Fairness and Bias in Algorithmic Hiring: a Multidisciplinary Survey

1 INTRODUCTION

• The survey treats algorithmic hiring as both a possible way to mitigate workplace bias and a risk for reinforcing or amplifying discrimination against vulnerable groups.It responds to a field often split between techno-enthusiasm and pessimism with a multidisciplinary analysis of systems, methods, resources, challenges, and opportunities.Source: §1 · p. 2
• Compared with earlier reviews, the survey broadens coverage beyond selected bias sources, non-empirical HR research, or nine mitigation articles to measures, datasets, and social and legal context.It also covers more hiring tasks and stages than recruitment-focused fair recommender-system work, which emphasizes ranking.Source: §1 · p. 2
• Its integrated coverage gives practitioners domain-specific bias factors, mitigation methods, practical guidance, and legal references, while supporting researchers with technology, measures, and datasets.The survey also aims to serve non-experts and identify gaps and future research directions at the intersection of computer science, law, and policy.Source: §1 · p. 2

2 THE ALGORITHMIC HIRING PIPELINE

• Algorithmic hiring comprises tools that automate or assist recruitment and evaluation decisions, organized here into four stages of a hiring pipeline.The stage boundaries are elastic and depend on use; for example, CV-scoring tools may serve sourcing or screening.Source: §2 · p. 3, Figure 1 · p. 3, Sourcing · p. 4

Sourcing

• Sourcing systems include job-ad tools, job-to-candidate and candidate-to-job search or recommendation, and social networks that affect candidate visibility and referrals.These systems use methods such as machine learning, Boolean matching, sorting, and information extraction; fairness is a critical, well-studied property of search, ranking, and recommendation.Source: Sourcing · pp. 3–4

Screening

• Screening tools narrow candidate pools through games, asynchronous video interviews, questionnaires, and chatbots that assess traits, infer performance, filter candidates, or schedule interviews.Video models may analyze visual, verbal, and paraverbal features; large language models may extend chatbot influence to other recruitment stages.Source: Screening · p. 4

Selection

• Selection follows screening and may use offer optimization, background checks, and team placement to support candidate offers and assignments.Offer tools estimate acceptance likelihood from compensation and benefits; checks may examine criminal records or social media.Source: Selection · pp. 4–5

Evaluation

• Post-hiring tools support employee development, monitoring, engagement analytics, and career progression, including training assignment, work allocation, and career development.Tracking supports monitoring and estimation; questionnaires and surveys measure satisfaction, commitment, and retention; succession plans and promotions can be partially automated, while turnover prediction often aims to favor retention.Source: Evaluation · p. 5
• Turnover data used to improve hiring systems can also be repurposed for employee termination; 98% of 300 surveyed US HR leaders said algorithms would assist layoff decisions in 2023.Source: Evaluation · p. 5
• Evaluation decisions can affect tenure, which is a key prediction target for sourcing and screening algorithms, while complex influencing factors may produce undesirable discrimination.Hiring-stage decisions also generate data that affect later stages and subsequent interactions.Source: Evaluation · p. 5

3 BIAS CONDUCIVE FACTORS

• Measurement errors, early-stage decisions, stereotypes, and sensitive-attribute proxies can encode group disparities in hiring data and algorithmic outcomes.Human ratings can bias employability targets, while earlier decisions can produce downstream samples that misrepresent groups. Stereotypes may be activated indirectly through correlated inputs such as video appearance, voice, names, and languages in CVs.Source: §3 · pp. 5–6

3.1 Institutional biases

• Institutional practices include direct discrimination, occupational and career segregation, cultural-fit evaluations, and preferences for elite educational credentials.Matched-pair and correspondence studies report discrimination against Black applicants, women, non-white, disabled, and older applicants, as well as preferences for standard American accents. Cultural-fit assessments can sustain uniform workforces, while segregation can reinforce wage gaps and underrepresentation in senior positions.Source: §3.1 · pp. 6–7
• Workplace evaluations, wage expectations, and social networks can reproduce institutional inequalities through stereotypes, homophily, and unequal access to referrals and connections.Gender and race stereotypes can bias employee evaluations, while manager-employee homophily may disadvantage minority employees in promotion and ratings. Salary expectations and negotiation can reinforce wage gaps, and referral and connection systems may favor well-connected candidates; one professional-network recommender under-recommended women.Source: §3.1 · p. 7

