{"id":1243,"date":"2022-02-28T11:05:59","date_gmt":"2022-02-28T10:05:59","guid":{"rendered":"https:\/\/trustyour.ai\/whitepaper\/certifying-fairness-of-ai-applications-an-impossible-task\/"},"modified":"2022-03-02T10:49:55","modified_gmt":"2022-03-02T09:49:55","slug":"certifying-fairness-of-ai-applications-an-impossible-task","status":"publish","type":"page","link":"https:\/\/trustyour.ai\/en\/whitepaper\/certifying-fairness-of-ai-applications-an-impossible-task\/","title":{"rendered":"Certifying Fairness of AI-Applications An Impossible Task?"},"content":{"rendered":"\t\t
\n\t\t\t\t\t\t
\n\t\t\t\t\t\t
\n\t\t\t\t\t
\n\t\t\t
\n\t\t\t\t\t\t\t<\/div>\n\t\t<\/div>\n\t\t\t\t\t<\/div>\n\t\t<\/section>\n\t\t\t\t
\n\t\t\t\t\t\t
\n\t\t\t\t\t
\n\t\t\t
\n\t\t\t\t\t\t
\n\t\t\t\t
\n\t\t\t\t\t\t\t\t\t

INTRODUCTION<\/h2>

Since more and more decisions that used to be made by humans are nowadays made either with the help of Artificial Intelligence (AI) or by AI alone, it is essential that the AI algorithms are \u201cfair.\u201d In this whitepaper, we discuss issues surrounding the fairness of AI applications, with a special focus on how it can be assessed independently and subsequently certified. We explain why an AI application cannot be classified \u2013 and subsequently certified \u2013 as \u201cfair\u201d or \u201cunfair\u201d in a general sense and propose an approach that makes it possible to classify it as \u201cfair\u201d or \u201cunfair\u201d under the (application)-specific\u00a0definition of fairness.<\/span><\/p>\t\t\t\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t\t<\/div>\n\t\t<\/div>\n\t\t\t\t

\n\t\t\t
\n\t\t\t\t\t\t
\n\t\t\t\t
\n\t\t\t\t\t\t\t\t\t

\"\"<\/a><\/p>

Download Whitepaper
<\/b><\/a>(pdf, 5 MB)<\/p>\t\t\t\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t\t<\/div>\n\t\t<\/div>\n\t\t\t\t\t<\/div>\n\t\t<\/section>\n\t\t\t\t

\n\t\t\t\t\t\t
\n\t\t\t\t\t
\n\t\t\t
\n\t\t\t\t\t\t
\n\t\t\t\t
\n\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\"\"\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t
\n\t\t\t\t
\n\t\t\t\t\t\t\t\t\t

Mission Impossible?<\/h2>

As Artificial Intelligence (AI) is becoming increasingly popular, concerns about its application are rising. One of the most controversial issues is that AI-based applications could result in an unfair outcome, e.g., by favoring men over women or people of a specific ethnicity over people of other ethnicities. Issues like that can be addressed by establishing ways of determining whether an AI-based application is \u201cfair\u201d1<\/a><\/sup>,2<\/a><\/sup>. However, this is a very difficult task since \u201cfairness\u201d is difficult to define.\u00a0<\/p>

Most people have a good sense of fairness and – even more so – unfairness when evaluating existing outcomes. For example, if an algorithm used for job applications consistently rejects applications by women, most people would call it unfair. However, the question \u201cWhat is fair?\u201d is much harder to answer, and people are typically less willing to provide quick and precise responses to it. One potential response in our example could be \u201cThe algorithm should pick an equal number of men and women.\u201d Intuitively, this sounds fair. But what if there are significantly more male applicants than female ones? 10 out of 10 women and 10 out of 50 men would appear unfair to most people. It may be tempting to make it right by taking the same fraction of each, i.e., 2 out of 10 women and 10 out of 50 men. But what if there are big differences in the educational backgrounds among female and male applicants? What would a fair outcome be in that case? Another intuitive answer could be \u201cMen and women should be treated equally.\u201d But how can equality be verified? And wouldn\u2019t this be contrary to the intuitive fairness idea that a roughly equal number of men and women should be considered? Even this simple example shows how challenging it is to find objective ways of certifying something as \u201cfair.\u201d\u00a0<\/p>

