope is also limited. The work examined diffusion-based image generators, so its results should not automatically be extended to language models or other generative architectures.

The most accurate reading is therefore a cautious one: large diffusion systems may make individual-source attribution harder, even when their training data includes a particular image or artist’s work. That may weaken one route for proving responsibility for a specific output, while leaving broader questions about training practices, copying, imitation, and intellectual-property rights unresolved.

The researchers also tested broader removals: all works by one artist and all photographs of a particular person. Their results indicate that these removals can likewise produce little or no measurable change in generated outputs when the model and dataset are large enough. The experiments covered 24 diffusion ensembles and datasets ranging from 256 to more than 160,000 images.

This does not mean that an image played no role in the model’s overall training. It means that, for a particular output and under the study’s removal test, its individual causal contribution may be too small to detect.

How the diffusion ensemble worked

Testing this question directly is difficult. A conventional approach would retrain a model after removing a source image, then compare the new output with the original. Repeating that process for individual images, artists, or subjects would be expensive, and approximate methods may not remove every downstream effect of the data.

Dai and Gifford instead designed a diffusion ensemble: multiple smaller diffusion components trained on different, overlapping data subsets. To test a source’s influence, the researchers could disable the components that had seen that source and generate the corresponding counterfactual output without retraining the entire system.

They compared the ensemble method with 24 conventional diffusion models trained on the same data, reporting broadly comparable output quality by standard measures.

The approach gave the researchers an exact removal test within their architecture: if the output remained unchanged after every component exposed to a source was switched off, the experiment provided no measurable evidence that source had caused that output.

Why the finding matters for copyright lawsuits

Attribution decay could make a specific type of claim more difficult: the argument that a particular generated image was caused by, or copied from, one plaintiff’s image or an entire artistic corpus. If removing that material does not change the output in an exact counterfactual test, the test offers little support for assigning that output to the removed source.

That issue could arise in disputes involving image-generation services or model developers, including companies such as Midjourney, OpenAI, and Microsoft. But the study itself does not decide any company’s legal liability. It provides evidence about the difficulty of establishing causal links between individual training data and individual outputs.

Several separate questions remain open:

• How was the training data obtained? The study does not determine whether collecting or using a copyrighted work for training was lawful.
• Is the output substantially similar to a protected work? Visual similarity may be relevant even when causal attribution to one training file is difficult.
• Did the system memorize or reproduce an example? A model may sometimes reproduce training material, which is different from proving that every output is attributable to one source.
• Was the system configured to imitate an artist? Targeted prompts, fine-tuning, or other model behavior may provide evidence not captured by a single-image removal test.

The practical implication is that courts and investigators may need a broader evidentiary record: data-acquisition records, provenance information, direct similarity analysis, targeted-prompt testing, model-behavior evidence, and expert analysis. That conclusion follows from the study’s finding that one-image-to-one-output attribution becomes unreliable at scale; it is not a legal rule created by the research.

Why this is not a free pass for AI developers

Attribution decay should not be read as proof that diffusion models never copy. The study’s result is that an individual source’s measurable influence can become negligible as training datasets grow. Models can still memorize or reproduce training examples in some settings.

The scope is also limited. The work examined diffusion-based image generators, so its results should not automatically be extended to language models or other generative architectures.

The most accurate reading is therefore a cautious one: large diffusion systems may make individual-source attribution harder, even when their training data includes a particular image or artist’s work. That may weaken one route for proving responsibility for a specific output, while leaving broader questions about training practices, copying, imitation, and intellectual-property rights unresolved.