3.2 Individual preferences

• Seemingly individual preferences can reflect group-level disadvantage, including lower tenure, self-promotion gaps, commute constraints, salary negotiation differences, and discouraging job descriptions.Discrimination and harassment can contribute to lower tenure for transgender, nonbinary, female, Black, and disabled workers, which tenure-maximizing algorithms may penalize. Gendered wording and unrealistic requirements can discourage applicants, while caregiving-related work gaps disadvantage women.Source: §3.2 · pp. 7–8

3.3 Technology blindspots

• Integrated hiring technologies can introduce or amplify disparities through biased ad delivery, inaccessible assessments, uneven tool performance, missing data, and human-technology interactions.Ad optimization can skew job-ad audiences even for neutral campaigns, while video interviews and language or vision tools may underserve disabled and minority candidates. Engagement gaps and missing profile data can create feedback loops, and ranking position or automation bias can reduce exposure and human oversight.Source: §3.3 · pp. 8–9

3.4 Overlaps and Interactions

• BCFs interlock across institutions, individual preferences, and technology, leaving people with intersectional identities especially vulnerable to compounding barriers.The example of a woman with children and a migration background illustrates barriers spanning ad reach, commute, profile data, screening, cultural fit, parental-leave assumptions, and wage setting. The authors argue that ameliorating one factor may generate positive spillovers that reduce others.Source: §3.4 · pp. 9–10

4.1 Dimensions of Fairness Measures

• The framework distinguishes outcome, accuracy, impact, process, and representational fairness, alongside conditionality, granularity, normativity, interpretability, and multinarity.Its shared notation describes non-sensitive features, targets, sensitive groups, scores, rankings, and desired representation.Source: §4.1 · pp. 10–11, §4.2 · p. 11

4.3.1 Outcome Fairness.

• Outcome fairness is most common; its measures address selection, ranking, scores, and group representation, while LRR estimates sensitive-attribute effects in external noncooperative audits.DI uses a selection-rate ratio, whereas DD uses a difference that can understate large disparities when rates are small; NDKL can obscure which groups benefit at particular ranks.Source: §4.3.1 · pp. 11–14

4.3.2 Accuracy fairness.

• Accuracy fairness compares groupwise model accuracy conditional on the target, so its measures inherit biases encoded in the ground truth.Examples include regression error measures such as MAE and classification measures such as BCRD and MID.Source: §4.3.2 · pp. 14–15

4.3.3 Impact fairness.

• Impact fairness is rare: Salary Difference is the only surveyed measure and estimates gender differences in average wages associated with matched vacancies.It uses an embedding-based model to match candidates to vacancies and compares average salaries for male and female candidates in the same industry.Source: §4.3.3 · p. 15

4.3.4 Representational fairness.

• Representational fairness remains understudied [1]; Gender Bias Score flags stereotypically gendered language in job descriptions using a gendered-trait word inventory.The measure counts stereotypically masculine and feminine words, making it a coarse indicator of bias in representations.Source: §4.3.4 · p. 15

4.3.5 Process fairness.

• Process measures assess sensitive-attribute information in other variables: sAUC evaluates proxy recoverability, GTR estimates sensitive-attribute relevance to the target, and MIA detects predictive amplification.sAUC and GTR are ex-ante, while MIA is ex-post and model-dependent; sAUC has also been extended to multiclass attributes.Source: §4.3.5 · pp. 16–17

4.4 Discussion

• Measure choice depends on sensitive-attribute cardinality, interpretability, target availability, and operating flexibility; fixed cutoffs favor coarse measures, while uncertain conditions favor granular ones.In the fictitious EquiHire example, these constraints lead to DI for primary monitoring and SD as a complementary wage-outcome measure.Source: §4.4 · pp. 17–18
• No measure gives a complete picture: outcome fairness abstracts from context, accuracy fairness is less interpretable, and process and impact fairness need broader, context-specific research.Measures can support actionable diagnoses, but diagnostic clarity varies: binary outcome and impact measures tend to be more interpretable than accuracy measures or multinary summaries such as RMS.Source: §4.4 · pp. 17–18
• The DI version computes the ratio of the worse-off group’s acceptance rate to the best-off group’s, ignoring structural inequality and affected communities.Source: §4.4 · p. 19
• Although gender and race are sensitive attributes given historical disadvantage and hiring biases against women and ethnic minorities, fairness research lacks consensus on favoring these groups.Source: §4.4 · p. 19

