AI in Hiring: Where It Helps, Where It's Biased, and the Rules to Set

A recruiter receives a ranked list of applicants. The ranking looks precise, so it becomes the shortlist. Nobody can explain which past decisions shaped the score, which writing style the tool prefers, or what happens to a candidate who needs an accommodation. That is how a time-saving tool becomes an unapproved employment policy.

AI is now easy to add to a recruitment workflow. It can draft a vacancy, extract experience from CVs, schedule interviews, answer routine questions and summarise notes. Those are useful capabilities. The mistake is to let the tool quietly decide who deserves attention, then call the result objective because it came from software.

This guide is for HR directors, founders and people leaders. It is governance guidance, not legal advice. The EU, UK and US sources below are useful benchmarks, but an organisation must check the employment, privacy, equality and accessibility rules that apply in its own country and obtain local advice where the decision is high-risk.

Start with one distinction: assist or decide?

“AI used in HR” is too broad to govern. Write down the exact task and its consequence.

The closer the output is to an employment decision, the stronger the controls must be. The EU AI Act is a useful signal even for organisations outside the EU: its Annex III identifies recruitment and selection systems, targeted job advertising, application filtering and candidate evaluation as high-risk use cases within its framework. That does not make the regulation local law, but it shows why “it is only a ranking tool” is not an adequate risk assessment.

Where AI can genuinely help

The best early uses reduce repetitive work while leaving the meaning of the decision visible.

Drafting and administration

Ask an approved tool to turn a signed-off role profile into a first draft of a vacancy, create a plain-language version, remove unexplained jargon, or prepare a checklist for interviewers. A person must check that the advert describes the real job and does not introduce unnecessary requirements or language that discourages qualified applicants.

Scheduling and candidate questions

Scheduling, reminders, location details and answers to routine process questions are usually more bounded than candidate scoring. Give the assistant a narrow knowledge base, a route to a human, and a rule that it must not invent a status, promise an outcome or request information that the process does not need.

Structured summaries

AI can turn structured interview notes into a consistent summary. It should not fill gaps with assumptions, infer personality from a face or voice, or convert a vague impression such as “not a good fit” into a polished paragraph. The interviewer remains responsible for the evidence and should be able to correct the summary.

Where bias enters the workflow

Bias is not only a malicious rule. It can enter through data, proxies, language, design choices, missing context or the way people respond to a score.

Historical decisions become a hidden target

If a tool learns from previous hiring decisions, it may learn the organisation’s old preferences rather than the capabilities the job actually needs. A history of hiring from a narrow network can make that network look like a definition of merit. A history of rejecting career breaks can make a gap appear to be a defect when it may have no bearing on performance.

Proxies look neutral

Names, schools, postcodes, employment gaps, language patterns, device signals and online behaviour can stand in for protected or sensitive characteristics. Removing a field does not remove its proxies. A model can also reward the format and vocabulary of applicants who have had better access to coaching, polished CV templates or the dominant business language.

Personality and emotion scores are especially fragile

Claims that a camera, voice recording or facial expression can reveal honesty, leadership or “culture fit” should trigger a stop-and-review decision. The signal may be weak, culturally narrow or inaccessible to a person with a disability. It can also encourage interviewers to trust a confidence score instead of asking job-related questions.

The US EEOC and Department of Justice warn that AI tools may screen out disabled applicants who could do the job with a reasonable accommodation. The practical lesson is wider than US law: accessibility cannot be an afterthought in an automated workflow. Ask how a candidate can request an accommodation, how the tool avoids penalising it, and who can override a bad result.

The human must stay in charge of five moments

“Human in the loop” is too vague if nobody can say what the human does. Name the moments where a person must review, question and own the result:

1. Before the advert is published: confirm the criteria describe the work, not an accidental profile of previous hires.
2. Before a candidate is screened out: check the reason against the stated criteria and look for missing context or an accommodation.
3. Before a shortlist is approved: review a representative sample, including people the tool ranked low.
4. Before an offer or rejection: make the decision from job-related evidence, not an unexplained score.
5. When a candidate challenges the process: provide a real route to human review and record what was reconsidered.

The ICO’s Recruitment Rewired work emphasises that human involvement should be consistent for candidates within the same hiring stage. A reviewer who checks only the borderline cases, or only the people who look promising, is not providing a reliable control.

What a fair-use policy should require

Use the following as the core of a two-page policy. Expand the controls when the tool influences selection, pay, promotion or dismissal.

Use the NIST AI Risk Management Framework as a useful vocabulary for validity, reliability, safety, security, transparency, explainability, privacy and fairness. It is voluntary and not a substitute for local law, but it helps an HR team ask better questions of a supplier and its own process.

Questions to ask an AI hiring vendor

• What exactly does the system do, and which employment decisions can it influence?
• What data trained or configured it? Is our candidate data used for general model training?
• Which languages, CV formats, assistive technologies and accessibility paths were tested?
• Can we see the factors or evidence behind a recommendation without exposing another candidate’s data?
• What happens when the model is updated? Will we receive a change notice and a new evaluation record?
• Where are candidate data, logs and backups stored, who can access them, and when are they deleted?
• Can the tool export a complete decision record and support a human override?

If the vendor cannot answer these questions, do not compensate with confidence. Narrow the use case to a lower-risk administrative task or pause the purchase.

A sensible 30-day starting plan

1. Week 1: inventory every AI feature already used in recruitment, including browser tools and informal staff use. Map each to assist, organise, recommend or decide.
2. Week 2: pause unapproved high-impact uses. Write the approved-use rules, name owners, define candidate notice and agree the accommodation route.
3. Week 3: test one bounded workflow against representative cases. Review errors, language, accessibility, unequal patterns and the quality of human overrides.
4. Week 4: publish the policy, train recruiters, create an incident and challenge log, and set a review date for the tool and the results.

A small organisation does not need a complicated committee to act fairly. It needs a visible decision rule, a human owner, a way for candidates to ask questions, and enough evidence to notice when the tool is treating people differently. Efficiency is worth having. Quietly automating an old bias is not.

For the wider operating model, combine this guide with the two-page AI policy template and the guide to responsible AI governance. If you need help reviewing an AI use case before it reaches candidates, contact Peter.