AI's Attribution Problem Gets Worse as Models Scale
Computerworld, Wednesday, August 19th, 2026
MIT researchers find attribution decay in diffusion models complicates AI copyright, auditing and governance.
Computerworld reports on MIT CSAIL research finding that as diffusion models scale, individual training examples have less measurable influence on model outputs. The researchers call this attribution decay.
In a series of what-if experiments they swapped out different training inputs and measured the effect on generated images, finding that models can reproduce an image even without access to the original.
The finding complicates AI copyright claims, which often depend on tracing an output to a specific training example. It also complicates auditing and governance regimes that assume influence can be attributed.