Eric Loomis pled guilty to driving a stolen vehicle & fleeing police
Before sentencing, the court looked at an ML assessment (COMPAS) of his likelihood to reoffend. Loomis was rated a high risk
COMPAS was not trained for sentencing, but for resource allocation after release
COMPAS is biased in favor of whites. Loomis is white
The court sentenced him to 8 years in prison, in part because of his high risk rating
Fair?
Houston SD purchased an ML tool (EVAAS) to evaluate teachers' effectiveness. The system is proprietary
Houston fired 25% of teachers with ineffective scores, based solely on their scores
There is no process for teachers to inspect, challenge, or correct their data or scores
Known errors cannot be corrected because of how the model normalizes: fixing one teacher's data has the potential to change all other teachers' scores
Fair?
MSU uses cookies, ML, and third party datasets to score possible students who visit their site
The scores favor those:
likely to accept an admissions offer to MSU
applying to majors the school is already strong in
wealthy out-of-state students
The scores are used to determine who to focus recruitment efforts on, not who to accept
Fair?
What's Fair?
Mariah A. Knowles
Tue, Dec 6, 2022
image: Diana Ejaita for Vox
When we say something is fair...
Fairness is a conclusion
Fair amounts to Good, Justified, Okay, Fine, Morally Acceptable, ...
Fair amounts to
Fair is a thick evaluation, colloquially
Fair amounts to
+ some set of morally relevant criteria
...
Problem 1: Fair covers a range of moreorless related concepts, some competing
Problem 2: It's easy to talk past each other
Technology creates risks
We have responsibilities w.r.t. those risks
One type of risk is:
People will be subject to an environment
that runs afoul
of fair terms of social cooperation
Some Fair Terms
Malice
Respect for Individuals
Acceptable Social Arrangements
Deception
Acceptable Judgment
We ought not hold malice towards other groups
We ought to respect others as individuals
Power and Freedom ought to be distributed justly
&
Harms and Goods ought to be distributed "equally"
We ought not deceive others
We ought not judge others on things outside their control
&
We ought not judge others on things that bear no relevance to the outcome
...
ty
Decision Tree
Model Cards
Design Practices
Loomis & EVAAS
MSU
Facebook
Car Insurance
Pasco
Algorithm is Innocent
Breast Cancer