Wikimedia Developer Summit/2017/AI ethics

Wikimedia Developer Summit 2017 Session Notes Template

Session Overview

Title

Algorithmic dangers and transparency -- Best practices

Day & Time

Tuesday, January 10th at 1:10PM PST

Room

Hawthorne

Phabricator Task Link

https://phabricator.wikimedia.org/T147929

Facilitator(s)

Aaron Halfaker

Note-Taker(s)

Michael H, Ejegg

Remote Moderator

Advocate

Pau

Stream link

https://www.youtube.com/watch?v=myB278_QthA

IRC back channel

#wikimedia-ai ( http://webchat.freenode.net/?channels=wikimedia-ai) )

 

Session Summary

Detailed Summary

https://etherpad.wikimedia.org/p/devsummit17-AI_ethics

Purpose

Advance our principles with regards to what's acceptable or problematic with regards to advanced algorithms.

Agenda

• What's an AI?
  • What do we use them for?
  • Why are we worried about AIs?
• Some thoughts:
  • Stack protocol: One-hand (new), two-hands (continue thread)
    • Gatekeeper: Please use your judgement
• What happens next?
  • Method: https://en.wikipedia.org/wiki/Grounded_theory
  • Results: https://meta.wikimedia.org/wiki/Research:Algorithmic_dangers_and_transparency
• Victory! _o/ \o/ \o_

Style

Problem solving (discussion)

Discussion Topics

1. People affected by a prediction model should have the means to talk about it, identify problematic-seeming aspects
2. training data given to models should be proper wiki artifacts so they can be edited and discussed
3. how do we implement policies based on the "protected class" concept?
4. any AI model should account for change in the dynamics of the social system -- models should reflect our values as they change.
5. With vandalism tools, there's a tradeoff between educating users about why they got flagged and training vandals to be more subtle