FOLLOW US

AI Amnesia may be an emerging problem for researchers, artists and regulators. They are all trying to determine where generative artificial intelligence systems get the information, images and styles reflected in their output.

New research from scientists at the Massachusetts Institute of Technology suggests that as diffusion models are trained on increasingly large datasets, it can become harder — and sometimes impossible — to identify a specific piece of training data responsible for a generated result.

The findings were published Aug. 18 in Nature Communications. There is a challenge to the basic assumption behind efforts to understand generative AI. Researchers cannot always trace an output back to the data that influenced it.

Researchers Zheng Dai and David K. Gifford of MIT's Computer Science and Artificial Intelligence Laboratory found that the connection between an individual training example and a generated image can effectively disappear as the amount of training data increases.

AI Amnesia and the Attribution Problem

The researchers call the phenomenon “attribution decay.” Their study found that the more data a diffusion model is trained on, the less attributable its generated samples become.

Diffusion models are widely used for generating images, video and audio. They also have applications beyond entertainment and creative work, including scientific research involving protein structure modeling and therapeutic development. These systems learn statistical patterns from enormous collections of data and use those patterns to generate new material.

To investigate whether individual training examples could be identified as causes of generated content, the MIT researchers used a technique called ablation. Rather than retraining an entire model each time, they examined what happened when specific portions of training data were effectively removed.

The results were striking. At sufficiently large training scales, removing an individual image, all images associated with a particular person or even artwork from a particular artist could leave a generated result essentially unchanged. If eliminating the data does not meaningfully alter the output, determining that the specific material caused the output becomes extremely difficult.

Why Bigger Training Datasets Change the AI Amnesia Equation

The study examined diffusion models trained on datasets ranging from hundreds of images to more than 162,000 images. Researchers found a consistent relationship between increasing training-set size and declining attribution.

The effect also appeared when researchers attempted to attribute generated material to people and artists rather than individual images. According to the study, this means that simply identifying a visually similar work may not establish that the work actually caused an AI-generated result.

That distinction could become important in disputes involving copyright, licensing and compensation.

Artists and other creators have sued AI companies over the use of copyrighted material in training datasets. In some cases, plaintiffs have sought information about the datasets used to train image-generation systems. They are arguing that copyrighted works were collected and used to reproduce recognizable creative styles.

Copyright and AI Regulation

The MIT findings could complicate those legal debates.

If an AI-generated image cannot reliably be attributed to a particular training image or artist, courts may have difficulty determining whether a similarity resulted from direct influence, broader statistical learning or coincidence.

The researchers said attribution has potential applications involving machine unlearning, data poisoning, model interpretability, fairness and privacy. It also has significant implications for commercial and legal questions surrounding generative AI.

Cornell Law School professor James Grimmelmann said the findings suggest that technological attribution may not provide a reliable way to determine whether similarities between an AI output and copyrighted material resulted from copying.

That could force courts and researchers to rely on other methods when evaluating claims involving AI-generated content.

Could AI Amnesia Become a Legal Shield?

The research raises a particularly unusual possibility: Larger models could become harder to hold accountable to individual pieces of training data precisely because their knowledge is distributed across enormous datasets.

The researchers, however, caution that unattributability does not mean AI systems never memorize or reproduce training examples. Their study focuses on whether an output can be causally attributed to a particular unit of training data. The paper notes that models trained on extremely large datasets can still occasionally produce near-identical copies of individual training examples.

The researchers theorize that important features may become redundantly distributed throughout a large training dataset. As a result, removing one source may have little measurable effect because other examples contain overlapping information.

That creates a paradox: The more information an AI model absorbs, the harder it may become to determine which individual sources mattered.  AI Amnesia continues to grow along with the use of AI.

Additional Information

• Nature Communications: “Outputs of generative diffusion models are often unattributable”
• MIT CSAIL: “When AI art has no author”
• MIT News: Study on AI-generated images and training data attribution
• USA Herald: AI Hallucinations — Why Artificial Intelligence Can Sound Convincing While Getting the Facts Wrong
• USA Herald: Artificial Intelligence coverage and 2026 developments