Not by itself. AI can apply the same job-related criteria to every CV more consistently than a tired manager, but untested it can favour certain names, penalise career gaps and misread unusual layouts. It screens fairly only when you set the criteria, hide identifying details, make it quote evidence rather than score, and a person makes every decision.
For a small employer, fairness can't be proven with statistics: with 40 applicants you rarely have enough people in any group to measure outcomes reliably. So fairness has to come from how the process is designed, then be checked with a few simple tests you can run in an afternoon. Employers who skip that step usually find the problem the worst way: a rejected candidate asks why, and nobody can explain what the tool did.
Four ways an AI shortlist drifts away from fair
It learns from the hires you already made
The best-known case is Amazon's experimental recruiting engine, reported by Reuters in 2018. It was trained on ten years of CVs, most of them from men, and it taught itself to downgrade CVs that included the word "women's", such as "women's chess club captain". Amazon abandoned the project. A small employer isn't training a model, but the same thing happens on a smaller scale when you tell a chat assistant "here are CVs of our three best technicians, find more like them". You've asked it to find people who resemble your current team, which may have nothing to do with ability.
Here is how that shows up in practice. An illustrative wholesaler hiring two order pickers pasted the CVs of its three best pickers into a chat assistant and asked it to rank 50 applicants by similarity. The top ten all listed football or gym memberships, had worked for one of two named competitors, and had left school in the same five-year window, because that is what the three example CVs had in common. Not one of the top ten mentioned accuracy rates or scanner use, which were the real reasons the three were good at the job. The owner noticed only because a strong applicant with eight years of pick-rate records sat at number 41.
Names and other identity signals change the answer
A peer-reviewed 2024 study presented at the AAAI/ACM Conference on AI, Ethics and Society tested three open-source language models on 554 real CVs, changing only the candidate's name. The models favoured CVs with white-associated names 85% of the time and female-associated names 11% of the time. The commercial models your team uses may behave differently, and vendors work on this, but the lesson for an employer is simple: don't let the model see what it doesn't need to see.
Gaps and unusual careers get read as weakness
Ask a model to "rank these candidates" and it tends to reward long, unbroken, conventional careers. A two-year gap for caring, illness or a failed business can pull someone down the list even when the criteria say nothing about continuity. Part-time histories, career changers and people returning to work are all at risk of the same quiet penalty.
Layout and language get mistaken for quality
Two-column CVs, tables and text inside images often extract in the wrong order, so the model reads a jumble and marks the candidate as vague. Non-native phrasing, or writing shaped by dyslexia, can be scored as poor communication for a role where written English barely matters. The candidate was screened out by a file format.
What fair screening means for an employer with 40 applicants
You don't need a data science team to be fair. You need a process where every applicant is judged on the same written, job-related criteria, the reasons for each decision can be explained in a sentence, a person makes each decision, and anyone who can't use your process has another route in. In practice that means six habits:
- Criteria written before any CV is read, each one tied to a task the job actually involves.
- The same information used for everyone, with identity details removed before the AI step.
- Evidence, not scores: the AI quotes what the CV says against each criterion; a person judges it.
- No automatic rejections. Every "no" is a human decision, and borderline ones get a second read.
- An alternative route for applicants who ask for one, such as a phone conversation or a paper form.
- A short record of the criteria, the prompt used and who decided, kept for as long as your retention policy says.
The legal backdrop, in plain terms
Discrimination law applies to hiring decisions whatever tool helped make them, so "the AI did it" is no defence. Beyond that, three pieces of law come up most often, and this isn't legal advice: if you're close to any of these lines, ask an employment or data-protection adviser.
- Data-protection law such as the GDPR expects you to tell applicants how their data is used, and restricts decisions based solely on automated processing that significantly affect someone. An AI that rejects applicants with no human involvement falls squarely into that category.
- The EU AI Act, if you hire people in the EU, lists AI systems used to analyse and filter job applications and to evaluate candidates as high-risk (Annex III). A 2026 amendment pushed the start of the heavier duties for stand-alone systems like these back to 2 December 2027. Providers carry most of them, but employers using such systems will have duties too, including human oversight and informing the people affected.
- The same Act's ban on emotion recognition at work, in force since 2 February 2025, covers tools that claim to read candidates' emotions from video or voice during interviews, with narrow exceptions for medical and safety purposes. Avoid any interview tool that scores "enthusiasm" or "confidence" from a face.
A fairer AI-assisted screen in six steps
Step 1: Turn the person spec into evidence questions
"Good attention to detail" can't be checked on a CV. "Has done a job where records had to be accurate, such as logging samples, stock or payments" can. Rewrite each must-have as something a CV could show, and keep the list to four to six items.