5.1 Pre-processing

• Preprocessing results varied: rule-based scraping moderately reduced TPRD [63], SHAP-based scraping yielded a sizeable sAUC drop with limited performance impact [194], and subspace projection did not reduce resume gender information as measured by sAUC [194].Balanced sampling can reduce the influence of job segregation on hiring algorithms [276]. Group Norming marginally improved DI, had limited fairness impact, risks unstable normalization across small intersectional groups, and is not recommended amid US employment-law controversy.Source: §5.1 · pp. 19–21

5.2 In-processing

• In-processing methods reduce sensitive information in representations, but outcome-fairness benefits depend on the method and the initial inequity.Adversarial inference improved selected fairness indicators in job recommendation and SKL in synthetic resume ranking. Face decorrelation reduced sensitive information, with positive DI effects for very unfair video models but limited effects for more equitable language and audio systems. Name Decorrelation improved gender, but not ethnicity, TPRD and RMS [119].Source: §5.2 · pp. 22–23

5.3 Post-processing

• Post-processing methods alter rankings or scores to meet fairness objectives, with approaches for representation quotas, equalized odds, and screening selection.DetGreedy prioritizes relevant candidates from underrepresented groups; DetCons and DetConsSort address possible quota violations, and DetGreedy is implemented in LinkedIn Recruiter with query-dependent gender targets. CDF Rescoring maps scores to achieve xEO = 0, while Spatial Partitioning rebalances a selected applicant set when sensitive attributes are unknown at test time.Source: §5.3 · p. 24

5.4 Discussion

• Mitigation methods target different bias sources: balanced sampling addresses under-representation, proxy reduction targets sensitive proxies, and output re-ranking addresses joint bias effects.The survey cautions that Group Norming should not substitute for understanding and improving the socio-technical measurement system that produced biased input features.Source: §5.4 · pp. 24–25
• Post-processing is easy to integrate but raises sensitive-data and legal concerns, while video-based hiring warrants particular caution.Post-processing typically needs sensitive attributes at runtime and may implicate disparate-treatment or positive-action rules; pre- and in-processing generally need them only during training, and modality-specific methods may not require them. Video systems show large disparities, and face-based systems may not generalize reliably; the survey cautions against claims that computer vision can infer candidate motivation or personality even when paired with fairness measures.Source: §5.4 · pp. 25–26
• Proxy-reduction methods are popular, but improve outcome fairness most clearly for large disparities and provide no deployment-time guarantee.They target process fairness, such as sensitive-attribute predictability, in pursuit of outcome fairness; the survey calls for more study of effects on vulnerable candidates.Source: §5.4 · p. 25

6.1 Text-based datasets

• Text-based resources range from short skill biographies with binary gender labels to large hiring records, including about 900,000 IT resumes linked to callback outcomes.Resumes and biographies can contain strong proxies for sensitive attributes, including names and addresses.Source: §6.1 · p. 26

6.2 Visual datasets

• Visual resources include real and simulated video interviews and synthetic FairCVs combining biographies, face images, numerical features, and bias-injected target scores.The video datasets include real job seekers and student interviewees; FairCVs support studying how biased targets and sensitive-attribute proxies can contribute to discriminatory models.Source: §6.2 · p. 27

6.3 Tabular data

• Tabular resources include synthetic data with deliberately introduced bias, pre-deployment gameplay audits, employee outcomes, and field studies of hiring tools.Other datasets capture candidate-job matches and recruitment outcomes, employee retention or performance, and differential job-ad delivery.Source: §6.3 · pp. 27–29

6.4 Discussion

• Datasets are concentrated in English and US or Dutch settings, with sourcing (11 of 20) and screening (8 of 20) dominant, no selection datasets, and only one for evaluation.The authors warn that later-stage algorithms may be deployed without clear understanding of their risks and limitations.Source: §6.4 · p. 29
• Disability, religion, and sexual orientation are missing despite workplace discrimination evidence; gender is usually binary, while race or ethnicity appears in 6 of 20 datasets.In textual and visual datasets, race and ethnicity are often annotated rather than self-reported because those modalities carry strong proxies for sensitive attributes.Source: §6.4 · pp. 29–30
• Hiring datasets use diverse targets whose validity and links to job performance are uncertain, making them risky ground truth, especially for conditional fairness measures.The targets range across hiring-pipeline stages and reflect different constructs, annotators, and data sources.Source: §6.4 · p. 30

