Fair, Ethical and Just: Can an AI Algorithm Check all the Boxes?

Global Affairs Academy
6 followers
1 Video View·Aug 29, 2026
Doaa Abu-Elyounes of Harvard Law School Berkman Klein Center for Internet and Society, speaking on Panel IX: 'AI'.
AI Generated·May contain errors
Good afternoon, everyone.
I hope you had a good lunch,
because otherwise it's not fair
to talk about fairness on empty stomach.
Thank you for joining me today.
So, all actors in the field are working intensively
on policy design for the use and development of AI.
All initiatives put in the center
the call for AI to be fair, ethical, and just.
But what does exactly it means?
It's an open questions
that I will try to unpack in this presentation
using a multidisciplinary approach.
So, the computer science literature
refer to more than 20 different notions of fairness.
Each paper spend a lot of time
on arguing why the suggested notion of fairness
is the fairest, the most just, and the most ethical.
In my paper, in my presentation,
I argue that the complexity of the various policy domains
that algorithms are implemented in
require different solutions.
Therefore, there is no hierarchy
between the different notions of fairness,
and one is not superior to the other.
So, the legal and social frameworks
surrounding each policy domain
require a notion that is tailored
to its unique characteristics.
What I'll do is that I will address
the three main groups of fairness notions
in the computer science literature.
I will address the sub-notions of each group,
determining to which policy domain
each one of the fairness notions is most suitable for.
Let's start with individual fairness, the first group.
The focus of this group is the individual,
regardless of his or her group affiliation.
And the idea is that everyone is equal in front of the law.
The first sub-notion in this group is the unaware approach.
It's called for equal protection,
but in a traditional way of viewing equal protection,
as colorblind.
So, the algorithm should be blinded
and unaware of any differences between people,
and the only prohibited attributes by law
such as sex or gender cannot be included in the algorithm.
Legally, we already know that colorblindness doesn't work,
because success is not just a matter of effort and talent,
but also a matter of access to resources
that you have or don't have
depends on your group affiliation.
Computationally, it's very easy to design an algorithm
that doesn't take into account race or gender,
but there are many factors that could serve as proxies
for the protected attributes.
I think that individual fairness in the unaware approach
can work only in cases where the group of people
that we are comparing between is quite homogeneous.
Still in the realm of individual fairness,
fairness through awareness
call for treating similarly situated people similarly,
not everyone.
Legally, the idea behind this approach
is that regulation by its nature
seeks to differentiate between individuals,
and the questions to ask are
when it's legal to classify people,
and what is the basis for the classification?
So, I will demonstrate with an example.
In the criminal justice,
imagine that there is a study that show that
if you are a black defendant
and if you have five priors or more,
you will be considered risky.
But if you are a white defendant
and you have two priors or more,
it's enough to consider you as risky.
The similarly situated individuals
will be black defendants with five priors or more
and white defendants with two priors or more.
This notion is quite promising,
but it's very hard to determine the metric,
the computational metric that will identify
the similarities between individuals.
Group fairness approaches is the second biggest group.
It include many sub notions
that the common denominator between all of them
is that not all groups
have the same starting point in life.
The best is to integrate differences in the equation
and not to ignore them.
Legally, group fairness equal affirmative action,
basically favoring a group of individuals
that have been historically disadvantaged
in order to give them an equal starting point.
Affirmative action is a civil mechanism
that has been approved in court and by the legislator
in a very narrow set of cases.
And these are the cases that can be used
in the computational sense as well.
The first group notion approach is decoupling.
It's called for creating one algorithm per group.
So COMPAS, a risk assessment tool using criminal justice,
have two versions, General COMPAS and COMPAS Women.
And the creators of COMPAS justify having that
because women compose a very small part
of the criminal justice.
So their claim is that if we don't have
a specific tool for women,
their unique characteristics and needs will be ignored.
So COMPAS takes into account economic marginalization,
trauma and other unique characteristics
that are special for women.
But the question is that,
do we want to allow that as a society?
Do we feel comfortable with having COMPAS Men
and COMPAS Women?
And how about having COMPAS Black and COMPAS White?
Okay, the second approach in group fairness
is statistical parity.
This one called for making sure that the outcome
is distributed equally in accordance
with the total population.
