Fairness and Bias in Algorithmic Hiring: A Multidisciplinary Survey

In the paper

“Representation in Positive Predicted (RPP) [4, 132, 199] is a measure used in noncooperative audits, where the overall population of interest for an algorithm is unknown.”

What the sources say

• Partly supported

  The auditing section describes outside auditors treating unavailable algorithms as black boxes, and the measuring-statistical-significance section (p. 199:13) measures representation as the fraction of an attribute group among people reached by an ad. The paper does not call this measure “Representation in Positive Predicted (RPP)” or say it is used specifically when the overall population of interest is unknown. (LLM, conf=0.80)

  cites [4] [4] Muhammad Ali, Piotr Sapiezynski, Miranda Bogen, Aleksandra Korolova, Alan Mislove, and Aaron Rieke. 2019. Discrimination through Optimization: How Facebook's Ad Delivery Can Lead to Biased Outcomes. Proc. ACM Hum. Comput. Interact. 3, CSCW (2019), 199:1-199:30. https://doi.org/10.1145/3359301 - sourceExperimental — accuracy not yet measured
• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [132]
• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [199]

In the paper

“which prohibit basing hiring decisions on protected attributes [2, 158]”

What the sources say

• Supported

  The section on the broader scope of direct discrimination describes US disparate treatment as addressing discriminatory intent or explicit classification. The Direct Algorithmic Discrimination section says that using protected characteristics in decisions is a clear-cut case of direct discrimination and discusses selection in job applications on the basis of race or sex. The paper also cautions that EU direct discrimination is broader than US disparate treatment. (LLM, conf=0.88)

  cites [2] [2] Jeremias Adams-Prassl, Reuben Binns, and Aislinn Kelly-Lyth. 2023. Directly discriminatory algorithms. The Modern Law Review 86, 1 (2023), 144-175. - source
• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [158]

In the paper

“This approach probably gained popularity because it aligns with legislation against disparate treatment (US) and direct discrimination (EU), which prohibit basing hiring decisions on protected attributes [2, 158] and because it does not require run-time access to sensitive attributes.”

What the sources say

• Partly supported

  The Introduction (page unavailable) says that treating people less favourably because of a protected characteristic constitutes direct discrimination, supporting the legal premise. The Direct Algorithmic Discrimination section (page unavailable) adds that using a proxy can also constitute direct discrimination. Our copy does not establish why proxy reduction gained popularity or whether it avoids run-time access to sensitive attributes. (LLM, conf=0.84)

  cites [2] [2] Jeremias Adams-Prassl, Reuben Binns, and Aislinn Kelly-Lyth. 2023. Directly discriminatory algorithms. The Modern Law Review 86, 1 (2023), 144-175. - sourceExperimental — accuracy not yet measured
• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [158]

In the paper

“The prevalent human vs. algorithm evaluation framework is of limited utility; to overcome it, more research on recruitermachine interaction is required [240], including candidate screening models [6] leading to more granular measures.”

What the sources say

• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [240]
• Partly supported

  The Introduction (section 1, page unavailable) notes calls to consider the human user's behavior, and the Conclusion (section 5, page unavailable) proposes further work on human decision-making in candidate screening. These support more research on recruiter–machine interaction, but the paper does not say that the prevalent human-versus-algorithm evaluation framework is of limited utility. (LLM, conf=0.83)

  cites [6] [6] Jose M Alvarez and Salvatore Ruggieri. 2023. The Initial Screening Order Problem. arXiv preprint arXiv:2307.15398 (2023). - sourceExperimental — accuracy not yet measured

In the paper

“The same model can lead to different results if used in different ways, including the use of different thresholds for classifiers or different cutoffs for rankers.”

In the paper

“Salary Difference (SD) [222] is the only measure of this family proposed in the literature.”

In the paper

“Most of the fairness measures considered in hiring focus on outcome equity; the field is strongly influenced by the 80% rule (Section 8.2), which is appreciated as a quantitative rule of thumb, but also often criticized [74, 267].”

What the sources say

• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [74]
• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [267]

In the paper

“This is especially true for measures of accuracy fairness such as BCRD and MID.”

