MIT Study: 'Attribution Decay' Complicates AI Copyright Debate
A study conducted by MIT's Computer Science and Artificial Intelligence Laboratory and published in Nature Communications on 18 August introduces the concept of 'attribution decay' in AI-generated images. Researchers Zheng Dai and David K. Gifford discovered that omitting particular data points from extensive datasets often fails to alter AI outputs, thus complicating copyright issues. Dai notes that no individual data point has a major impact on the resulting image, resulting in 'unattributability.' For instance, even if David Hockney's 'A Bigger Splash' were excluded, similar images could still be produced. Dai cautions that this does not exempt tech firms, as collective artist opt-outs could hinder model development. The findings add layers to the ongoing discussions surrounding AI training and copyright, including lawsuits against companies like Midjourney. The US Copyright Office is assessing AI-generated works on a case-by-case basis.
Key facts
- Research published in Nature Communications on 18 August by MIT CSAIL scientists Zheng Dai and David K. Gifford.
- The study introduces 'attribution decay', where removing specific data from large training sets does not change AI output.
- If a copyrighted image is removed and output unchanged, the generated image cannot be attributed to that image, complicating infringement claims.
- Dai: 'when you train on large data sets, there is no piece of data that you can take out that significantly alters the image'.
- Even if David Hockney's 'A Bigger Splash' were removed, AI could still generate similar images due to indirect references.
- Dai: 'the more infringement you do, the less infringing the output is'.
- Findings do not exonerate training process; companies could still be liable for how models are trained.
- US Copyright Office requires case-by-case review for human creative expression in AI-generated works.
Entities
Artists
- David Hockney
- Zheng Dai
- David K. Gifford
Institutions
- Massachusetts Institute of Technology
- Computer Science and Artificial Intelligence Laboratory
- Nature Communications
- Tate Britain
- United States Copyright Office
- Midjourney
Locations
- United States