AI bias in hiring: the hidden risk companies miss 

We know that 72% of HR leaders believe AI can reproduce or amplify human bias. Yet most of them are still increasing their use of AI in hiring. That isn’t necessarily a contradiction, but it is a misunderstanding of how AI bias actually operates.

The assumption running beneath most AI adoption decisions is that a tool designed to be objective will be objective. But good intentions don’t count for much when platforms scale preexisting biases and develop biases of their own. This is the AI maturity gap in practice . adoption is outpacing the governance needed to use these tools fairly.

AI bias in hiring occurs when the data, model design, or operational use of an AI tool produces outcomes that systematically disadvantage candidates based on characteristics unrelated to job performance. It’s distinct from human bias in one critical way: where human bias is inconsistent and visible, algorithmic bias is stable, scalable, and often invisible until it has already shaped thousands of decisions. 

Why AI can both reduce and increase bias in recruitment 

AI can remove certain kinds of bias from hiring. It applies the same criteria to every candidate, scores responses using the same algorithm, and eliminates the variability that comes with tired interviewers and split-second judgements. For organisations where inconsistent human decision-making has been the primary driver of unfair outcomes, that consistency is an improvement. 

But consistency and fairness aren’t the same thing. An algorithm trained on biased historical data will apply those patterns consistently – to every candidate, in every role, across every hiring decision the organisation makes. It doesn’t correct for the bias in its training data. It learns it, and then it scales it. 

Our recent survey found that HR leaders believe AI can both amplify bias and reduce it. And they’re right — it depends entirely on what the tool was built on and how it’s governed. For a broader overview of both the opportunities and risks AI brings to HR, see The promise and pitfalls of AI integration in HR. 

How bias enters AI hiring systems 

Bias doesn’t appear in AI systems fully formed. It enters at three specific points: 

• In the data. AI learns from historical information – past hiring decisions, past performance ratings, past promotion patterns. If that history reflects structural bias in who was hired, developed, and retained, the model learns those patterns as signals of success. A tool trained on ten years of hiring data from an organisation that promoted men disproportionately will learn that the characteristics of those men are predictors of performance. They aren’t. They’re proxies for the decisions of the people who promoted them.

• In the model. Even with clean training data, models can identify proxy variables – factors that correlate with protected characteristics without naming them directly. Postcodes may correlate with ethnicity. University attended can correlate with socioeconomic background. Vocabulary and sentence structure often correlate with both. A model optimising for predictive accuracy will use these signals if they improve its outputs, regardless of whether they’re job-relevant. 

• In the interaction. Prompts and criteria embed assumptions about what good looks like. A recruiter who asks an AI tool to find candidates similar to the company’s top performers is prioritising whatever characteristics those performers share – including the ones that have nothing to do with the job. The instruction sounds neutral but the outcome often isn’t. And because the bias enters through the prompt rather than the model, it can be invisible even to the people operating the tool. 

Why algorithmic bias is harder to detect than human bias

Human bias is inconsistent. A biased interviewer has good days and bad days, makes different calls in different moods, and produces a pattern that’s messy enough to be challenged or dismissed. Algorithmic bias is stable. The same inputs will produce the same outputs, every time, across every candidate the system processes. 

That stability is what makes it dangerous. A biased algorithm doesn’t reveal itself through occasional bad decisions – it reveals itself through aggregate outcomes that only become visible at scale. By the time the pattern shows up in hiring data, it has already shaped hundreds or thousands of individual decisions. Each one looked reasonable in isolation but none of them were random. 

Active monitoring is the only reliable defence. This is why Assessio uses demographic data to identify and correct adverse impact across our assessments. After all, waiting for a pattern to become visible in outcomes means waiting until the damage is already done. 

Real examples of AI bias in hiring: HireVue and LLM resume screening

In 2020, HireVue’s AI-powered video interview system came under sustained criticism for analysing candidates’ facial expressions and body language as proxies for job-relevant traits.  

In one documented case, a furloughed employee was rejected when she reapplied for her own job – scoring well on her responses but poorly on her body language. The tool had been consistently making these decisions at scale, before anyone noticed. HireVue removed the facial analysis feature in 2021, acknowledging the science was flawed. 

A 2025 study by researchers at the University of Hong Kong and the Chinese Academy of Sciences tested five leading large language models on resume screening. All five showed systematic bias – scoring female candidates higher than male candidates, and consistently scoring Black male candidates lower than white male candidates with identical qualifications.  

The researchers noted these biases appeared deeply embedded in how the models evaluate candidates – and that some had likely been introduced or reinforced during debiasing attempts. Tools explicitly designed to reduce bias had encoded new patterns in the process. 

How structured AI reduces bias in recruitment 

The solution to AI bias is not to abandon AI entirely — it’s to implement AI processes built on real psychometric science. Structured AI doesn’t make assumptions about job fit, but bases its decisions on validated data that has been tested to confirm it really does predict performance. 

Three standards define what that looks like in practice: 

• Predictive validity – the tool’s outputs correlate with how candidates actually perform in the role, not with proxy variables that happen to separate high scorers from low ones. 
• Reliability – the same candidate would receive a consistent result over time, reflecting stable job-relevant characteristics rather than transient circumstances. 
• Relevance – the assessment measures only what matters for the role, tested for adverse impact across the demographic groups in the candidate population. 

A tool that meets all three can be trusted to measure what it claims to measure – consistently and without systematically disadvantaging groups of candidates in the process. 

Five ways to prevent AI bias in hiring 

Use validated data. Training data should be tested for representativeness across gender, age, ethnicity, and other protected characteristics before it’s used to build or calibrate a model. Historical hiring data is almost never a safe starting point without intervention. 

1. Monitor bias continuously. Pre-launch testing tells you whether a tool was fair when it was deployed. It doesn’t tell you whether it remains fair as the candidate pool shifts, the role evolves, or the model updates. Adverse impact analysis should be an ongoing practice, not a one-time sign-off. 

2. Remove demographic inputs. Any data the model doesn’t need for predicting job performance is data that creates bias risk. Names, postcodes, educational institutions, and photos should be removed from inputs unless there is a specific, job-relevant reason for their inclusion. Even then, the outputs should be tested for disparate impact. 

3. Add human oversight. AI should surface data and recommendations. Humans should make the final decisions, particularly at consequential points in the hiring process. An organisation where AI can reject candidates without any human review has removed the check that might catch systematic error before it scales. 

4. Ensure explainability. If you can’t explain to a candidate what the tool assessed, why those factors were chosen, and how their result was reached, the tool isn’t ready for responsible use. Explainability isn’t just a regulatory requirement under the EU AI Act, it’s the standard that forces vendors to justify their methodology rather than hide behind proprietary algorithms. 

For tools built on validated psychometric science, these aren’t new requirements – they’re the starting point. The question isn’t whether your AI hiring tools can meet this standard. It’s whether they have the structure and governance to make sure they continue to meet these standards indefinitely. 

👉 This article draws on research from The Maturity Gap, Assessio’s data-driven guide to AI adoption, governance, and trust in HR. Download the full report to explore the findings in depth. Download The Maturity Gap here.