In the paper

“For example, the version of DI reported in Equation (3) computes the ratio between the acceptance rates of the worse-off group (min) and the best-off group (max), ignoring prior knowledge of structural inequality and affected communities.”

In the paper

“Although it is commonly held, especially in fairness research, that disparities against disadvantaged groups should be mitigated, there is less consensus on how to treat algorithms that happen to favor historically disadvantaged communities.”

In the paper

“Sensitive AUC (sAUC) [33, 119, 120, 194, 196, 222] is a widely used measure of the information about sensitive attributes stored in (proxy) non-sensitive attributes.”

What the sources say

• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [33]
• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [119]
• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [120]
• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [194]
• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [196]
• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [222]

In the paper

“Differently from sAUC and GTR, MIA is an ex-post , model-dependent measure.”

In the paper

“In algorithmic hiring, the literature focuses on the amount of information on sensitive attributes 𝑠 encoded by non-sensitive features 𝑥 and the target variable 𝑦 , with the goal of minimizing it.”

In the paper

“Departing from this approach, process fairness is operationalized with reference to the information on sensitive attributes encoded in the remaining variables.”

In the paper

“Measures of process fairness such as sAUC and MIA detect the presence of sensitive attribute proxies.”

In the paper

“the target variables in Table 4 are very diverse”

In the paper

“Focusing on screening datasets, we find different constructs (e.g. communication skills vs. commitment) annotated by people with different competencies (AMT workers vs. experts) from disparate data sources (YouTube videos vs. mock interviews)”

In the paper

“datasets with geographical information primarily represent US citizens (6 datasets) and the Netherlands (3 datasets)”

In the paper

“Gender is by far the most common sensitive attribute, overwhelmingly encoded as binary”

In the paper

“We found no datasets (and consequently no studies) for selection, and only one for evaluation ( IBM HR Analytics )”

In the paper

“finding one additional dataset (IBM HR Analytics) and no additional article.”

In the paper

“Biased employee evaluation has different forms and derives from multiple causes.”

In the paper

“Since discrimination has shaped residential patterns and influenced public transportation, this factor may have disproportionate impacts along racial and ethnic lines [144].”

What the sources say

• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [144]

In the paper

“This BCF has several concurrent causes, including expectations reinforced by the status quo and lower success in salary negotiation [68, 110, 163].”

What the sources say

• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [68]
• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [110]
• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [163]

In the paper

“This is especially problematic in the hiring domain, where agency is deemed important for leadership positions and it is harder to demonstrate for women without giving an impression of low communion [160].”

What the sources say

• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [160]

In the paper

“While algorithmic hiring represents an opportunity to mitigate these biases, it also runs the risk of reinforcing and amplifying them, causing harm and hindering trustworthiness [168].”

What the sources say

• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [168]

In the paper

“When translated into data and dignified with a ground truth status, vertical segregation leads to models that reinforce wage gaps [222] and lack of diversity in high-status positions.”

What the sources say

• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [222]

In the paper

“Whether, and more importantly what types of , algorithmic hiring can be less biased and more beneficial to society than low-tech alternatives currently remains unanswered, to the detriment of trustworthiness.”

In the paper

“Prior experience is considered a fundamental predictor of suitability for a given position [101].”

What the sources say

• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [101]

In the paper

“First, hiring datasets display measurement errors whose severity varies with protected groups, most critically in target variables related to employability, reflecting current biases in human ratings.”

In the paper

“Finally, there are some higher-order effects caused by technological blindspots and biases in external tools integrated into the hiring pipeline.”

In the paper

“Moreover, biased decisions at early stages result in downstream samples misrepresenting certain groups.”

In the paper

“While algorithmic hiring represents an opportunity to mitigate these biases, it also runs the risk of reinforcing and amplifying them, causing harm and hindering trustworthiness [168].”

What the sources say

• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [168]

In the paper

“Models that predict the likelihood that candidates will accept an offer can reinforce existing wage gaps.”

In the paper

“Two key drivers are stereotypes and employee-manager relationships.”

In the paper

“Gender and race stereotypes about competence drive people's perception on the workplace and potentially bias supervisor evaluations against female and black employees [116, 229, 233, 242].”

