4 ms·

So I'm way behind on developments in AI ethics. Are there actually any promising avenues for mitigating harmful impacts of AI? I'm possibly damaged by reading T

by Joeboy 9y ago

So I'm way behind on developments in AI ethics. Are there actually any promising avenues for mitigating harmful impacts of AI? I'm possibly damaged by reading The Guardian, but it seems like the only options are a) making AI fairer and b) using humans instead. And option a doesn't really work because people can't agree on what's fair.

sdenton4 9y ago
FWIW, humans aren't terribly fair either. You can create a shite review system, implemented by thousands of inconsistent humans, and end up with something crappy where it is far harder to point at any single decision and say 'this is crappy' because the system itself is masked by thousands of individual decisions. (Think 'customer service call centers' here.) Or you can automate, and end up with a _transparently_ bad decision system backed by a tweak-able model, which - at least in this moment in history - is far easier to write outraged front-page guardian articles about. The sorts of problems that we're deploying machine learning on often don't have single right answers. We've quickly gone from wonder that these things work at all to public outrage that they aren't providing perfect answers, but instead answers which often mirror our own societal shortcomings. The design choices made by the capitalists promote revenue over well-being, and the eigenvector of desire often points in the direction of the horrific. The systems, as they stand today, are like any other tool, amplifying human ethical choices, for better or worse.
- contextfree 9y ago
  "... which - at least in this moment in history - is far easier to write outraged front-page guardian articles about." which could be considered a positive feature! Sort of reminds me of how you'll occasionally see articles highlighting the excessive permissions demanded by mobile apps, but they rarely mention that under the application models still used by PC operating systems (most notoriously on Windows pre-UWP, but also unsandboxed Mac OS applications, and on mainstream Linux distributions), most apps run with full user rights and therefore have more excessive permissions than mobile apps are even able to ask for.

by Joeboy 9y ago

sdenton4 9y ago

FWIW, humans aren't terribly fair either. You can create a shite review system, implemented by thousands of inconsistent humans, and end up with something crappy where it is far harder to point at any single decision and say 'this is crappy' because the system itself is masked by thousands of individual decisions. (Think 'customer service call centers' here.) Or you can automate, and end up with a _transparently_ bad decision system backed by a tweak-able model, which - at least in this moment in history - is far easier to write outraged front-page guardian articles about. The sorts of problems that we're deploying machine learning on often don't have single right answers. We've quickly gone from wonder that these things work at all to public outrage that they aren't providing perfect answers, but instead answers which often mirror our own societal shortcomings. The design choices made by the capitalists promote revenue over well-being, and the eigenvector of desire often points in the direction of the horrific. The systems, as they stand today, are like any other tool, amplifying human ethical choices, for better or worse.

contextfree 9y ago
"... which - at least in this moment in history - is far easier to write outraged front-page guardian articles about." which could be considered a positive feature! Sort of reminds me of how you'll occasionally see articles highlighting the excessive permissions demanded by mobile apps, but they rarely mention that under the application models still used by PC operating systems (most notoriously on Windows pre-UWP, but also unsandboxed Mac OS applications, and on mainstream Linux distributions), most apps run with full user rights and therefore have more excessive permissions than mobile apps are even able to ask for.