Thesis – Algorithmic Bias in AI-Assisted Recruitment: An Australian Employer Perspective

Abstract

Artificial intelligence (AI) screening tools are increasingly embedded in Australian recruitment, yet employer understanding of the discrimination risks they carry remains underexplored. This thesis investigates how Australian employers perceive, govern and monitor algorithmic bias in AI-assisted recruitment. Semi-structured interviews were conducted with 14 human resources (HR) and talent acquisition leads from organisations spanning banking, retail, government, resources, health and other sectors, and were analysed using reflexive thematic analysis. Four themes were developed: black-box procurement, over-trust in algorithmic rankings, absent bias auditing, and consistency upsides. Participants procured tools with minimal visibility of training data or validation evidence, rarely overrode ranked shortlists and conducted no scheduled adverse impact testing, despite obligations under Commonwealth anti-discrimination legislation and the Fair Work Act 2009 (Cth). The thesis argues that accountability for algorithmic fairness is presently displaced onto vendors, and proposes procurement, human oversight and monitoring practices aligned with Australian Human Rights Commission guidance.

Introduction

Automated screening has shifted from novelty to default practice in high-volume Australian recruitment. Large employers can receive several thousand applications for a single graduate program, and AI-assisted tools promise to compress screening through CV parsing and ranking, chatbot pre-screening, game-based psychometrics and scored asynchronous video interviews (Tambe et al., 2019). Vendors position these products as faster, cheaper and fairer than manual shortlisting, and adoption has accelerated across the private and public sectors (Hunkenschroer & Luetge, 2022).

The efficiency case sits alongside a well-documented discrimination risk. Models trained on historical hiring records can learn and reproduce the preferences embedded in those records, including preferences that disadvantage protected groups (Barocas & Selbst, 2016). The Australian legal exposure is direct. The Racial Discrimination Act 1975 (Cth), the Sex Discrimination Act 1984 (Cth), the Disability Discrimination Act 1992 (Cth) and the Age Discrimination Act 2004 (Cth) prohibit discrimination in deciding who is offered employment, and the general protections in the Fair Work Act 2009 (Cth) prohibit adverse action on protected grounds. None of these statutes excuses an employer because the contested decision was generated by third-party software. The Australian Human Rights Commission (AHRC, 2021) has accordingly warned that automated systems can produce unlawful outcomes at scale, recommending transparency, meaningful human oversight and regular auditing wherever AI informs significant decisions.

Despite this exposure, little empirical research has examined how Australian employers actually understand and govern these tools. This thesis addresses that gap through three research questions:

1. How do HR and talent acquisition leaders in Australian organisations understand algorithmic bias risk in AI-assisted recruitment?
2. What governance, oversight and monitoring practices surround these tools in practice?
3. What conditions would support fairer and more accountable adoption?

This extract presents the condensed literature review, methodology and principal findings, followed by implications for Australian practice.

Literature Review

Sources of algorithmic bias

Bias enters recruitment algorithms through several channels. Training data drawn from past hiring decisions encodes historical patterns of who was hired, so a model optimised to predict successful candidates may simply predict resemblance to previous hires (Barocas & Selbst, 2016). Facially neutral features such as postcode, career gaps or extracurricular vocabulary can operate as proxies for gender, age, disability or cultural background, and measurement bias arises where scoring instruments, including speech analysis in video interviewing, perform unevenly across groups. Köchling and Wehner’s (2020) systematic review documents discrimination risk at every automated stage of the funnel, including the widely reported CV model that downgraded applications signalling female gender.

Audit evidence

Independent audits of deployed systems remain scarce. Bogen and Rieke (2018) mapped predictive tools across the hiring funnel and concluded that equity risks concentrate in sourcing and screening, precisely the stages where volume pressure makes automation attractive. Raghavan et al. (2020) examined the public claims of algorithmic hiring vendors and found validation evidence rarely disclosed, bias mitigation claims largely unverifiable and methodological detail treated as proprietary. Purchasers therefore typically cannot evaluate the fairness properties of what they are buying.

The fairness metrics tension

Even willing adopters confront a structural difficulty: fairness has multiple formal definitions that cannot generally be satisfied at once. Where base rates differ between groups, a scoring system cannot be simultaneously calibrated and equal in error rates across groups (Chouldechova, 2017; Kleinberg et al., 2017). Selecting a fairness metric is therefore a normative choice rather than a technical inevitability, yet it is usually made silently by vendors. Ajunwa (2020) describes the resulting paradox of automation as anti-bias intervention, in which tools marketed as removing human prejudice can entrench discrimination while insulating it from scrutiny.