What the sources say

• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [116]
• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [229]
• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [233]
• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [242]

In the paper

“These patterns put some job seekers at a systematic disadvantage based on sensitive attributes such as age, disability, gender, religion or belief, racial or ethnic origin, or sexual orientation.”

In the paper

“Elitism in recruiter evaluations favors candidates educated at prestigious institutions who can also list specific extracurricular accomplishments correlated with socioeconomic status [218], further entrenching family status and social class [51, 184].”

What the sources say

• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [218]
• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [51]
• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [184]

In the paper

“For instance, video interviews and resumes contain a wealth of information on gender, race, and other sensitive attributes [63, 69, 189, 218, 222], which can lead algorithms to learn stereotypical associations between sensitive and target variables encoded in the data.”

What the sources say

• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [63]
• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [69]
• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [189]
• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [218]
• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [222]

In the paper

“Job segregation (i.e. the world today) plays a fundamental role in hiring decisions (i.e. the world tomorrow).”

In the paper

“Predominantly analyzed with respect to binary gender [53, 82], recent studies have also highlighted glass ceiling effects for non-binary workers [62], racial minorities [117], and intersectional identities [273].”

What the sources say

• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [62]
• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [117]
• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [273]

In the paper

“Overall, disparities may exist in data sample composition and feature values across sensitive groups that reflect and amplify problematic structural differences in society.”

In the paper

“Workplace proximity and availability of reliable commutes can influence job satisfaction [201] and candidate-employer interactions [262], with amplifying effects introduced by technology [142].”

What the sources say

• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [201]
• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [262]
• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [142]

In the paper

“The average job posting yields more than 100 candidates [101, 206].”

What the sources say

• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [101]
• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [206]

In the paper

“Cultural fit is often considered predictive of a candidate's ability to conform and adapt to the core values and collective behaviors of an organization.”

In the paper

“Sensitive attribute proxies allow models to learn and reflect the diverse BCFs described in the following section.”

In the paper

“For example, agency (orientation towards leadership and goal attainment) is stereotypically associated with men, while communion (warmth and propensity for care) is frequently associated with women [72, 118], with far-reaching effects of gender roles and expectations in employment [118, 215].”

What the sources say

• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [72]
• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [118]

In the paper

“Unfortunately, most work in this space provides partial treatment, often constrained by two competing narratives, optimistically focused on replacing biased recruiter decisions or pessimistically pointing to the automation of discrimination.”

In the paper

“The debate on this topic is often polarized between techno-enthusiasm [139, 177] and pessimism [9, 132] due to partial perspectives on a field that is large and complex.”

What the sources say

• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [139]
• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [177]
• ?Not checked — no full text of the sourcecites [9] [9] Lori Andrews and Hannah Bucher. 2022. Automating Discrimination: AI Hiring Practices and Gender Inequality. Cardozo L. Rev. 44 (2022), 145. -
• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [132]

In the paper

“By one count, there are over 250 Artificial Intelligence (AI) tools for HR on the market [99]”

What the sources say

• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [99]

In the paper

“This dimension is strongly connected to world views [100, 123], and the extent to which differences in variables between sensitive groups are influenced by measurement errors and unjust social structures.”

What the sources say

• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [100]
• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [123]

In the paper

“This approach may be easier to interpret [286], but fails to capture large disparities when acceptance rates are small.”

What the sources say

• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [286]

In the paper

“For example, an increased representation of a minority group in a company can improve job satisfaction, increase their importance and visibility, and encourage other minority applicants in future hiring rounds.”

What the sources say

• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [155]
• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [174]

In the paper

“This fact may be reflected in data sets as a lower tenure for these groups, which can be penalized by algorithms trained to maximize tenure in order to reduce hiring costs and retain human capital.”

In the paper

“It should be noted that the platform(s) chosen to run a campaign can introduce a further bias in favor of its predominant demographics; campaigns run on platforms that cater to younger users, for example, are less likely to reach older segments of the population.”

In the paper

“Asynchronous video interviews can produce specific patterns from candidates with speech impairment (e.g. short answers) or mismeasure input from candidates with sight impairment (e.g. eye contact), which are judged unfavorably by algorithmic recruitment models [189].”

What the sources say

• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [189]

In the paper

“Fitting models to quantify the influence of a sensitive feature on an outcome variable is typical of situations with limited access to the model(s) responsible for the outcomes. This is a common setting for external noncooperative audits.”

