You are looking at 1 - 1 of 1 items

  • Author or Editor: Paul Ohm x
Clear All Modify Search
You do not have access to this content

Paul Ohm

This chapter offers two ‘throttling metrics’, quantitative measurements of some aspect of a machine learning system. These metrics can be used to regulate our ability to control the way machine learning systems impinge on important human values. The first metric measures the ratio between the performance of an artificial intelligence system to the human decision-maker who today serves the same purpose. Too often we implicitly (or sometimes explicitly) set the target of this metric at the value of one, at simple equivalence. Any AI that performs ‘as well as’ a human is ready to be deployed. Instead, we would often be better off setting this metric at a higher value, requiring an artificial intelligence to be two or three or even fifty times better than a human before it should be deployed. The second metric measures the incompleteness of the data used to train a machine learning system. Machine learning experts have been given access to more information about human behavior than ever before. In this environment, anything that forces a researcher to analyze less than the maximum amount of available data, for example a privacy rule allowing an individual to opt-out of the research, has begun to be characterized as deviant or anti-science. We should instead sometimes consider the forced contraction of training datasets as a second metric, something we should protect and preserve rather than try to argue out of existence.