So if I have 100 loans to grant,
I should give 50 to men and 50 to women
because women compose a 50% of the society.
Computer scientists dismiss notions that are based on,
dismiss this notion quickly
and consider the following example.
So if I'm the CEO of an IT company
and I'm trying to hire a couple of individuals,
I might use algorithm that is based on statistical parity.
But one might tell me that it's not a good algorithm
because I will have to admit less qualified women
in order to satisfy the 50-50 percentage.
But I want to argue that using statistical parity
will actually solve other problems
because we all know that the reason why we have less women
in IT-related jobs, prestigious IT-related jobs,
is not because they're less qualified,
but because maybe the ads that advertise this kind of jobs,
women don't reach them.
They're not published for women or when they are,
they are kind of designed in a very hostile way for women.
So I think it's a good example
and the same could be applied for school admissions.
Equalized odds is the next group fairness notion.
It recognizes the fact that there is no algorithm
that predict 100% correct in all of the cases
and it's called for equalizing the errors
that the algorithm make.
So the errors that the algorithm make
are called false positives and false negatives.
And equalizing them is complicated
because society value them differently.
If I'm a utilitarianist, again, in criminal justice,
I care about the total happiness
of the greatest number of people
and I care about public safety.
I wouldn't mind putting a bit more people in jail
in order to make sure that public safety is there.
But if I'm focusing on the individual,
I'm an egalitarianist, I probably remember
that our criminal justice is centered
on the beyond the reasonable doubt standard
and I have to be the most fair toward the individual.
What is the error rate that our society
is willing to tolerate is a very hard question
both legally and computationally.
The third group, causal reasoning based approaches,
was developed because of two unique characteristics
of machine learning that have been discussed here before.
The first one is that machine learning algorithms
are based on correlations
and correlation does not imply causation.
And also the fact that machine learning algorithms
lack explainability.
And the combination between those two can harm due process.
So causal reasoning based approaches
will only take into account the factors
that it has been proven by experts
that they cause the outcome.
It shouldn't be surprising that we can't satisfy
all notions of fairness.
So campus, the tool for risk assessment
that I mentioned before,
it gives the dependent each one a score
from low to high that determines the risk of recidivism.
ProPublica, a news outlet in the US,
conducted an investigation into campus
and concluded that it's biased toward black
because among dependents who were released from jail
after two years, among black dependents,
people who didn't commit any other crime,
42% were labeled mistakenly as high risk
and the algorithm did the same mistake
among white dependents only in 22% of the cases.
Of course, the developers of campus
descended from this conclusion
and they said that campus is fair.
And the core argument between ProPublica and Northpoint,
the developers of campus,
is a misconception of fairness.
ProPublica wanted equal opportunity.
They wanted black and white dependents,
low risk dependents, to have the opportunity
to be classified as so.
And Northpoint wanted individual fairness.
They are claiming the law doesn't allow us
to treat black dependents differently
and given the difference in the base rate
between black and white dependents,
the gap that ProPublica is referring to
will always be there.
But in fact, none of their approaches
address the most important issue
and it is the fact, the differences in the base rate
and the fact that black dependents are overrepresented
in the criminal justice because of historical reasons.
So what developers can do in order
to better address algorithmic fairness?
The first and most important point
is to clarify their approach to fairness
and to basically show whether other notions of fairness
were examined before the chosen one
was before the one that has been developed was chosen.
And also having much more information about the society
that the algorithm would be implemented in is crucial.
From the point of view of the policymakers,
it's very important to clarify the laws and regulations
but this is a tricky one because law by its nature
is designed to be broad in order to accommodate
as many cases as possible.
But for the algorithm, it's the opposite.
The more specific you are,
better the prediction that you will get.
So balancing between those two is very important.
It's also very important to audit
the results of the algorithm.
So to conclude, all models are wrong but some are useful
is a quote used to teach students in statistics classes
that the models are wrong because they represent
a simplified version of reality.
But what I hope I started to convince you
in my presentation is that most models are right
but it depends how we use them.
The variety of fairness notion is big
and I expect it to grow, this diagram.
The notion that is working for credit scoring
might not be the same notion
that will work for criminal justice.
And it's very important to be aware
of the very specific details of each case.
Thank you.