In the paper

“Different interpretations have been advanced, including lower risk aversion and the perceived chance of success [110, 122].”

What the sources say

• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [110]
• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [122]

In the paper

“On the contrary, job descriptions with unrealistic requirements discourage job seekers who lack one or more requirements, with a repulsive effect that can be stronger for candidates from vulnerable groups [101, 179].”

What the sources say

• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [101]
• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [179]

In the paper

“This happens even though advertisers design neutral campaigns and is exacerbated if they target specific attributes [258], increasing the opportunity for bias and opacity at the sourcing stage.”

What the sources say

• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [258]

In the paper

“Explicitly mentioning requirements such as community outreach can signal the posture of an employer and act as a pull factor for minority job seekers [192, 234].”

What the sources say

• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [192]
• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [234]

In the paper

“This measure is closely related to equal opportunity [114] and separation [17] and presupposes the availability of a ground-truth variable 𝑦 to condition on:”

What the sources say

• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [114]
• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [17] [17] Solon Barocas, Moritz Hardt, and Arvind Narayanan. 2019. Fairness and Machine Learning: Limitations and Opportunities . fairmlbook.org. http://www.fairmlbook.org. -

In the paper

“Salary negotiation differences between men and women are documented, including in propensity [163] and strategy [110].”

What the sources say

• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [163]
• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [110]

In the paper

“For example, while a measure providing a contextualized and accurate operationalization of a construct defined in the law displays normativity, the measure will inherit all the limitations of the underlying law.”

In the paper

“Well-connected candidates have an inherent advantage over candidates with lower centrality; this advantage may be amplified by algorithms for link prediction [264].”

What the sources say

• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [264]

In the paper

“More general examples include position bias that causes underexposure risks for candidates that are not at the top of a ranking [60] and automation bias leading to over-reliance on technology and reduced human oversight [106, 156].”

What the sources say

• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [106]
• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [156]

In the paper

“These measures are rare in the literature, as they require a broader understanding and modeling of the sociotechnical system around the decision-making algorithm.”

In the paper

“This reflects on the visibility and perceived competence of candidates at the different stages of the hiring pipeline; it can be especially difficult to subvert for women due to the social and economic penalties incurred for violating female gender stereotypes [182].”

What the sources say

• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [182]

In the paper

“Measures in this family lack the notion of preferable outcomes and candidate benefits.”

In the paper

“Moreover, platforms are more likely to have rich profiles and metadata for their common users, while lacking information for less engaged job seekers, who are penalized due to missing data.”

In the paper

“Furthermore, people with a migration background are less likely to own a motor vehicle [146].”

What the sources say

• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [146]

In the paper

“These databases contain a wealth of pre-conviction information, such as arrest data [157] whose validity as a proxy for crime is questionable [97].”

What the sources say

• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [157]
• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [97]

In the paper

“With time, this contributes to reducing vertical segregation, with minority employees in key positions shaping company culture, and attracting more candidates with a similar background.”

What the sources say

• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [155]
• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [174]

In the paper

“Intersectional identities are particularly vulnerable as they are likely to be affected by multiple BCFs.”

In the paper

“It is worth noting that the next three sections are based on a systematic review of the literature summarized in Appendix A.”

In the paper

“It is important to note that these factors are interlocking and overlapping insofar as BCFs are mutually reinforced by the structure of institutions (Section 3.1), individual preferences (Section 3.2), and technological blind spots (Section 3.3).”

In the paper

“This causes an overrepresentation of the former in the training data, leading to rich-get-richer dynamics in job platforms [185].”

What the sources say

• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [185]

In the paper

“This difference may be related to gender roles with respect to household and childcare responsibilities.”

In the paper

“Social networks topology influences job seekers' awareness of openings [96], and their likelihood of being successfully screened by recruiters due to a successful referral [136, 284].”

What the sources say

• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [96]
• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [136]
• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [284]

In the paper

“Accessibility issues and ableist norms can discourage disabled people from job applications [226] and tilt evaluations against them [245, 246].”

What the sources say

• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [226]
• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [245]
• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [246]

In the paper

“Off-the-shelf algorithms from these domains integrated into hiring pipelines [141] are likely to underserve minority candidates for feature extraction and negatively affect algorithms that are based on these feature.”

