If you are reading that Economist interview with an AI [@openAI's GPT-2], you absolutely need to read the tweet below for context.
— Gary Marcus (@GaryMarcus) November 26, 2019
Economist's claim that interview was unedited is grossly misleading; coherence was in the eyes of human, not the output of the machine. pic.twitter.com/X0gDyZVt7g
For anyone not familiar with the fairness literature, as a start, I recommend the classic “Fairness through Awareness” paper on how to make use of the protected class attributes to ensure fairness. https://t.co/2xZXFAZq5T https://t.co/aiYpWPg06Y
— Delip Rao (@deliprao) November 20, 2019
This is such an important and underappreciated point. Simple, rule-based approaches can be good for consistency, fairness, & transparency. https://t.co/6ReoUNs3Pn
— Rachel Thomas (@math_rachel) November 19, 2019
Sampling bias occurs when not all elements have the same probability of appearing in the datasets. Police records are not a representative sample of crime due to variation in:
— Rachel Thomas (@math_rachel) November 17, 2019
- reporting rates
- police attention
- arrest rates (eg Black people more likely arrested for marijuana) pic.twitter.com/VWfEJW0pUJ
So many AI hiring systems are appallingly bad.
— Janelle Shane (@JanelleCShane) November 14, 2019
Better to draw the resumes at random out of a hat; at least then you know it won't be copying bias. https://t.co/0FKY69l7Jx
I was interviewed by @Slate’s Aaron Mak about the Apple Card discrimination problem: https://t.co/OFs0RDv4pB
— Cathy O'Neil (@mathbabedotorg) November 12, 2019
A few pieces of advice for Apple:
— Cathy O'Neil (@mathbabedotorg) November 11, 2019
1. don't deny this algorithm is a problem, and don't pretend it isn't your problem.
2. figure out what people actually want to know; the answer isn't your valuable source code, so please don't claim you cannot address it by hiding behind IP. https://t.co/RjpO4VX86a
Thread: DHH receives credit limit 20x higher than his wife for @AppleCard, even though it turns out his wife has a higher credit score than him. When they try to find out why, are told by Apple "I swear we're not discriminating. It's just the algorithm." https://t.co/4RRKcJghKq
— Rachel Thomas (@math_rachel) November 9, 2019
There's been some distributing ML applications on my TL recently: ETS's automated essay grading w/ demonstrable bias against African American students & the paper attempting to automate criminal sentencing (ht @there_are_two_) for example.
— Rachael Tatman (@rctatman) November 6, 2019
Here're my thoughts (as a flow chart). pic.twitter.com/ffzzPNbi5g
More generally, algorithmic accountability cannot be automated. https://t.co/WD1QVqe1oe
— Cathy O'Neil (@mathbabedotorg) November 5, 2019
The Compas Recidivism Algorithm:
— Rachel Thomas (@math_rachel) November 3, 2019
- it's no more accurate than random people (Amazon Mechanical Turk)
- it's a black box with 137 inputs but no more accurate than linear classifier on 2 vars
- Wisconsin Supreme Court upheld its use (it is still used in other states as well) pic.twitter.com/cMdGOkQPP5
My talk on "Getting Specific about Algorithmic Bias"https://t.co/K5MOawggb4 pic.twitter.com/efNQtoVClJ
— Rachel Thomas (@math_rachel) November 3, 2019