7.1 Data modalities

• Mitigation choices vary by modality: dictionaries suit text, adversarial methods can remove sensitive visual representations, and pre-processing offers options for tabular data.Text dictionaries can be straightforward standalone components that draw on HR domain knowledge, but their task-specific terms transfer poorly across problems, languages, and geographies. LLMs can reproduce gender and nationality stereotypes in job recommendations and gendered pronouns in generated reference letters; prompting “in an inclusive way” can elicit they/their pronouns [35]. Video-based systems may be unreliable under varying conditions, and proposed workplace restrictions on emotion or intention inference support caution; tabular options include feature removal, balanced sampling, Fairlearn, AI Fairness 360, and synthetic data generation [21].Source: §7.1 · pp. 30–31

7.2 Tasks

• Ranking systems need ranking-aware fairness measures and ranking-specific mitigation rather than simply reusing classification methods through thresholding.Classification mitigation is relatively straightforward because freely available libraries provide measurement and mitigation implementations. Fair ranking evaluation should account for user browsing behavior, suitably modeled; methods such as adversarial inference can apply to both task types, while DetGreedy is ranking-specific.Source: §7.2 · p. 31

7.3 Scalability and efficiency

• Post-processing can be easier to integrate and scale, but its computational cost, like that of other mitigation methods, depends on the method and system.Geyik et al.’s model-agnostic re-ranking can run as a standalone component and process only top-k outputs, though it adds cost per output and requires care under anti-discrimination law. Re-training and adversarial methods can be more resource-intensive, while dictionary lookups are cheap and LLM re-training is at the resource-intensive end [167].Source: §7.3 · p. 32

7.4 Sensitive attribute data availability and usage

• Sensitive-data availability constrains method choice: pre-processing can use attributes available only at training time, while rule-based approaches need no individual sensitive attributes.Privacy regulations such as the GDPR and the difficulty of obtaining consent can limit access; age and gender may be recorded, whereas attributes such as religious beliefs may need to be explicitly requested. Post-processing that requires every candidate’s sensitive attributes at inference time may be unrealistic; pre-processing can operate on a consenting subset in controlled batch settings, with examples including adversarial inference and synthetic-data rebalancing.Source: §7.4 · pp. 32–33

7.5 Fairness definitions and intervention targets

• Fairness targets require context-specific, multi-stakeholder decisions guided by ethical, social, and legal considerations.LinkedIn’s DetGreedy target matched the top-k gender distribution to qualified job seekers, with company goals, interpretability, and stakeholder collaboration informing the choice; targets vary by product and pipeline context. Disparate impact and the EEOC’s 80% rule, adopted by Pymetrics, are popular partly because they are interpretable or appear legally grounded.Source: §7.5 · p. 33

7.6 Fairness vs. utility trade-offs

• DetGreedy improved fairness without significant utility impact, rebalancing training data could improve utility, and overcompensation can reduce downstream human bias.DetGreedy’s result came from online A/B tests, while the rebalancing result came from an offline experiment with real data from an international HR company. Overcompensation means artificially over-representing a gender in ranking output.Source: §7.6 · p. 33

8 LEGAL LANDSCAPE

• The section covers EU and US regulations and non-discrimination provisions, emphasizes EU law as less cited in computer science literature, and closes with technology-policy challenges.Source: §8 · pp. 33–34

8.1 European law

• In the EU, general non-discrimination law applies absent algorithmic-hiring-specific rules, with protections grounded in human-rights instruments and four directives.Together, the directives protect six grounds—age, disability, gender, religion or belief, racial or ethnic origin, and sexual orientation—and prohibit direct and indirect discrimination.Source: Non-discrimination law · p. 34

