What did Zheng Dai and David Gifford’s MIT CSAIL study, published in Nature Communications, discover about “attribution decay” in diffusion based AI image generators—how did their diffusion ensemble architecture test the effect of removing specific training images across datasets

What attribution decay showed

• With larger datasets, the counterfactual radius declined: removing a single image increasingly made little or no appreciable difference to a generated output.
• At sufficiently large scales, the same could hold not only for a particular image, but for all images by a given artist, or all photographs depicting a particular person: removing that whole category often did not appreciably change the generated sample.
• This is a finding of non-attribution in the tested counterfactual sense. If removing a unit does not change an output, that unit cannot be assigned causal responsibility for that output under the study’s measure.

Copyright implications

• The work could make a particular theory of infringement harder to establish in litigation against image-model companies—including Midjourney, OpenAI, or Microsoft—when the claim is that a disputed output was caused by a named artist’s works or copies that artist’s style. A claimant may face difficulty proving a direct, output-specific causal contribution if omission of the artist’s complete corpus leaves the output effectively unchanged.
• But it does not decide those cases or establish a legal rule. Copyright disputes can concern unauthorized copying during training, substantial similarity between particular outputs and works, access, market harm, contractual terms, or other issues beyond whether one item was individually necessary for one output.

Critical limits

• Not a no-memorization result: An output can be unattributable to any one training item under this deletion test while a model has nevertheless memorized material, can reproduce it under some prompts or sampling conditions, or generates an output substantially similar to a protected work. The paper tests causal sensitivity to removing units, not every form of memorization or infringement.
• Not yet a result about LLMs: The study specifically concerns diffusion models, which generate audiovisual media. Its ensemble/ablation technique and empirical scaling result have not thereby been demonstrated for transformer-based large language models.
• Not a replacement for other evidence: Courts and technical investigators would still need prompt-based reproduction tests, similarity and provenance analysis, training-data and system records, model audits, and applicable legal analysis. Attribution decay says that a single-source causal story may fail at scale; it does not show that copying cannot be proved by other evidence.