What the sources say

• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [141]

In the paper

“Note that minority job seekers may suffer a lower quality of service, leading to disengagement and triggering a negative feedback loop by iteratively lowering training representation, quality of service, and engagement.”

In the paper

“Background check tools allow employers to obtain additional information on candidates from public records [244].”

What the sources say

• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [244]

In the paper

“They are all conditional on the target variable 𝑦 , inheriting the biases encoded in the ground truth.”

In the paper

“Similar criticism can be moved to accuracy fairness, with the additional limitation that it is less interpretable and less aligned with the desiderata of job seekers.”

In the paper

“Dedicated proxy reduction approaches (Section 5) can mitigate the resulting risk of discrimination.”

In the paper

“this method is based on the observation that most intrinsically gendered information in word embeddings, such as the difference between the vectors for 'mother' and 'father', is contained within a small gender subspace.”

What the sources say

• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [32]

In the paper

“In terms of outcome fairness, Name Decorrelation is found to improve TPRD and RMS with respect to gender but not ethnicity.”

What the sources say

• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [119]

In the paper

“this approach achieves mitigation on a held-out test set during development, but provides no guarantee at deployment.”

In the paper

“Balanced sampling, for example, can counter the effect of under-representation in training sets caused by factors such as horizontal and vertical job segregation.”

In the paper

“It consists of short resumes containing a binary gender indicator (he/she) and a two-sentence description of job skills for IT, finance, and administration.”

What the sources say

• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [283]

In the paper

“Impact fairness can highlight nuanced aspects of hiring beyond selection rates; SD, for instance, captures wage gaps and salary negotiation differences prompting further analysis into remuneration packages.”

In the paper

“While this method can mitigate large outcome disparities, it does not produce convincing improvements when the inequity is less significant [33, 119, 120].”

What the sources say

• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [33]
• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [119]
• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [120]

In the paper

“The dataset comprises approximately 900,000 resumes (without names, emails, and URLs) in the historical hiring records of eight IT firms based in the US, relevant to just over 6,000 job postings from the IT sector.”

What the sources say

• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [196]

In the paper

“To exemplify, a combination of job segregation, work gaps, and disparate performance of language processing tools, even if relatively weak on their own, may trigger a very low (unfair) value of DI when they compound.”

In the paper

“These are likely two key factors explaining why post-processing is less popular and why hybrid approaches combining different types of fairness interventions, e.g. mitigating via preand post-processing, remain under-explored [11].”

What the sources say

• Not checked yetcites [11] [11] Adam Mehdi Arafan, David Graus, Fernando P Santos, and Emma Beauxis-Aussalet. 2022. End-to-End Bias Mitigation in Candidate Recommender Systems with Fairness Gates. In 2nd Workshop on Recommender Systems for Human Resources, RecSys-in-HR 2022 . CEUR-WS, 1-8. -

In the paper

“notably, their sAUC is on par with adversarial methods explicitly trained to predict gender and ethnicity,”

What the sources say

• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [120]

In the paper

“DPG Resumes [222] contains over 10 million vacancies (including salary range, working hours, and job descriptions) and just under 1 million resumes augmented with job categories of interest and gender inferred from first names.”

What the sources say

• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [222]

In the paper

“Bias in Bios , [63] is composed of textual biographies written in English and extracted from the Common Crawl dataset.”

What the sources say

• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [63]

In the paper

“Representational fairness (GBS) captures biases in representations; when applied to job descriptions, extreme values highlight stereotypically gendered language, signaling a risk of culture-based avoidance.”

In the paper

“Having access to sensitive attributes at runtime enables precise and reliable interventions, whereas strategies that do not rely on this data must be continuously monitored to ensure they are meeting their intended goals.”

In the paper

“Proxy reduction methods (e.g. decorrelation, adversarial inference) target sensitive attribute proxies, reducing the overreliance of models on sensitive information.”

In the paper

“This approach represents an intermediate view on conditional discrimination: it prohibits inter-group discrimination based on a specific feature, while allowing it as a basis for intra-group discrimination.”