Non-discrimination law

• Direct discrimination treats someone less favorably because of a protected characteristic and is prohibited except for narrowly defined cases.The passage gives refusing to recruit people of a certain ethnicity as an example, while a women’s clothing brand may hire only female models for advertising.Source: Non-discrimination law · p. 34
• Indirect discrimination arises when a neutral practice disadvantages a protected group and lacks objective justification; the discriminator’s intent or knowledge is irrelevant.A practice may be justified by a legitimate aim only when its means are appropriate and necessary.Source: Non-discrimination law · p. 34
• Whether a neutral hiring rule is indirectly discriminatory depends on job context: rejecting spelling errors may be unjustified for cleaning work but justified for lawyers’ writing duties.The hypothetical cleaning company could not justify its rule because spelling accuracy was not necessary to being a good cleaner, whereas precise writing is central to many lawyers’ work.Source: Non-discrimination law · p. 34
• An organization remains responsible to victims when an algorithmic system it uses discriminates illegally, even if an outside company supplied the system.The organization may later sue the developer, but it remains liable toward the victim; general non-discrimination law covers indirect and accidental algorithmic discrimination.Source: Non-discrimination law · pp. 34–35
• GDPR generally bans processing special-category personal data, with exceptions including explicit consent, which is generally invalid when requested by an employer.This restriction makes sensitive data such as ethnicity difficult to use for auditing or training algorithmic systems.Source: Other relevant law · p. 35

Other relevant law.

• The proposed EU AI Act would impose requirements on high-risk systems, including HR algorithms, and permit sensitive-data use for debiasing under a new GDPR exception.At the time of writing, the Act had not been officially adopted and its final text had not been published.Source: Other relevant law · p. 35

8.2 US law

• US employment-discrimination law includes Title VII and federal statutes addressing age, disability, and genetic information, alongside practically important nonbinding EEOC guidelines.Title VII prohibits employment discrimination based on race, color, religion, sex, and national origin; case law also plays a crucial role.Source: Federal law · p. 35

Federal law

• Disparate treatment requires an adverse action based on a protected attribute and proof of intent, which injured people may find difficult to establish in court.The US doctrine resembles, but does not perfectly mirror, the EU concept of direct discrimination.Source: Federal law · p. 35
• Disparate impact requires no intent and covers neutral practices that significantly disadvantage protected groups; EEOC guidance uses an 80% (4/5) evaluation-rate threshold, but business necessity can justify a practice.A practice may therefore be allowed even if it disproportionately disadvantages a protected group.Source: Federal law · p. 35
• Without proof of intent, algorithmic discrimination claims are likely to rely on disparate impact, though courts assess legally relevant disadvantage using more than statistical scores alone.Source: Federal law · p. 36
• Federal AI-specific discrimination legislation appears unlikely soon; NYC and Illinois have instead enacted laws regulating automated hiring and AI video interviews.Since July 5, 2023, NYC Local Law 144 has required independent bias evaluations for covered tools; Illinois requires notice and consent before AI video analysis, permits deletion requests, and requires race and ethnicity reporting when employers rely solely on AI.Source: Other relevant law · p. 36

Other relevant law.

• Practitioners must follow local rules, and affirmative-action law may constrain fairness interventions that exceed what is needed to remedy unjustified discrimination.The passage identifies this as a complex issue in both the US and EU legal frameworks.Source: Other relevant law · p. 36

8.3 Operational challenges

• Whether hiring algorithms pursue legitimate aims through appropriate means requires contextual assessment and, where possible, comparison with hypothetical non-algorithmic outcomes.The passage raises uncertainty about inferring candidates’ motivation as a legitimate aim and using estimates trained on biased ground-truth variables as appropriate means.Source: §8.3 · p. 36
• Human oversight of post-processing and evaluation of its effects on groups and especially vulnerable individuals are advisable amid unsettled affirmative-action law.Direct discrimination is generally harder to justify, and the legality of post-processing choices such as targeting TPR parity rather than DI remains uncertain; US and EU affirmative-action law is in flux.Source: §8.3 · p. 36

9.1 Algorithmic Uptake

• During COVID-19, delivery companies adopted algorithmic recruitment to avoid contagion risks and potentially reduce HR costs, amid laxer emergency regulation.The authors argue that non-technological or reduced-technological options should also be considered.Source: §9.1 · p. 37
• Hiring technology can reduce bias relative to human decision-making, yet often reinforces bias, while institutional investment encourages companies to retain it beyond the emergency.The authors call for considering the balance of ethical harms if non-technological procedures are reintroduced.Source: §9.1 · p. 37

