search

webmachinelearning / webmachinelearning-ethics

😇 Ethical Principles for Web Machine Learning
https://www.w3.org/TR/webmachinelearning-ethics/
Other
16 stars 3 forks source link

Add a "human rights considerations" or similar section #2

Closed jyasskin closed 2 years ago

jyasskin commented 2 years ago

Machine learning systems have often learned biased behavior, from the Google "gorillas" incident to face recognition that misses black people and women to re-offending predictors that deny bail to black people.

While this spec can't prevent people from building bad ML-based systems, it can and should give some advice about how to do them better. https://datatracker.ietf.org/doc/draft-irtf-hrpc-guidelines/ proposes a standard "Human Rights Considerations" section, and this information could naturally fit there, but you could instead add a more specific section. I don't think this fits well in the existing Security or Privacy Considerations sections.

anssiko commented 2 years ago

Thanks @jyasskin for raising this important issue!

We discussed ethical issues at TPAC and are looking for cross-disciplinary contributors to help document these issues and applicable mitigations per our charter commitments.

We have started a separate repo for this effort (https://github.com/webmachinelearning/ethical-webmachinelearning) so I'll transfer this issue there.

I think what you outlined are good concrete "bad examples" to be incorporated. Perhaps the ethical considerations document should be linked from the affected specs when we've fleshed it out a bit better.

Shameless plug: let us know if you're aware of other Googlers, or others, who have expertise in this space and would like to contribute. I'm not an ethics expert, but I care about this topic personally and wanted to help get this work started.

(For context, "this spec" refers to the Web Neural Network API.)

jyasskin commented 2 years ago

https://ainowinstitute.org/ is probably a good group to invite to participate.

Apparently ISO SC42 is working on ethical standards too.

I've also invited an appropriate Googler.

anssiko commented 2 years ago

@jyasskin I happened to find a connection to the AI Now Institute at New York University through my network, so we'll explore that path -- thanks for the suggestion! We're expecting a new participant to join this effort shortly (whose GH handle I don't yet have) who is well positioned to establish the AI Now Institute connection.

Anyone we could tag from ISO SC42 you know of who could help introduce the large body of documentation (behind a paywall) ISO has produced in this space?

bbcjamesfletcher commented 2 years ago

@jyasskin @anssiko just working through the GitHub issues before I hand the doc over - as this predates my involvement, I'm presuming the general thrust of this issue, that the spec "can and should give some advice about how to do them better" on issues such as bias etc. has been covered with the approach we have developed in the document. I'll close this issue, but if you feel the current approach misses any specific elements, please flag them.