Article 10. Fairness reviews: identifying essential attributes

Spoken (by a human) version of this article.

When we're checking for fairness in our algorithmic systems (incl. processes, models, rules), we often ask:

What are the personal characteristics or attributes that, if used, could lead to discrimination?

This article provides a basic framework for identifying and categorising these attributes.

About this podcast

A podcast for Financial Services leaders, where we discuss fairness and accuracy in the use of data, algorithms, and AI.

Hosted by Yusuf Moolla.
Produced by Risk Insights (riskinsights.com.au).

Episode Transcript

(00:00):
This article was published inOctober 2024, and it's titled
Fairness Reviews, IdentifyingEssential Attributes.
in a few discussions and workingout exactly what to do for
clients, , this is somethingthat's come up we want to do
a fairness review, what exactlyare we going to be looking for?
Because there's variousdefinitions, etc.

(00:21):
And so, we've come up withan approach and this article
explains what that approach is.
So here we go.
In a previous article,we discussed fairness
in algorithmic systems,equity and equality.
When we're checking for fairnessin our algorithmic systems,
which include processes,models and rules, we often

(00:43):
ask the question, what arethe personal characteristics
or attributes that, if used,could lead to discrimination?
This article provides a basicframework for identifying and
categorizing those attributes.
Anti discrimination laws existin most jurisdictions, so
that's a good place to start.
If none apply to yourcountry, so for example,

(01:05):
South Korea, Japan, at somepoint anyway, they had no
anti discrimination laws.
You could use existinghuman rights laws or perhaps
one of the internationalcovenants or conventions.
Let's talk aboutthe legal landscape.
There's no shortageof definitions when it
comes to discrimination.
For example, in Australia,there are at least five

(01:26):
relevant federal laws.
Each state and territory has itsown set of rules, legislation.
And the definitions vary,but there's some overlap.
One example of a definitionis detailed in the 2014 guide.
Produced by the Australian HumanRights Commission, a quick guide

(01:47):
to Australian discriminationlaws and its states.
The Australian HumanRights Commission Act 1986
specifies discrimination onthe basis of race, colour,
sex, religion, and gender.
Political opinion, nationalextraction, social origin,
age, medical record, criminalrecord, marital or relationship

(02:09):
status, impairment, mental,intellectual or psychiatric
disability, physical disability,nationality, sexual orientation,
and trade union activity.
That's a lot to take in, andthat's just one definition,
simplifying the approach.

(02:31):
To make this easierto work with, we can
group these attributesinto 5 main categories.
So as opposed to the previouslist, which was a little
bit difficult to followbecause it was just a list
of items all together.
If we categorize them, itstarts to make it a little
bit easier to manage.
1.
Age, 2.

(02:53):
Race, 3.
Sex or Gender, 4.
Disability, 5.
Activity slash Beliefs.
Each of these containsseveral attributes.
Detailing the attributes canhelp provide context and support
our efforts to reduce bias.
So in the article, there'sa table, three by five,

(03:14):
and it lists each categorywith its attributes.
And I'll just call thatout really quickly.
First category is age, so that'sage including age specific
characteristics, and that willvary depending on the nature
of the work that you do.
Race, so race, color,descent, nationality, Origin,

(03:35):
so that's ethnic origin ornational origin or ethno
religious origin, immigrantstatus and physical features.
Sex and gender includes gender,sex, gender identity, intersex
status, sexual activity,sexual orientation, marital
or relationship status,parental status, pregnancy

(03:57):
or potential pregnancy.
Breastfeeding or bottlefeeding, family or
carer responsibilities.
Disability would includephysical, intellectual,
psychiatric, sensory,neurological or learning
disability, then physicaldisfigurement, disorder,
illness or disease thataffects thought processes,

(04:17):
Perception of reality,emotions or judgment, or
results in disturbed behavior.
And in presence of organismscausing or capable of
causing disease or illness.
Activity and beliefscould include religious or
political beliefs, religiouspolitical activity or
affiliation, profession,trade, occupation, industrial

(04:41):
or trade union activity.
There are some additionalconsiderations, so there are
a few more to think about.
These are less frequentlyobserved, but they
need to be considered.
They don't, I mean, theycould fit into one of the
above categories, not veryneatly, but, you might be
able to squeeze them in.
And those are three, soirrelevant medical records.

(05:03):
Irrelevant criminal records,and discrimination based on
association with someone whohas any of the above attributes.
Putting it into practice.
Consider whether and howeach of the attributes might
be influencing decisions.
Some key questions to ask.
Are we collecting data onany of these attributes?

(05:24):
Could our systems be indirectlyusing these attributes?
Are we using externaldata or models that
use these attributes?
Do our policies or procedurestreat people differently
based on these attributes?
And finally, are our staff,including, importantly,
data scientists, aware ofthese potential biases?

(05:46):
We could go a bit deeper.
So going a bit deeper,we may ask, I've got five
questions again here.
Are we inadvertently proxyingprotected attributes through
seemingly neutral data?
Are we prepared to explainour fairness approach to
customers, regulators,or other stakeholders?
Can we explain how ouralgorithmic systems
work end to end?

(06:08):
Do we regularlyaudit our systems?
And then lastly, howquickly can we respond
if we detect unfairnessin our deployed systems?
Regularly revisiting thesequestions helps ensure our
systems remain fair andequitable as they evolve.
There's a disclaimer in thearticle and I'd leave that to
you to read, but basically theinformation in this article,

(06:30):
this episode does not constitutelegal advice and it may not be
relevant to your circumstances.
That's the end of the article.
Thanks for listening.