In the paper

“Visual signals have been removed from several products [175] because they inevitably encode sensitive information such as race and gender while lacking a solid foundation to justify their use in the hiring domain.”

What the sources say

• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [175]

In the paper

“Even if found to be accurate and fair in a specific evaluation, hiring algorithms based on face analysis are unlikely to predictably generalize and maintain accuracy or fairness under variable conditions, or at least they are less likely to do so than algorithms based on more established data modalities and better-understood correlations with job performance.”

In the paper

“Finally, output re-ranking can mitigate the joint effect of different bias conducive factors.”

In the paper

“The effectiveness of Face Decorrelation for outcome fairness is more nuanced, showing positive effects on DI for very unfair models based on the video modality and limited effects on more equitable systems based on language and audio.”

What the sources say

• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [120]

In the paper

“this result suggests that targeting this type of process fairness can improve outcome fairness in highly imbalanced situations, while only providing limited benefits in situations with lower inequity.”

What the sources say

• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [119]
• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [120]

In the paper

“Gender and race, for example, are considered sensitive attributes due to the recurrent structural disadvantages incurred by women and black people.”

In the paper

“Furthermore, specific data modalities, such as text and vision, facilitate specialized methods like subspace projection and face decorrelation, which function without sensitive attribute information about data subjects.”

In the paper

“Group norming seems applicable to counter systematic biases encoded in input features against protected groups, such as biased employee evaluations.”

In the paper

“Conversely, pre-processing and in-processing approaches, such as importance-based scraping and adversarial inference, require access to sensitive attributes only during training, not testing.”

In the paper

“No single measure provides a complete picture. Different fairness flavors have complementary strengths and weaknesses.”

In the paper

“Indeed, outcome fairness is based on a narrow view of algorithmic hiring as a single decision point abstracted away from its context, where equity is purely a function of algorithmic estimates and (sometimes) their similarity to target variables approximating a ground truth.”

In the paper

“while Peña et al. [199] demonstrate SKL improvements in a multimodal setting based on synthetic resume ranking .”

What the sources say

• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [199]

In the paper

“Impact fairness quantifies the downstream impacts of algorithmic outcomes on different populations; modeling important aspects of the surrounding socio-technical system, it goes beyond brittle algorithmic abstraction.”

In the paper

“It is composed of applications, i.e. pairs of resumes and job postings for engineers (software and hardware) and technical scientists.”

What the sources say

• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [58]

In the paper

“It is worth noting that post-processing explicitly takes into account sensitive attributes to change algorithmic outcomes for job candidates, which may be critical for disparate treatment and affirmative action (US) as well as direct discrimination and positive action (EU) [22, 112].”

What the sources say

• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [22]
• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [112]

In the paper

“Both variants of Adversarial Removal (NS, MSE) are shown reduce the sensitive information encoded in latent representations:”

What the sources say

• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [120]

In the paper

“Both sAUC and GTR are ex-ante measures, i.e. they can be computed from the data before the final model is trained; therefore, the granularity dimension does not apply to them.”

In the paper

“Geyik et al. [103] decouple their post-processing method from specific model choices and properties of input data, which means their method can naturally scale across the different (ranking) systems at LinkedIn”

What the sources say

• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [103]

In the paper

“datasets that target employee tenure and performance, such as Walmart Employees and IBM HR Analytics , signal an interest from companies in understanding the factors that predict future productivity and loyalty”

In the paper

“Artificial target scores are generated for each CV, as a linear combination of numerical features; biased scores are derived from these with an ethnicity- and gender-dependent additive penalty emulating biases in the data.”

What the sources say

• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [199]

In the paper

“This can be especially problematic when using conditional fairness measures whose very definition hinges on this so-called ground truth”

In the paper

“GPT-4 is found to more frequently use female pronouns in reference letters generated for female-dominated occupations (such as nannies), and male pronouns for male-dominated occupations (such as plumbers) [35]”

What the sources say

• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [35]

In the paper

“This challenge may explain in part the popularity of disparate impact as a fairness metric, due to its ease of interpretation across a broad and diverse stakeholder group, or the popularity of complying with the EEOC's 80% rule [77] as a fairness target (adopted e.g. by Pymetrics [271]), as it appears, at least superficially, a legally grounded target”

What the sources say

• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [77]
• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [271]