Australian regulatory guidance

Australian guidance has developed faster than Australian evidence. The AHRC (2021) recommends that organisations using AI in significant decisions conduct impact assessments, ensure decisions are explainable, preserve meaningful human review and audit outcomes for bias. Australia’s voluntary AI Ethics Framework similarly emphasises fairness, transparency, contestability and accountability (Department of Industry, Innovation and Science, 2019). Missing is evidence on whether employer practice reflects this guidance, the gap this thesis addresses.

Methodology

Research design

The study adopted a qualitative, interpretivist design. Because the research questions concern how practitioners perceive risk and enact governance, rather than the statistical behaviour of particular tools, semi-structured interviews were used, covering procurement, oversight and monitoring consistently while leaving room to probe unexpected practices.

Participants and data collection

Purposive sampling, supplemented by snowball referral, recruited 14 senior HR and talent acquisition leaders. Eligibility required direct responsibility for selecting or operating at least one AI-assisted screening tool within an Australian organisation. Interviews of 45-70 minutes were conducted by video conference between August and November 2025, recorded with consent, transcribed verbatim and de-identified as P1 to P14. A participant profile is presented in Table 1.

Ethics

The study received approval from the University Human Research Ethics Committee (approval 2025/318) and was conducted in accordance with the National Statement on Ethical Conduct in Human Research (National Health and Medical Research Council, 2018). Participants received a plain language statement, gave written consent and could withdraw before publication. Transcripts were stored on encrypted university systems, and organisation and vendor names were redacted.

Data analysis

Transcripts were analysed using reflexive thematic analysis (Braun & Clarke, 2006, 2021). Figure 1 illustrates the process: repeated reading for familiarisation, systematic coding in NVivo, generation of candidate themes, review against coded extracts and the full dataset, and definition and reporting of final themes. A reflexivity journal recorded analytic decisions, and themes were treated as patterns of shared meaning rather than topic summaries.

Findings

As Table 1 shows, participants spanned a broad cross-section of Australian employers. CV parsing and ranking was the most common tool category (11 of 14 organisations), followed by chatbot pre-screening (five), scored video interviewing (three) and game-based psychometrics (two).

Table 1: Participant and organisation profile (n = 14)

Analysis developed four themes, summarised in Table 2 and elaborated below.

Table 2: Themes developed from the analysis

Theme 1: Black-box procurement

Twelve participants described purchasing screening tools without access to training data provenance, feature lists or validation results. Procurement decisions turned on demonstrations, reference sites and integration with the existing applicant tracking system, while fairness claims were accepted on vendor assurance. P4 (banking) recalled: “We asked what the model was trained on and we got a marketing deck. There was no technical answer in the room.” Several participants could not state which applicant attributes their tool actually used, and contracts rarely included audit rights or notification of model changes.

Theme 2: The ranking feels objective

Nine participants described strong deference to algorithmic rankings. Recruiter attention concentrated on candidates above the cut-off, and overriding the ranking required written justification while accepting it required none, a design that inverts the intended direction of human oversight. P9 (technology) was candid: “If the system scores someone below sixty, honestly, nobody looks at them.” Participants often framed the ranking as neutral because the algorithm “does not see gender”, a belief at odds with the proxy mechanisms documented in the literature.

Theme 3: Nobody is auditing

Eleven participants confirmed that no scheduled bias or adverse impact testing occurred in their organisation. Recruitment analytics tracked time-to-fill, cost-per-hire and offer acceptance, but not selection rates by gender, age or cultural background, and most assumed auditing was the vendor’s responsibility. Awareness was nonetheless rising. P7 (state government) reflected: “After the Robodebt Royal Commission, our board asked whether any of our systems could do that to people. For recruitment, we did not have an answer.”

Theme 4: A more consistent first gate

Eight participants argued that the tools improved on the manual status quo. Uniform criteria were applied identically to every application, removing variation from tired or time-poor screeners, and structured scoring displaced informal shortlisting based on gut feel or personal networks. Several noted that configuring the tool forced hiring managers to articulate selection criteria before advertising, a discipline they regarded as a fairness gain in itself.

Discussion

The findings depict a governance vacuum rather than indifference. Legal responsibility for discriminatory outcomes rests with the employing organisation under Commonwealth legislation, yet the practical capacity to inspect, test and correct the screening instrument sits with vendors who disclose little. This displacement of accountability mirrors the audit findings of Raghavan et al. (2020) and gives the black-box procurement theme its significance: employers are carrying legal risk they cannot delegate on the strength of marketing material.

The over-trust theme extends automation bias into recruitment practice. Human review is the central safeguard in both AHRC (2021) guidance and vendor rhetoric, but a loop in which disagreeing with the machine requires justification while agreeing requires none is oversight in name only. The interface design of ranked shortlists actively channels attention away from the candidates most likely to have been wrongly excluded.