9.2 Algorithmic Fairness

• A pandemic shock can sharply reduce JS’s hiring prospects despite fair technical solutions, as demand falls, competition rises, and intersecting disadvantages compound.Gendered childcare duties, fewer connections for referrals [284], and reduced social-safety-net support linked to migration background intensify the effects.Source: §9.2 · p. 37
• Fairness measures confined to one system and moment can miss a rapidly changing socio-technical context, while shocks undermine data representativeness and make changes hard to quantify.Pre-deployment evaluations can become obsolete; fresh data is needed, but changing incentives, sensitive-data complexity, and delayed hiring feedback complicate evaluation.Source: §9.2 · pp. 37–38

10.1 Bias and Validity

• Algorithms can mitigate some biases but may encode recruiters’ and societal biases, while hiring targets such as job performance remain difficult to define or predict.The authors recommend evaluating applicant representation and outcomes through and after the hiring pipeline, studying individual fairness and exploratory policies, and scrutinizing AVIs and personality prediction through validity theory. They also note that automation may reduce avoidance by vulnerable candidates, while some systems predict scores assigned by human reviewers; job performance is difficult to define and predict.Source: §10.1 · pp. 38–39

10.2 Broader Context

• Fair hiring requires contextualized evaluation of the whole hiring process and recruiter-machine interaction, not just fairness scores for individual algorithms.Human recruiters can complement algorithms by handling special cases and changing conditions, while stage-by-stage fairness does not guarantee fair outcomes across the pipeline. The authors call for studying impacts on job seekers and utility for different employers, including small businesses, and caution that job-performance or tenure models could enable termination decisions, raising ethical and social concerns. They also call for research beyond human-versus-algorithm comparisons, including screening models [6].Source: §10.2 · p. 39

10.3 Data

• Fairness research can inform organizational practice, but limited data coverage and privacy and intellectual-property constraints impede evaluation for vulnerable groups.The authors recommend more geographically and linguistically diverse data and better coverage of relevant sensitive attributes, including disability status and sexual orientation, supported by applicant surveys and data-donation campaigns. They also call for innovative, IP-friendly auditing protocols.Source: §10.3 · p. 40

10.4 Law

• Clearer regulation could shape fairer hiring, but legal constraints on sensitive-data processing complicate bias assessment and fairness interventions.The authors call for multidisciplinary work to reconcile privacy and non-discrimination trade-offs across jurisdictions; they identify multiparty computation, sample-level estimators, and noise injection as promising approaches.Source: §10.4 · p. 40

• Algorithmic fairness is now necessary for business-as-usual product development and frontier research as the social, technological, and legal landscape evolves.Source: §11 · p. 40
• Practitioners and researchers should use contextualized measures on appropriate data to understand bias factors and deploy suitable mitigation strategies.Source: §11 · p. 40
• A balanced understanding can help avoid legitimizing questionable applications through misguided analyses and support shared benefits for society.Multidisciplinary work with legal scholarship is critical to guide policy and define technically achievable desiderata.Source: §11 · pp. 40–41

A SYSTEMATIC REVIEW METHODOLOGY

• The survey combines a CS perspective with Human Resource Management, Industrial and Organizational Psychology, Philosophy, and Law, while Sections 4–6 draw on a systematic literature review.The review also analyzes influential technical reports, industry white papers, and legal literature using mixed methods.Source: A SYSTEMATIC REVIEW METHODOLOGY · p. 55
• The search used IEEE Xplore, ACM Digital Library, and Google Scholar, querying article titles with expanded algorithmic, hiring, and fairness terms.The expanded query included terms targeting evaluation stages such as interviews and appraisals.Source: A SYSTEMATIC REVIEW METHODOLOGY · p. 55
• The reviewers screened the top 100 results manually, including only quantitative fairness audits or novel methods addressing fairness in algorithmic hiring.They excluded related work focused on freelancing, human perceptions, qualitative aspects, or accuracy without fairness consideration.Source: A SYSTEMATIC REVIEW METHODOLOGY · p. 55
• The review supplemented its search with forward and backward snowballing and related surveys or dedicated venues, finding one additional dataset and no additional article.The additional dataset was IBM HR Analytics; related surveys considered were [173].Source: A SYSTEMATIC REVIEW METHODOLOGY · p. 55