AI in Recruitment: How to Use It Without Building In Bias
A practitioner's guide to using AI in hiring responsibly, covering fair data, adverse-impact testing, transparency with candidates and meaningful human review.

AI in recruitment can speed up sourcing, screening and scheduling without increasing bias, but only if fairness is designed in. That means choosing low-risk use cases, keeping protected attributes and their proxies out of models, testing outcomes for adverse impact across groups, telling candidates how AI is used and keeping trained humans accountable for decisions.
Hiring teams are under real pressure: high application volumes, thin recruiter capacity and hiring managers who want shortlists yesterday. AI helps with all three. It can also quietly replicate every historical bias in your past hiring data, at scale and with a veneer of objectivity. This guide covers how to get the benefits while keeping the risk under control.
Where AI in recruitment helps, and where it is risky
Not every use of AI in hiring carries the same risk. The closer a system gets to deciding who is rejected, the higher the stakes.
| Use case | Typical value | Bias risk | Recommended control |
|---|---|---|---|
| Writing and de-biasing job descriptions | Clearer, more inclusive ads | Low | Human edit and approval |
| Interview scheduling and candidate FAQs | Faster response, less admin | Low | Clear escalation to a person |
| Sourcing and search across profiles | Wider talent pools | Medium | Audit who is surfaced by group |
| CV parsing and structured extraction | Consistent data for recruiters | Medium | Accuracy checks on diverse CV formats |
| Ranking or scoring candidates | Faster shortlisting | High | Adverse-impact testing, human review of every rejection |
| Video or voice analysis of candidates | Contested | Very high | Avoid unless validation is exceptionally strong |
| Automated rejection without review | Speed | Very high | Do not do this |
A sensible approach is to start at the top of this table and move down only when your governance is mature enough to support it.
How bias gets into hiring models
Bias rarely enters because someone intended it. It usually comes from one of four places.
Historical training data
If a model learns from past hiring decisions, it learns past preferences. If a company historically hired mostly from a handful of colleges, a model trained on those outcomes will favour those colleges, and whatever demographic skew comes with them.
Proxy variables
Removing gender, age, religion or caste from the data is necessary but not sufficient. Postcodes, college names, gaps in employment, graduation year, extracurricular activities and even writing style can act as proxies. A model can reconstruct a protected attribute from them without ever seeing it.
Label choice
What does "good hire" mean in your data? If the label is "was promoted within two years", and promotions themselves were biased, the model inherits that bias. Choosing an outcome to predict is an ethical decision as much as a technical one.
Generative models
Large language models used to summarise CVs or score answers can carry biases from their training data. They can also respond differently to names, dialects or phrasing. Prompting a model to "be fair" does not make it fair; only testing tells you, which is why structured LLM evaluation matters here.
A step-by-step process for fair AI in hiring
This is the process we recommend for any system that influences who progresses in a hiring funnel.
- Define the decision and its owner. Write down exactly what the AI does, which stage it affects and which named person is accountable for outcomes.
- Base criteria on the job, not the past. Build scoring around validated, job-related requirements from a structured job analysis, rather than "people like our previous hires".
- Audit and prepare the data. Remove protected attributes and review likely proxies. Check that the data represents the candidate population you actually want to reach.
- Choose explainable approaches where possible. A simpler model whose factors a recruiter can understand is often better than a marginally more accurate black box.
- Test for adverse impact before launch. Compare selection rates across groups (see below) on historical or pilot data.
- Pilot in shadow mode. Run the model alongside human decisions without acting on it, and compare outcomes and disagreements.
- Design meaningful human review. Recruiters must see the reasons behind a score, be able to override it and be trained to do so.
- Tell candidates. Explain where AI is used, what it assesses and how to request an alternative process or human review.
- Monitor continuously. Re-run fairness tests on live data at a fixed cadence and whenever the model, prompt or job family changes.
- Document everything. Keep records of data sources, tests, results, decisions and changes. You will need them for audits and for your own learning.
Adverse-impact testing in practice
Adverse impact means a selection process disadvantages members of a particular group, whether or not anyone intended it. Testing for it is the core technical control.
The four-fifths rule of thumb
A long-established heuristic in US employment practice compares selection rates between groups. If one group's selection rate is less than four-fifths (80 per cent) of the rate for the highest-selected group, that is treated as a signal of possible adverse impact worth investigating. For example, if 40 per cent of one group passes a screening stage and only 28 per cent of another does, the ratio is 0.7, which falls below the threshold.
