A network's error peaks where it can just fit its training data, then falls again
Feed a network 4.5 times the training data and it can get worse, a limit one of the two 2019 papers behind double descent measured directly.
A different attack turns 1% into 88% on the same safety test
A single Cisco score branded DeepSeek R1 unsafe, and it came from 50 requests pushed by one attack.
Rename the people in a word problem and the model does worse
A model loses a little accuracy when a problem's names change and a lot when it gains one true but useless sentence, and whether that proves it cannot reason is the part still in dispute.
Tesla's lawyers said Elon Musk's recorded words might be a deepfake
The denial that a real recording is a deepfake is already being tried in court and campaigns, and the only experiment to test it found it works on text scandals but largely fails on video, the very evidence it was meant to discredit.
Houston pushed out teachers over a score they could not check
The district rated its teachers with SAS's private EVAAS formula and ended some contracts on the results, until a 2017 federal ruling found that a score no one could recompute could not lawfully cost a teacher their job.