AI Fairness and Big Data

Egalitarianism

The view that equality is a fundamental moral value

stop using AI

AI can make biased decisions

Humans can make biased decisions too

Humans can be held accountable

AI can not be held accountable

---

AI can make mistakes, which can lead to serious consequences

We can improve AI by adjusting its weights and using feedback

we cannot solve all of them completely (even in the medical field)

what makes discrimination wrong

The agent’s intention, beliefs and values

statistical discrimination

Discrimination can be wrong because it judges an individual based on generalization about a group they belong to

Leads to stereotypes

challenges that statistical discrimination is not always wrong

Objection: the individual may not fit the group pattern

a statistical average does not tell us what the particular person would do

It may reinforce stereotypes about the group

a statistical pattern can become a fixed assumption about everyone in that group

data

Information about objects, events, or qualities that can be stored and analyzed

Act as proxies for real-world objects and events or qualities

data as proxies

ex: heart rate records represent your past heartbeats

proxy attributes

indirectly

using ZIP codes to guess socioeconomic or racial background

analog representation

changes continuously to represent information

digital representation

distinct values to represent information

big data

Extremely large amounts of digital data that can be collected, stored, and analyzed using computers

Identify correlations and make predictions, but understanding causation is often necessary when we to want change outcomes

datafication

turning real world activities events, qualities into data can be stored and analyzed

limits of big data

correlation is not causation: identifying patterns doesn't necessarily mean understanding why things happen

the limits of historical data: past data cannot always predict future possibilities

descriptive facts are not normative judgements: knowing what is likely to happen doesn’t tell us what we ought to do