It is a screening signal, not a verdict. Small sample sizes produce noisy ratios, so pair it with statistical significance tests and look at trends over time.
What to test
- Each stage separately: sourcing, screening, assessment and interview, not just final offers.
- Intersectional groups where data allows, since bias can hide in combinations such as older women or candidates from particular regions.
- Score distributions, not only pass or fail outcomes, to see where thresholds create disparities.
- Model versions and prompts, whenever anything changes.
Where you lack demographic data
Many organisations, particularly in India, do not collect demographic data from candidates, and should think carefully before they do. Options include voluntary, clearly consented self-identification stored separately from hiring decisions, and testing on carefully constructed synthetic or matched CV sets where only one attribute varies. Take privacy advice before collecting sensitive data, and handle personal data in line with the DPDP Act 2023 principles of purpose limitation, consent and minimisation. Our guide to DPDP Act compliance for AI covers this further.
Transparency and human review
Candidates increasingly expect to know when software is assessing them. Good practice includes a plain-language notice at the point of application, a short explanation of what the AI evaluates and a named route to request human review.
Human review has to be real. A recruiter clicking "approve" on a list of AI rejections without seeing the reasoning is automation with extra steps. Effective review means:
- Recruiters see the key factors behind each recommendation.
- Rejections from high-risk stages get sampled and reviewed regularly.
- Override rates are tracked. Near-zero overrides can mean the team is rubber-stamping; very high overrides can mean the model is not useful.
- Recruiters are trained on the model's limitations and on common human biases too.
The regulatory picture at a high level
Regulators increasingly treat AI in hiring as high risk. In New York City, Local Law 144 requires employers using automated employment decision tools to commission an independent bias audit, publish a summary of results and notify candidates. The EU AI Act classifies AI systems used for recruitment and selection as high risk, bringing obligations around risk management, data governance, human oversight, transparency and record-keeping. In India, the DPDP Act 2023 governs how candidate personal data is collected and processed, and existing anti-discrimination principles still apply to hiring decisions regardless of the tools used.
This is general information, not legal advice. If you hire across jurisdictions, take advice from employment and data protection counsel before deploying AI screening.
A pre-launch checklist
- A named owner is accountable for the system's outcomes.
- Scoring criteria trace back to a documented job analysis.
- Protected attributes are excluded and proxy variables have been reviewed.
- Adverse-impact testing has been done at every affected stage, with results recorded.
- The system ran in shadow mode and disagreements with humans were analysed.
- Candidates receive a clear notice and a route to human review.
- Recruiters can see reasons, override decisions and have been trained.
- Monitoring, re-testing cadence and a rollback plan are in place.
- Vendor contracts give you access to audit results and the data needed for your own testing.
Frequently asked questions
How is AI used in recruitment?
Common uses include writing job descriptions, sourcing candidates, parsing CVs, screening and ranking applicants, scheduling interviews and answering candidate questions through chat. The lower-risk uses support recruiters with admin and consistency. The higher-risk uses influence who progresses and need strong fairness controls.
Can AI remove bias from hiring?
AI can reduce some forms of inconsistency, for example by applying the same structured criteria to every candidate. It cannot remove bias on its own, and it can amplify historical bias if trained on past decisions. Fairness comes from job-related criteria, testing and human accountability, not from the technology alone.
What is adverse-impact testing?
It compares selection rates across groups at each stage of a hiring process to see whether any group is disproportionately screened out. The four-fifths rule is a common first check, supported by statistical tests. Any flagged disparity should be investigated and either justified by job-related criteria or fixed.
Do candidates need to be told that AI is screening them?
In some jurisdictions, such as New York City and under the EU AI Act, notification or transparency obligations apply. Even where not required, telling candidates is good practice and builds trust. Offer a clear route to request human review.
Should we buy an AI hiring tool or build our own?
Most companies should buy for common tasks such as scheduling and sourcing, and be selective about ranking tools. Whichever route you take, insist on access to bias audit results and the ability to run your own adverse-impact tests. Responsibility for hiring decisions stays with you, not the vendor.
How Sunday Labs can help
Sunday Labs helps HR and talent teams design, evaluate and govern AI in hiring, from choosing sensible use cases to building adverse-impact testing and monitoring into the workflow. Our work is delivered by senior engineers who treat fairness and explainability as engineering requirements, not afterthoughts. We can also review vendor tools before you commit. If you are planning to bring AI into recruitment, start a conversation with us.