Another fundamental issue is that such concepts as fairness are deeply rooted in cultural beliefs, i.e., what is considered fair or unfair varies significantly between different countries, regions, ethnic groups, age groups, etc. In addition, the notion of fairness can – and most probably will \u2013 evolve and change over time. All this makes certifying something as \u2018fair\u2019 independently and objectively almost impossible.<\/p>

Nevertheless, there clearly is a demand for ascertaining the fairness of applications since fairness is closely linked to non-discrimination, which is a fundamental right according to Article 21 of EU\u2019s3<\/a><\/sup> Charter of Fundamental Rights. Non-discrimination also plays an important role in the recently proposed EU AI regulation, and the EU\u2019s Fundamental Right Agency (FRA) has published several reports on algorithmic fairness and related issues4<\/sup><\/a>,5<\/a><\/sup>,6<\/a><\/sup>. For businesses, preventing legal problems and public setback is an important incentive for ensuring that their AI applications are fair and non-discriminatory.\u00a0<\/p>

In this paper, we provide an overview of ways to certify the fairness aspects of AI applications and demonstrate how the problem of objective fairness certification can be solved.<\/p>\t\t\t\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t

\n\t\t\t\t
\n\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\"\"\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t\t<\/div>\n\t\t<\/div>\n\t\t\t\t\t<\/div>\n\t\t<\/section>\n\t\t\t\t
\n\t\t\t\t\t\t
\n\t\t\t\t\t
\n\t\t\t
\n\t\t\t\t\t\t
\n\t\t\t\t
\n\t\t\t\t\t\t\t\t\t

Defining Fairness<\/h2>

In the introduction, we have mentioned some fundamental issues with defining fairness (e.g., different intuitive fairness definitions that contradict each other and the general idea of fairness being significantly affected by culture).<\/p>

Based on the realization that different concepts of fairness may be contradictory, two main concepts of fairness have been established: group fairness and individual fairness7<\/a><\/sup>.<\/p>

Group fairness means that groups as a whole are treated equally. In our example of job application algorithm described in the introduction, this means that the group of women as a whole is treated equally to the group of men. In this case, \u201cequal treatment\u201d could mean that an equal fraction of both is accepted or that, on average, the quality of predictions is the same for both groups (e.g., the fraction of excellent candidates erroneously rejected by the algorithm).<\/p>

Individual fairness means that individuals are considered treated equally if they are equal regardless of the attributes (e.g., gender, ethnicity). The meaning of \u201cequal individuals\u201d depends on the context and the application. In our example of job application algorithm, individual fairness means that individuals with equal attributes relevant for the job (e.g., education, experience, grades) are treated equally and independently of their gender, etc.<\/p>

Individual fairness, which is closely related to how non-discrimination is legally defined, sounds very intuitive at first. However, there are lots of issues associated with it. The available attributes (e.g., experience in our example) cannot always be assessed objectively due to existing inequalities. For example, degrees obtained from certain universities may be valued more than others even if objectively there seems to be no difference. In addition, even if they are assessed objectively, differences could originate from existing inequalities: a person may have less job experience because of discriminatory employment practices, just like a less educated person may have less education not for the lack of talent and motivation but rather because they belong to a certain population group whose access to education is limited. The last example leads to another very important dilemma: in our application example, where do we draw the line when assessing fairness? If two groups have very different levels of education due to systematic discrimination and yet a certain education level is essential for assessing the candidate\u2019s suitability for a certain job, is the algorithm that is based on that desired education level and will consequently select more people from the group with higher average higher education level (the \u201eprivileged\u201c group) unfair? While from the societal point of view this would mostly be considered unfair, it is very difficult to decide whether the party that uses the algorithm is responsible for ensuring the fairness that goes beyond the scope of its hiring algorithm (equal chances for that particular job based on the education level or a wider scope of equal educational opportunities) and whether it is responsible for \u201cremedying\u201d the educational inequalities artificially by favoring individuals from the underprivileged group (in the case of gender, the EU law would explicitly permit this (Art. 21, p. 2 of the Fundamental Rights Charter8<\/a><\/sup>)).<\/p>

The most important thing to realize is that individual fairness and group fairness are generally contradictory. Despite the ongoing philosophical discussion on whether those two concepts are in fact separate concepts or mainly a matter of worldview9<\/a><\/sup>, in practical settings they are clearly distinct.<\/p>\t\t\t\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t\t<\/div>\n\t\t<\/div>\n\t\t\t\t\t<\/div>\n\t\t<\/section>\n\t\t\t\t