In the paper

“this allows these methods to be deployed and run in isolation from production systems in controlled batch scenarios, which can be another important practical benefit”

In the paper

“a third family of bias mitigation methods that can operate without requiring access to sensitive attributes at all are the rule-based approaches described above, which are commonly used for mitigating bias in textual data such as job descriptions or resumes”

In the paper

“Defining fairness and formulating an intervention target is not a one-off task, as algorithmic hiring components will be deployed in different, product-specific contexts and stages of the hiring pipeline”

In the paper

“Geyik et al. [103] find their DetGreedy method improves their fairness metric, with no significant impact on utility”

What the sources say

• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [103]

In the paper

“Significantly, the Equal Employment Opportunity Commission (EEOC) issues nonbinding, but practically important guidelines to interpret Title VII.”

In the paper

“Examples of these bias mitigation methods used by practitioners include adversarial inference [222], and re-balancing training data with synthetic data generation [11]”

What the sources say

• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [222]
• Not checked yetcites [11] [11] Adam Mehdi Arafan, David Graus, Fernando P Santos, and Emma Beauxis-Aussalet. 2022. End-to-End Bias Mitigation in Candidate Recommender Systems with Fairness Gates. In 2nd Workshop on Recommender Systems for Human Resources, RecSys-in-HR 2022 . CEUR-WS, 1-8. -

In the paper

“These methods rely solely on domain and task-specific gazetteers or dictionaries, and hence can be used when access to sensitive attributes of individuals is not available or desirable”

In the paper

“This type of adversarial bias mitigation can, however, be applied over a wider set of data modalities”

In the paper

“the challenge in defining fairness and formulating intervention targets in algorithmic hiring is exacerbated by the multi-stakeholder nature of the hiring domain”

In the paper

“pre-processing bias mitigation methods may prove more useful, as they (only) require access to sensitive attributes at training time, and can operate on attributes even when available for only a subset of job seekers”

In the paper

“tabular data is a common data modality in algorithmic hiring and machine learning more broadly”

In the paper

“the validity of these estimates and their connection with job performance has been called into question [19, 214]”

What the sources say

• ?Not checked — no full text of the sourcecites [19] [19] Murray R Barrick and Michael K Mount. 1991. The big five personality dimensions and job performance: a metaanalysis. Personnel psychology 44, 1 (1991), 1-26. - source
• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [214]

In the paper

“It comprises AVIs with 154 students from Utrecht University carrying out a mock interview with a fictitious oil company.”

What the sources say

• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [139]

In the paper

“For classification tasks, getting started with bias mitigation is relatively straightforward, as there exist several open source or otherwise freely available software packages and libraries, as mentioned above, that provide different implementations of bias measurement methods and mitigation algorithms, specifically designed for classification tasks”

In the paper

“they are consistent with previous research reporting that most efforts to improve fairness in artificial intelligence are influenced by the Global North [187, 221], particularly the US [135], and focus on English language resources [210]”

What the sources say

• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [187]
• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [221]
• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [135]
• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [210]

In the paper

“Once data modality, task, infrastructural and technical choices, and access to sensitive data are set, one important practical challenge is that of defining 'fairness' and formulating the intervention target of a bias mitigation method”

In the paper

“Importantly, disparate treatment requires not only an adverse action based on a protected attribute, but also the proof of intent on the part of the discriminating individual - unlike the EU variety of direct discrimination [254].”

What the sources say

• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [254]

In the paper

“Candela et al. [39] present the (evolving) framework used at LinkedIn that guides definitions and operationalizations of AI fairness across their different (types of) products”

What the sources say

• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [39]

In the paper

“Adversarial inference methods, which have been applied to a variety of data modalities and tasks, even require training multiple neural networks simultaneously, which means they can incur substantially higher costs compared to their non-mitigated counterparts”

In the paper

“The two most common tasks in algorithmic hiring are classification and ranking”

In the paper

“Developers of high-risk AI systems must, for instance, ensure that the training data are appropriate and do not lead to unlawful discrimination.”

