Is AI CV screening biased? How to check before you rely on it

AI CV screening can be biased, because it learns from past hiring and reads proxies such as names, universities and graduation years. Check it before you rely on it: test it on past roles, compare outcomes across candidate groups where you lawfully can, remove proxy fields, and keep a named person deciding every shortlist and rejection.

#Key takeaways

• Bias in screening usually comes from the data and the criteria, not from the model being careless.
• Test any screening tool on roles you have already filled before it touches a live candidate.
• Look for proxies: names, photos, nationality, graduation years, career gaps and university names.
• A person decides every shortlist and rejection, and the reason is recorded in their words.
• Re-test after every change to the model, the prompt or the criteria.

#How does bias get into AI screening?

The best-known case is still instructive. According to Reuters via Euronews, october 2018, Amazon dropped an experimental recruiting tool after it learned to penalise CVs containing the word 'women's', because it had been trained on a decade of mostly male hires. Nobody told it to discriminate; the history did.

Modern language models are not trained on your hires, but they still pick up patterns. Ask one to rank CVs for 'a strong fit' and it applies its own idea of strong, which can favour certain universities, unbroken careers or English fluency for roles that do not need it.

#What rules apply to AI in hiring?

Rules differ by country, so treat this as a map, not advice. Lewis Silkin, april 2024 explains that the EU AI Act treats AI used in recruitment and selection as high-risk, with most employers acting as deployers. The European Commission AI Act service desk, checked november 2026 gives a job matching and ranking tool that filters candidates as an example of a high-risk use.

In New York City, employers and employment agencies may not use an automated employment decision tool without a bias audit, according to the NYC consumer and worker protection FAQ, checked november 2026. Gulf employers also handle candidate data under the UAE personal data law and the Saudi PDPL. Ask your counsel which apply to your hiring.

#How do you test a screening tool before using it?

1. write the criteria first: the must-haves for the role, agreed with the hiring manager before any CV is read.
2. remove proxy fields: strip names, photos, dates of birth, nationality and marital status before the model sees the CV, and check graduation years and gaps.
3. replay past roles: run the tool on anonymised CVs from roles already filled and compare its shortlist with the recruiter's.
4. compare across groups: where the law lets you hold the data, compare pass rates for different candidate groups and investigate any gap.
5. swap tests: change one detail, such as a name or a university, on otherwise identical CVs and see whether the result moves.
6. check hidden text: look for white text or instructions inside CVs, a prompt injection risk ranked LLM01 by the OWASP top 10 for LLM applications.

Illustrative example, the numbers are assumptions: if the tool passes 40 of 100 CVs from one group and 25 of 100 from another with similar qualifications, that gap needs an explanation before go-live, not after a complaint.

#What should a person always decide?

Every shortlist and every rejection. The tool summarises each CV against the agreed criteria and shows its evidence; a recruiter reads the summary, opens the CV where needed and decides. The NIST AI risk management framework, january 2023 treats human oversight and accountability as part of managing AI risk, and in hiring that means a named decision-maker on every outcome.

#When is a simpler approach enough?

If you hire a few people per quarter, skip automated screening. A structured scorecard, two readers and anonymised CVs give you most of the fairness benefit with no model involved. Summarising CVs in an approved chat tool is fine, as long as nobody asks it to rank people.

#Where does Stennir fit?

Stennir's AI consulting service includes governance: the audit maps where AI touches your hiring, who decides, what is logged and which tests run before and after go-live. Where screening support is worth building, AI implementation builds it with proxy stripping, evidence for every summary and a recruiter approving each decision.

AI training teaches recruiters what screening output gets wrong and how to challenge it.

#What should a head of talent do next?

List every place AI already touches a candidate decision, including tools your ATS switched on by default. For each, note who decides and whether anything was tested. Read AI for operations, the resume screening worked example and what the board should see before rollout. Then book a 30-minute call.

#Common questions