\n\t\t\t\t\t\t
\n\t\t\t\t\t
\n\t\t\t
\n\t\t\t\t\t\t
\n\t\t\t\t
\n\t\t\t\t\t\t\t\t\t

Measuring Fairness<\/h2>

In response to the rising awareness of AI-related fairness issues, the research community has developed a wide range of quantitative fairness and bias metrics10<\/a><\/sup> many of which are already integrated open-source toolkits11<\/a><\/sup>,12<\/a><\/sup>,13<\/a><\/sup>,14<\/a><\/sup>, which assess group fairness, individual fairness or a balanced combination of the two. However, as evidenced by a large number of proposed and applied fairness metrics, there exists no single \u201cideal\u201d fairness measure due to the challenge of defining fairness described above.<\/p>

Equality-in-Outcome Approach<\/h5>

Statistical parity difference (spd) is a group-fairness measure defined as:<\/p>

\"\"<\/p>

which can also be written as probabilities:<\/p>

\"\"<\/p>

How exactly the distinction between privileged and underprivileged groups is made depends on the context of the application. spd compares the fractions of each group classified as 1 (a positive classification, meaning an invitation to a job interview in our hiring algorithm example). This comparison is independent of the group size. In this metric, perfect equality is reached at a value of zero, which means that the same fraction of people is chosen from both groups. For example, if 20% of the overall population are invited for a job interview, in order to be fair 20% of each group should be invited regardless of the group size. This is often described as a \u201cwe-are-all-equal\u201d approach since it assumes that there is no difference between the groups even if there is a difference in the data. One main drawback of spd is that it does not address true or false classification of individuals.<\/p>

Equality-in-Quality Approach<\/h5>

The average odds difference (aod) is another group fairness measure. In contrast to spd, it does not consider classifications as such (e.g., acceptance rates in our job application algorithm example) but rather the quality of the classifications and measures whether the quality rate defined via the false positive (the fraction that is incorrectly classified as 1) and the positive rate (the fraction that is correctly classified as 1) differs between the groups.<\/p>

\"\"<\/p>

Perfect equality is indicated by a zero. This is often described as a \u201cwhat-you-see-is-what-you-get\u201d worldview since it does not assume that an equal fraction from all groups must be accepted, only that the quality of the predictions should be the same for each group. Under this approach, the algorithm can choose only a very small fraction of one of the groups as long as it corresponds to the underlying data. In other words, this approach assumes that the data itself is not biased and accurately reflects reality, ensuring that no bias is introduced by the AI algorithm.<\/p>

Besides these two examples, there are many other fairness notions, such as conditional statistical parity, equal opportunity, fairness through unwareness, fairness through awareness and counterfactual fairness (https:\/\/arxiv.org\/pdf\/1908.09635v1.pdf).<\/p>

Below is a summary of issues related to fairness of AI applications:<\/p>

  1. There is no single concept of fairness.<\/li>
  2. Definitions of fairness often contradict each other.<\/li>
  3. Setting the boundary for fairness issues in an application is non-trivial and ambiguous.<\/li>
  4. All of the above points are tightly coupled with ethical considerations and worldviews that depend on many factors, such as region and culture, and can change over time.<\/li><\/ol>

    In the remaining part of this paper, we discuss how to overcome those obstacles when certifying AI applications in term of fairness.<\/p>\t\t\t\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t\t<\/div>\n\t\t<\/div>\n\t\t\t\t\t<\/div>\n\t\t<\/section>\n\t\t\t\t

    \n\t\t\t\t\t\t
    \n\t\t\t\t\t
    \n\t\t\t
    \n\t\t\t\t\t\t
    \n\t\t\t\t
    \n\t\t\t\t\t\t\t\t\t

    Certifying Fairness<\/h2>\n

    When developing a certification process, it is customary to establish general requirements for a broad range of applications and subsequently test if each application meets them. However, due to the above-mentioned intricacies surrounding fairness, this method cannot be applied to certifying the fairness of AI applications. In this case, a much more flexible approach is required. Developers of an AI application cannot simply test their application against the fairness criteria since they first have to define them for their application15<\/a><\/sup>. This has to be accomplished in a transparent and comprehensible way. Moreover, the details of the application and, most importantly, the setting in which the application is intended to be used must be considered. Questions that need to be answered as a prerequisite for certifying fairness are:<\/p>\n