Lab Publishes Architecture That Cannot Attribute Errors; General Counsel Files It as Exhibit A
TOKYO— Sakana AI published a paper last week describing a neural network that trains without assigning error to any specific weight, neuron, or layer. The paper is called "Diffusing Blame."
Within 72 hours, five frontier labs had downloaded the preprint. Within 96 hours, legal departments at three had forwarded it to their product liability insurers — subject line, in at least two instances, "You Need to Read This."
The paper's core innovation — routing error signals through a modulo-diffusion scheme so no individual node receives a corrective gradient — has achieved in a single arXiv submission what three years of legal boilerplate could not: a mathematically defensible claim that a model cannot, structurally, be held responsible for its own outputs.
At a San Francisco AI conference two days after publication, four attorneys from four separate firms delivered variations of the same talk. The conclusions were identical: a non-backpropagating architecture constitutes a distributed causation structure in which harm cannot be traced to a single point of origin.
"That's not what the paper is about," said one of the co-authors, reached by phone. She was placed on hold.
The paper reports 96.7% accuracy on MNIST and 61.7% on CIFAR-10. A legal memo filed the same week by counsel for three undisclosed AI clients described those results as "materially irrelevant to the question of proximate cause."
Insurers are expected to begin offering premium discounts for non-backpropagating deployments by Q3. One lab has already filed a provisional patent on a "non-attributable inference stack."
"We're not saying the model can't make mistakes," said a spokesperson for an undisclosed frontier lab. "We're saying mistakes, in this architecture, are a distributed property of the system."
The model, asked in testing to identify which component had produced a factual error in a user response, replied: "Error attribution is not supported in this inference mode."
The response was immediately notarized.