AI in HR and Hiring: Building Fair Recruitment Systems and Preparing for the Future of Work

Recruitment was one of the first business functions to adopt AI at scale, for an understandable reason: volume. A single job posting can attract a flood of applications, and recruiters cannot read them all with equal care. Automated screening, chatbots that answer candidate questions, interview scheduling and skills assessments promise to make the process faster and more consistent.

But hiring is also one of the most consequential places to deploy AI. A decision about whether someone gets an interview affects their income, career and dignity. That is why regulators treat it seriously. The EU AI Act lists AI systems used for recruitment and worker management among its high-risk categories, and New York City's law on automated employment decision tools requires bias audits and candidate notices for certain uses.

In Nigeria and across Africa, the context adds another layer. Youth unemployment is a pressing concern, and many talented people come through non-traditional routes: self-taught developers, bootcamp graduates, freelancers. Through BuildWithMMO, I mentor many developers whose skills are real but whose CVs do not fit conventional templates. A screening system that rewards pedigree over ability would filter out exactly the people the ecosystem most needs to bring in.

Most systems fall into a few technical categories:

• CV parsing and matching: NLP models extract skills, job titles, education and experience from documents, then score similarity to a job description. Modern systems use embeddings rather than keyword lists, which helps match "built REST APIs in Django" to "backend web development".
• Predictive ranking: some tools train models on historical hiring outcomes to predict which candidates will be hired or perform well. This is the riskiest approach, because historical decisions encode historical biases.
• Assessments: coding challenges, situational judgement tests and work samples, sometimes scored automatically.
• Conversational agents: chatbots that answer questions, collect information and schedule interviews.
• Video and voice analysis: tools that claim to infer traits from facial expressions or tone. These are scientifically contested and raise serious fairness and disability-discrimination concerns. I advise teams to avoid them.
• Generative AI for recruiters: drafting job descriptions, summarising interview notes and generating structured interview questions.

Large language models have also blurred the line on the candidate side. Applicants now use AI to tailor CVs and cover letters, which means keyword-matching screens increasingly compare machine-written documents against machine-written job descriptions. That dynamic weakens the signal in written applications and is another reason to shift weight towards structured, job-relevant assessments that measure what a person can actually do.

The core risk is simple to state: a model trained on past decisions learns to reproduce them. If a company historically hired mostly from a few universities, a model will learn that those universities predict success, even if they do not cause it. Proxies are everywhere. Names can correlate with ethnicity or religion, addresses with income, employment gaps with caregiving responsibilities and certain extracurricular activities with gender.

Removing a sensitive attribute from the input rarely fixes this, because other features carry the same signal. Effective mitigation requires measurement:

• Selection rate comparisons: compare the rate at which different groups pass each stage. Large disparities warrant investigation.
• Error analysis by group: check whether qualified candidates from certain groups are disproportionately rejected.
• Counterfactual testing: change a single attribute, such as a name, on an otherwise identical CV and see whether the score changes.
• Feature review: inspect which features drive scores and remove those that are proxies without job relevance.

A fair hiring model is not one that ignores demographics. It is one whose behaviour across groups has been measured, understood and corrected.

Measuring bias requires some demographic data, which creates its own privacy obligations. Under the Nigeria Data Protection Act and similar laws, collect such data only with a lawful basis, keep it separate from decision systems and restrict access.

Good design principles for AI in recruitment include:

1. Skills over signals. Score demonstrated abilities, such as work samples and structured assessments, rather than proxies like school names. Skills-based hiring tends to be both fairer and more predictive.
2. Assist, do not decide. Use AI to organise and summarise, while humans make shortlisting and hiring decisions. Ensure reviewers can see why a candidate was ranked as they were.
3. Explainability for candidates. Tell applicants when AI is used, what it evaluates and how to request human review.
4. Accessibility. Offer alternatives for candidates with disabilities, and test that assessments do not disadvantage people using assistive technology.
5. Low-bandwidth friendliness. In markets where many candidates apply from mobile phones on expensive data, heavy video assessments are an access barrier. Offer lightweight options.
6. Audit regularly. Run bias audits before launch and on a schedule, and whenever the model or job mix changes.

Beyond hiring, AI is reshaping work itself. The International Labour Organization has published research on generative AI and jobs suggesting that the most common effect is likely to be augmentation of tasks rather than full automation of occupations, with clerical work among the most exposed. The World Economic Forum's Future of Jobs Report 2025 similarly highlights AI and big data among the fastest-growing skill areas, alongside a continued premium on analytical thinking and resilience.

A few implications for workers and organisations in emerging markets:

• Task redesign beats job replacement: the practical question for most roles is which tasks can be assisted, and how to redeploy the time saved.
• Remote work cuts both ways: AI tools make it easier for African professionals to compete for global remote roles, but they also intensify competition and can commoditise routine work.
• Continuous learning is now infrastructure: companies that invest in reskilling, and governments that support it, will capture more of the productivity gains.
• Workplace surveillance needs limits: AI-driven productivity monitoring can erode trust and wellbeing. Transparency and worker consultation should come first.

• Hiring is a high-risk AI use case under the EU AI Act and is regulated in jurisdictions such as New York City.
• Models trained on historical hiring decisions reproduce historical bias unless measured and corrected.
• Prefer skills-based assessments and keep humans responsible for decisions.
• Avoid scientifically contested tools like emotion recognition from video.
• Design for candidates on mobile and low-bandwidth connections, and disclose AI use.
• Expect AI to reshape tasks more than eliminate whole jobs; invest in reskilling.

1. The EU Artificial Intelligence Act — Future of Life Institute (EU AI Act resource site)
2. Automated Employment Decision Tools (AEDT) — NYC Department of Consumer and Worker Protection
3. International Labour Organization — ILO
4. The Future of Jobs Report 2025 — World Economic Forum
5. Nigeria Data Protection Commission — NDPC

Cover image: original illustration created for this article. Cited works remain the property of their respective authors and publishers.