What the sources say

• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [84]

In the paper

“In general, post-processing approaches offer an advantage in terms of system integration, but they should be applied with care due to antidiscrimination law”

In the paper

“Ranking is a common task in the sourcing and screening stages of hiring, performed by search engines or recommender systems”

In the paper

“video-based systems tend to yield more disadvantageous effects for protected groups, also due to the strong encoding of sensitive information”

In the paper

“SHL Interviews [231] is a proprietary AVI dataset with more than 5,000 videos from 810 real job seekers from the US, UK, India, and Europe answering behavioral and domain knowledge questions.”

What the sources say

• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [231]

In the paper

“EU law distinguishes two categories of prohibited discrimination: direct and indirect”

In the paper

“incoming regulation proposals that prohibit inference of emotions or intentions from facial data in workplace contexts”

In the paper

“These types of methods are widely available in open-source bias mitigation software packages such as the Fairlearn Python package [268] or the AI Fairness 360 toolkit (available in Python and R) [21]”

What the sources say

• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [268]
• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [21] [21] Rachel K. E. Bellamy, Kuntal Dey, Michael Hind, Samuel C. Hoffman, Stephanie Houde, Kalapriya Kannan, Pranay Lohia, Jacquelyn Martino, Sameep Mehta, Aleksandra Mojsilovic, Seema Nagar, Karthikeyan Natesan Ramamurthy, John Richards, Diptikalyan Saha, Prasanna Sattigeri, Moninder Singh, Kush R. Varshney, and Yunfeng Zhang. 2018.

In the paper

“These mitigation strategies tend to be technically straightforward to implement, as they can be developed as standalone components and need not be integrated in complex algorithmic hiring pipelines”

In the paper

“alignment with the company's goals and values, interpretability (or ability to explain), and collaboration across different stakeholders were mentioned as key factors in guiding the eventual target distribution”

What the sources say

• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [103]

In the paper

“algorithmic efficiency and computational costs will vary across bias mitigation methods in nature and magnitude, with simple rule-based scraping methods that involve dictionary look-ups on the cheap end of the spectrum, and bias mitigation methods that involve re-training LLMs at the resource-intensive end [167]”

What the sources say

• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [167]

In the paper

“This also means that such a method can be developed and deployed as a standalone component or micro-service, which decreases development time, effort, and alignment, when compared to pre-processing or in-processing methods that may need to be designed and integrated in complex algorithmic hiring pipelines”

In the paper

“Companies interested in algorithmic hiring solutions pick one or more of these variables, balancing different priorities such as efficiency and quality of hire, with unpredictable effects on algorithmic fairness, a phenomenon called multi-target multiplicity [266]”

What the sources say

• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [266]

In the paper

“Student Interviews [33] consists of video interviews with 733 upper-level undergraduate students,”

What the sources say

• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [33]

In the paper

“There is a proposal for an AI Act in the EU, with many requirements for 'high-risk' systems, including algorithmic systems for HR [84].”

What the sources say

• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites [84]

In the paper

“the prevalence of classification-based bias metrics in algorithmic hiring (discussed in more detail in Section 4) is reflected in the aforementioned open source packages, which means that measuring and mitigating bias in ranking systems tends to follow the practice of re-purposing classification methods through, e.g., thresholding rankings”

In the paper

“access to sensitive attributes may be at odds with privacy regulations such as the GDPR”

In the paper

“Direct discrimination is always prohibited, with a few narrowly defined and specific examples”

In the paper

“Many other sensitive attributes (e.g., religious beliefs) will not be easily available, and will have to be explicitly requested for bias measurement and mitigation purposes”

In the paper

“Textual data is a common data modality encountered in the sourcing stage, typically in the form of textual CV or job description data, or in the screening stage, e.g. transcripts of video interviews”

In the paper

“different data modalities enable different bias mitigation strategies”

In the paper

“Mitigating bias in tabular data can be done through pre-processing approaches that directly intervene on the features, e.g. by removing features that are highly correlated with sensitive attributes via importance-based scraping, or increasing the representation of vulnerable populations in training sets with balanced sampling”

In the paper

“In the sourcing or screening stages, depending on the intervention target, outcome fairness may come at the cost of utility; this can happen, for example, when optimizing for gender parity in heavily male or female-dominated industries”

In the paper

“The proposals in the AI Act also contain a new exception to the GDPR, to enable the use of sensitive data for debiasing algorithmic systems.”