Bias in AI Hiring: A Guide for Recruiters

AI bias in hiring is when AI-assisted hiring processes create systematic and unfair differences in how candidates are evaluated, ranked, or recommended.

Bias can disadvantage candidates based on protected characteristics such as race, sex, age, or disability. However, it can also arise from seemingly unrelated factors that shouldn't influence whether someone is suitable for a role.

A 2024 review of algorithmic bias in hiring identifies multiple ways bias can enter AI systems, including:

• Historical bias: Past inequalities are reflected in the data used by AI.
• Representation bias: Certain groups are underrepresented in the data.
• Measurement bias: The information used to evaluate candidates doesn't accurately represent what the system intends to measure.

AI isn't automatically objective simply because it applies the same process to everyone. Fairness also depends on the data, criteria, and methods used to evaluate candidates.

Bias doesn’t always start with the algorithm. In fact, it can occur for many reasons at different stages of an AI-assisted process.

AI systems learn patterns from data. If that data reflects previous inequalities or biased human-led decisions, then those patterns can become part of how the system evaluates future candidates.

Research into language-model risks has found that models trained on human-generated data can reproduce (and sometimes amplify) existing social biases. Instead of asking whether AI treats everyone the same, you should understand exactly what data the AI considers a sign of a good candidate.

Data isn’t the only factor that creates bias in AI hiring. The choices humans make when designing an AI hiring system can also shape its outcomes. A 2026 study of generative AI in hiring found different outcomes when HR professionals used standard AI compared with AI specifically designed around inclusion.

When AI is being designed for hiring, developers should consider:

• What counts as a strong candidate?
• Which evidence matters?
• How are candidates scored?
• What is the AI optimizing for?
• How is fairness tested?

These decisions are what can determine whether AI reinforces or counters existing biases in hiring.

Even when humans make the final decision, AI can still influence their judgment. A 2025 study of AI-assisted resume screening found that participants’ hiring decisions changed when AI recommendations favored particular racial groups, despite them not showing the same preferences without the biased recommendations.

Therefore, human oversight doesn’t automatically remove AI bias, and you should always evaluate the recommendations that AI provides closely before making a final decision.

Not all bias in AI hiring fits in traditional categories such as race, sex, or disability. Recent research into AI self-preferencing found that LLMs evaluating resumes may favor applications generated by the same model. This means some candidates may be disadvantaged depending on which AI they used to create their resume rather than being evaluated solely on job fit.

The effects of bias in AI hiring can mean more than one unfair hiring decision. Depending on where and how AI is used, potential consequences include:

• Unfair candidate outcomes: Qualified candidates may be ranked lower, rejected, or overlooked for reasons unrelated to their ability to do the job.
• Missed talent: Employers risk filtering out strong candidates and narrowing their available talent pool.
• Reduced diversity: If certain groups are consistently disadvantaged, fewer candidates from those groups may progress through the hiring process.
• Influence on recruiters: Biased AI recommendations can influence human judgment rather than simply support it.
• Legal and compliance risks: Discriminatory outcomes can expose employers to greater scrutiny and potential legal risks.

One concern with AI is that bias can be repeated at scale. A 2026 study of Algorithmic Monocultures in hiring, involving around 4 million job applications, found that 14.74% of applications from Asian candidates and 25.87% from Black candidates were submitted to positions where the screening algorithm adversely impacted their respective groups. This means Asian or Black candidates could be less likely to progress past algorithmic screening, creating unfair biases and causing employers to overlook potentially qualified talent.

Although research suggests that AI can reduce certain forms of hiring bias, this doesn’t necessarily mean simply introducing AI makes recruitment unbiased. Bias in AI hiring depends on what the system evaluates, how it is designed, and what safeguards are built around it.

One potential benefit of using AI to reduce bias in hiring is more consistency. AI can help recruiters:

• Apply predefined criteria to every candidate
• Compare candidates against the same requirements
• Use standardized evidence and scoring
• Reduce reliance on intuition or “gut feeling”

But consistency is only helpful when the criteria itself are fair and job-relevant.

AI can be deliberately designed to steer hiring decisions towards evidence that matters for the role. A 2026 study of HR professionals found that inclusion-focused generative AI helped to reduce disability-related bias, particularly when hiring decisions became more complex.

The AI system encouraged decisionmakers to focus on concrete, job-relevant competencies rather than disability-related stereotypes, showing how AI can reduce hiring bias in recruitment.

Note: This study tested a specific inclusion-focused AI intervention and disability bias. It therefore doesn’t prove that all AI hiring tools reduce discrimination.

Another approach is to limit how much an AI system can use information related to protected characteristics, such as gender.

The LEACE study tested a method for removing gender information that an AI model could easily detect from its data. Researchers then asked the model to predict people's professions.

After removing this gender information:

• The gap in prediction accuracy between genders was substantially reduced.
• The model's overall profession-prediction accuracy fell only slightly, from 79.3% to 77.3%.

In short, the model became less reliant on gender-related information while staying almost as effective at its main task.