The absence of auditing cannot be excused by the fairness metrics tension. No system can satisfy every fairness definition simultaneously (Chouldechova, 2017; Kleinberg et al., 2017), but impossibility results constrain perfection, not measurement. Routine adverse impact analysis using applicant-flow data is feasible with information employers already hold and would surface precisely the patterns anti-discrimination law makes costly.

Finally, the consistency theme deserves to be taken seriously. Structured, criterion-based screening can genuinely outperform idiosyncratic manual shortlisting (Köchling & Wehner, 2020). The appropriate response is therefore not prohibition but conditional adoption: retain the standardisation benefits while building the verification practices participants currently lack, mindful of Ajunwa’s (2020) warning that unexamined automation converts bias into infrastructure.

Implications for Practice and Policy

Three implications follow for Australian employers. First, procurement should become the primary control point. Before contracting, organisations should require answers to questions such as:

• What data was the model trained on, and does validation evidence cover Australian applicant populations?
• What subgroup performance testing has been conducted across gender, age, disability and cultural background, and can results be disclosed?
• Which fairness definition does the tool optimise, and who made that choice?
• Does the contract provide audit rights, notification of model changes and support for candidate explanation requests?

Second, human review should be redesigned so that it can actually catch errors. Blind re-screening of sampled candidates below the algorithmic cut-off, rotating requirements to justify acceptance as well as rejection, and recruiter training on automation bias would make oversight meaningful, consistent with AHRC (2021) recommendations.

Third, monitoring should be routine. Quarterly adverse impact analysis of selection rates by protected attribute, reported to the executive alongside time-to-fill, would align recruitment analytics with the positive duty imposed on employers under the Sex Discrimination Act 1984 (Cth) following the Respect at Work reforms, and with the merit principle governing Australian Public Service selection. For public sector adopters especially, the Robodebt experience demonstrates the institutional cost of automated systems that no one is empowered to question.

Conclusion

This thesis examined how Australian employers govern algorithmic bias in AI-assisted recruitment through interviews with 14 HR and talent acquisition leaders. The picture that emerged combines dependence with blindness: organisations rely on screening algorithms they cannot inspect, defer to rankings they rarely challenge, and measure everything about recruitment except its fairness. At the same time, participants identified genuine consistency benefits that a blanket rejection of these tools would forfeit. The contribution is a grounded account of an employer perspective largely missing from the Australian debate. The study is limited by its sample of 14 predominantly large organisations and its reliance on self-report; applicant experiences and technical audits of tools deployed in Australia remain priorities for future research. Until such evidence accumulates, the prudent employer position is conditional adoption: contract for transparency, design oversight that can disagree with the machine, and measure outcomes as if the law applied to algorithms, because it does.

References

Ajunwa, I. (2020). The paradox of automation as anti-bias intervention. Cardozo Law Review, 41(5), 1671-1742.

Australian Human Rights Commission. (2021). Human rights and technology: Final report.

Barocas, S., & Selbst, A. D. (2016). Big data’s disparate impact. California Law Review, 104(3), 671-732.

Bogen, M., & Rieke, A. (2018). Help wanted: An examination of hiring algorithms, equity, and bias. Upturn.

Braun, V., & Clarke, V. (2006). Using thematic analysis in psychology. Qualitative Research in Psychology, 3(2), 77-101.

Braun, V., & Clarke, V. (2021). Thematic analysis: A practical guide. Sage.

Chouldechova, A. (2017). Fair prediction with disparate impact: A study of bias in recidivism prediction instruments. Big Data, 5(2), 153-163.

Department of Industry, Innovation and Science. (2019). Australia’s artificial intelligence ethics framework. Australian Government.

Hunkenschroer, A. L., & Luetge, C. (2022). Ethics of AI-enabled recruiting and selection: A review and research agenda. Journal of Business Ethics, 178(4), 977-1007.

Kleinberg, J., Mullainathan, S., & Raghavan, M. (2017). Inherent trade-offs in the fair determination of risk scores. Proceedings of the 8th Conference on Innovations in Theoretical Computer Science, 1-23.

Köchling, A., & Wehner, M. C. (2020). Discriminated by an algorithm: A systematic review of discrimination and fairness by algorithmic decision-making in the context of HR recruitment and HR development. Business Research, 13(3), 795-848.

National Health and Medical Research Council. (2018). National statement on ethical conduct in human research. Australian Government.

Raghavan, M., Barocas, S., Kleinberg, J., & Levy, K. (2020). Mitigating bias in algorithmic hiring: Evaluating claims and practices. Proceedings of the 2020 Conference on Fairness, Accountability, and Transparency, 469-481.

Tambe, P., Cappelli, P., & Yakubovich, V. (2019). Artificial intelligence in human resources management: Challenges and a path forward. California Management Review, 61(4), 15-42.