Active Fairness Instead of Unawareness
aa r X i v : . [ c s . A I] S e p Active Fairness Instead of Unawareness
Boris Ruf and Marcin Detyniecki ∗ AXA Research, Paris, FranceFebruary 2020
Abstract
The possible risk that AI systems could promote discrimination byreproducing and enforcing unwanted bias in data has been broadly dis-cussed in research and society. Many current legal standards demand toremove sensitive attributes from data in order to achieve “fairness throughunawareness”. We argue that this approach is obsolete in the era of bigdata where large datasets with highly correlated attributes are common.In the contrary, we propose the active use of sensitive attributes with thepurpose of observing and controlling any kind of discrimination, and thusleading to fair results.
Systematic, unequal treatment of individuals based on their membership of asensitive group is considered discrimination. There is broad consensus in oursociety that it is unfair to make a distinction on the ground of a personal charac-teristic which is usually not a matter of choice. Therefore, most legal frameworksprohibit such actions. When it comes to non-discrimination in the EU, for ex-ample, the
Convention for the Protection of Human Rights and FundamentalFreedoms defines the “Prohibition of discrimination” in Article 14 [1]. This prin-ciple is further contained in the
Charter of Fundamental Rights of the EuropeanUnion which states in Article 21 that “Any discrimination based on any groundsuch as sex, race, colour, ethnic or social origin, genetic features, language, reli-gion or belief, political or any other opinion, membership of a national minority,property, birth, disability, age or sexual orientation shall be prohibited.” [2]In automatic decision-making, the traditional approach to fight discrimina-tion is known as “anti-classification” among legal scholars [3]. This principleattempts to obtain “fairness through unawareness” by simply excluding anysensitive attributes as features from the data. In EU law, this is enforced onthe level of data protection: The General Data Protection Regulation (GDPR) ∗ { boris.ruf,marcin.detyniecki } @axa.com When it comes to AI systems, the concept of removing sensitive attributes fromdata in order to prevent algorithms from being unfair has proven particularlyinsufficient: Such systems are usually backed by high-dimensional and stronglycorrelated datasets. This means that the decisions are based on hundreds oreven thousands of attributes whose relevance is not obvious at first glance forthe human eye. Further, some of those attributes usually contain strong linkswhich again are difficult to spot for humans. Even after removing the sensitiveattributes, such complex correlations in the data may continue to provide manyunexpected links to protected information. In fact, heuristic methods exist toactively reconstruct missing sensitive attributes. For example, the Bayesian Im-proved Surname Geocoding (BISG) method attempts to predict the race giventhe surname and a geolocation [6]. While the reliability of this method is gen-erally disputed, it demonstrates that prohibiting to collect sensitive attributesdoes not prevent any possible misuse just by technical design.But even without any bad intent to discriminate, there is a danger of hiddenindirect discrimination which is very difficult to detect in the results [7, 3].To illustrate the problem, we imagine an AI system which analyses CVs inorder to propose starting salaries for newly hired staff. We further assume thatwomen were discriminated in the past because their salaries were systematicallylower compared to those of their male colleagues. Historical bias of this kindcannot be overcome by excluding the sensitive attribute “gender” in the learningdata since many links to non-sensitive attributes exist. For example, somehobbies or sports may be more popular among women or men. In languages withgrammatical gender, the applicant’s gender may be revealed through genderinflections of nouns, pronouns or adjectives. And yet more complex, in a countrywith compulsory military service exclusively for men, the entry age at university2ould provide a hint to the gender, too. Even when trying to additionally adjustfor all of those identified correlations manually, it remains impossible to establisha sufficient degree of ”unawareness” which could guarantee discrimination-freedecisions.
We acknowledge the risk to privacy protection when storing sensitive attributesand we understand the motivation behind the current rules to address thoseconcerns. However, based on the considerations in the previous sections, weconclude that the current practice of trying to ignore the existence of sensitivesubgroups by omitting sensitive attributes bares greater risk than any privacyconcerns related to the data collection. We therefore suggest to re-examine thestatus quo and propose the active use of sensitive attributes in AI systems to make sensitive subgroups visible and account for them with the purpose ofverifiable fair results. Such a paradigm shift would allow for statistical measuresto detect any type of discrimination and make it possible to mitigate unwantedbias in the data. Allowing to collect the sensitive attributes would be a helpfulstep towards verification mechanisms for AI stakeholders to test for imbalancedresults, as well as for third parties such as regulators who could audit the data toensure non-discrimination of underprivileged subgroups. To protect the sensitiveattributes from misuse, new technical security mechanisms such as restrictedaccess rules could be established.In a nutshell, the principle of “active fairness” as opposed to “fairnessthrough unawareness” would lay the foundations for tools which ensure thatthe standards of fairness and non-discrimination in AI systems are respected.Ultimately, access to such instruments would clear the way for increased trustin AI systems in society and could contribute to overcome a problem which hasplagued humanity ever since: human bias. Fixing biases in algorithms remainsa technical problem which is complex but still far easier to solve than correctingcognitive bias. If we succeed in developing automated systems which can help ustaking fair, impartial decisions, the potential contribution for human progressand the protection and support of disadvantaged groups will be enormous.
References [1] Council of Europe.
Convention for the Protection of Human Rights andFundamental Freedoms , volume Rome, 4.XI. Council of Europe, 1950.[2] European Union.
Charter of Fundamental Rights of the European Union ,volume 53. Brussels, 2010.[3] Sam Corbett-Davies and Sharad Goel. The measure and mismeasure offairness: A critical review of fair machine learning.
CoRR , abs/1808.00023,2018. 34] European Union. Regulation (eu) 2016/679 of the european parliament andof the council of 27 april 2016 on the protection of natural persons withregard to the processing of personal data and on the free movement of suchdata, and repealing directive 95/46/ec (general data protection regulation).
Official Journal of the European Union , L119:1–88, 5 2016.[5] Alex Gano. Disparate impact and mortgage lending: A beginner’s guide.
U.Colo. L. Rev. , 88:1109, 2017.[6] Marc N Elliott, Allen Fremont, Peter A Morrison, Philip Pantoja, and NicoleLurie. A new method for estimating race/ethnicity and associated dispar-ities where administrative records lack self-reported race/ethnicity.
Healthservices research , 43(5 Pt 1):1722—1736, 10 2008.[7] Sara Hajian and Josep Domingo-Ferrer. A methodology for direct and indi-rect discrimination prevention in data mining.