However, the technique couldn't guarantee that every signal associated with gender had been removed. This is an important limitation: hiding or removing protected characteristics doesn't automatically make an AI system unbiased, because other information may still indirectly reveal or correlate with those characteristics.

It’s impossible to remove every form of bias in AI hiring with one single method. Instead, mitigation involves addressing potential bias at different stages, from deciding what the AI evaluates to measuring how it affects candidates.

The NIST AI Risk Management Framework supports this approach by treating AI risk management as an ongoing process throughout the AI lifecycle. In practice, how bias is mitigated in AI hiring systems involves a combination of careful design, testing, human oversight, and continuous monitoring.

AI should evaluate candidates against criteria that genuinely relate to the role, such as relevant skills, competencies, experience, or qualifications.

As discussed earlier, AI should focus on concrete, job-relevant competencies to reduce reliance on irrelevant characteristics or stereotypes.

The information an AI learns from or uses to evaluate candidates can introduce bias into its outputs.

Ask:

• Are candidate groups adequately represented?
• Does the evidence accurately measure what it is intended to measure?
• Could seemingly neutral information indirectly reveal or correlate with protected characteristics?

It can also be useful to avoid reducing candidate suitability to a single source of evidence. For example, Hire-Match combines information from resumes with assessments and structured interview evidence to build a broader picture of a candidate, rather than evaluating suitability from resume keywords alone.

However, using more evidence doesn't automatically remove bias. Each source of information and the way it is interpreted still needs to be evaluated.

People overseeing an AI system should be able to understand what evidence contributed to its scores, rankings, or recommendations.

NIST identifies explainability and interpretability as important characteristics of trustworthy AI. Greater transparency makes it easier to scrutinize an output and figure out whether a recommendation is supported by relevant evidence.

This is also an area where AI hiring tools can differ considerably. Hire-Match gives recruiters explanations of candidate strengths and concerns alongside its recommendations, giving them more information to examine than an unexplained ranking alone.

Having a human make the final decision is an important safeguard, but this doesn’t guarantee fairness on its own.

As the 2025 study showed, biased AI recommendations can influence human judgment. Therefore, recruiters should question, investigate, and override recommendations when the evidence is not supported.

Instead of choosing an AI tool based on claims that it is designed to reduce bias, you should evaluate exactly what the tool does to mitigate bias and how it does it.

A good AI hiring tool should be able to provide evidence of how it identifies and reduces specific sources of bias, as well as how it monitors for new biases that could appear over time.

Unfortunately, there is no simple yes or no answer. Research shows how algorithmic screening can produce disparities and repeat them when hiring at scale. In contrast, other studies indicate that AI can reduce certain biases by helping decisionmakers focus on job-relevant competencies.

Together, these findings show that AI can reproduce, amplify, or introduce bias, but it can also be designed to reduce specific forms of human bias.

The question isn’t whether AI is used in hiring. It's what the AI evaluates, how it is designed and tested, whether its recommendations can be scrutinized, and how potential bias is monitored over time.

AI alone won't make hiring unbiased. But when fairness is treated as an ongoing process rather than a feature that can simply be switched on, AI has the potential to support fairer hiring decisions.

• Albaroudi, E., Mansouri, T. and Alameer, A. (2024) ‘A comprehensive review of AI techniques for addressing algorithmic bias in job hiring’, AI, 5(1), pp. 383–404. 
• Belrose, N., Schneider-Joseph, D., Ravfogel, S., Cotterell, R., Raff, E. and Biderman, S. (2023) ‘LEACE: Perfect linear concept erasure in closed form’, Advances in Neural Information Processing Systems, 36.  
• Bommasani, R., Bana, S.H., Creel, K.A., Jurafsky, D. and Liang, P. (2026) ‘Algorithmic monocultures in hiring’, Proceedings of the 2026 ACM Conference on Fairness, Accountability, and Transparency, pp. 6351–6382. 
• Weidinger, L., Mellor, J., Rauh, M., Griffin, C., Uesato, J., Huang, P-S., Cheng, M., Glaese, M., Balle, B., Kasirzadeh, A., Kenton, Z., Brown, S., Hawkins, W., Stepleton, T., Biles, C., Birhane, A., Haas, J., Rimell, L., Hendricks, L.A., Isaac, W., Legassick, S., Irving, G. and Gabriel, I. (2021) ‘Ethical and social risks of harm from language models’. arXiv. 
• Wilson, K., Sim, M., Gueorguieva, A-M. and Caliskan, A. (2025) ‘No thoughts just AI: Biased LLM hiring recommendations alter human decision making and limit human autonomy’, Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society, 8(3), pp. 2692–2704. 
• Xu, J., Li, G. and Jiang, J.Y. (2025) ‘AI self-preferencing in algorithmic hiring: Empirical evidence and insights’, Proceedings of the 5th ACM Conference on Equity and Access in Algorithms, Mechanisms, and Optimization, p. 288. 
• Yang, M.M. (2026) ‘Mitigating disability bias in hiring: The role of inclusion-focused generative AI in complex HR decisions’, Human Resource Management Journal, 36(3), pp. 553–565.