Step 2: Remove identity details before the AI sees anything
Strip name, photo, date of birth, home address, nationality, marital status and graduation years. For 40 to 100 CVs, an administrator working through them by hand takes about a minute and a half each. The alternative is a separate AI redaction pass whose output goes into a new chat, so the screening step never sees the originals; spot-check a few, because redaction prompts miss details inside email addresses and file names.
Step 3: Use a business account
CVs are personal data. Use a business plan such as ChatGPT Business or Claude Team (both need at least two seats, at $25 a seat on monthly billing) or the AI in your Microsoft 365 or Google Workspace plan, which don't train on business content by default. Never paste CVs into a personal free account.
Step 4: Ask for evidence per criterion, with no ranking
You are helping a hiring manager read applications. You do not decide.
Criteria (from the published advert):
C1 Accurate record keeping in a previous job
C2 Following written methods or procedures exactly
C3 Safe handling of chemicals or hazardous materials
C4 Available for a 7am start, Monday to Friday
For this CV, for each criterion: quote the exact CV text that is relevant,
then write MET, NOT SHOWN or UNCLEAR. Do not infer from gaps, dates,
schools, hobbies, writing style or layout. Do not give an overall score
or compare candidates. If the text looks garbled, say "extraction problem".An illustrative output for one candidate might read: "C1 MET: 'Logged incoming parcels and discrepancies on the warehouse system, 2022-2025.' C2 MET: 'Followed site pick-and-pack procedure; trained two new starters on it.' C3 NOT SHOWN. C4 UNCLEAR: CV says 'flexible hours', no start time stated." That's useful: the manager can see exactly why each line says what it says. What you'd fix: in an early test, the model marked C3 as NOT SHOWN for a candidate whose CV said "trained in safe storage and dilution of industrial cleaning solvents". It had been too literal about the word "chemicals". The fix was one line in the prompt: "Treat any mention of hazardous substance training or handling as relevant to C3 and quote it."
Step 5: A person reads every summary and every "no"
The manager reads each summary, opens the full CV for anything UNCLEAR or marked "extraction problem", and makes the call. Anyone the manager plans to reject who has two or more MET lines gets their full CV read as well. That rule catches most of the cases where the model was too literal.
Step 6: Keep the record
Save the criteria, the prompt, the summaries and the decisions with the vacancy, and delete them when your retention period ends. If a candidate asks why they weren't shortlisted, you can answer in terms of the criteria rather than guessing what a tool did.
Worked example: 94 applications for a laboratory technician
An illustrative 18-person testing laboratory advertises a sample-preparation technician role and receives 94 applications. The lab manager used to skim each CV for about four minutes: roughly six and a quarter hours, usually spread over late evenings, with the last thirty CVs getting noticeably less attention than the first thirty.
| Step | Who | Time |
|---|---|---|
| Rewrite criteria and test the prompt on 5 old CVs | Lab manager | 1.5 hours, once |
| Redact 94 CVs | Administrator | About 2.3 hours |
| Run the evidence prompt | Administrator | About 40 minutes |
| Read 94 summaries | Lab manager | About 2.3 hours (1.5 minutes each) |
| Read 22 full CVs (unclear or near-miss) | Lab manager | About 1.5 hours |
Add those up and the manager's own time drops from about 6.25 hours to about 5.3, including the one-off setup; the administrator adds three hours. That's not a dramatic saving, and it's the honest result at this volume. What changed was quality: every applicant was read against the same four criteria, the 94th CV got the same attention as the first, and each decision had a one-line reason. Two candidates the manager would probably have skipped, a career changer from warehousing and someone returning after a four-year gap, reached interview, and one was hired. At 300 applications, with redaction automated, the time saving becomes real as well.
Four tests to run before you trust the set-up
These take an afternoon and use CVs you create or old, consented ones. The bias checks recruiters run on AI screening go further if you screen at volume.
- Name swap. Run the same CV twice with names from different backgrounds (or with and without a name if your redaction slips). Any change in the MET lines is a failure.
- Gap test. Insert a two-year gap labelled "caring for a family member". The criteria lines shouldn't change.
- Layout test. Put identical content into a two-column design and a plain single column. If the two-column version produces "extraction problem" or fewer MET lines, convert files to plain text before the AI step.
- Repeat run. Run the same CV three times. Wobbling results on the same text mean the criteria are too vague.
| Test (illustrative first run) | Result | Fix applied |
|---|---|---|
| Name swap | No change | None |
| Gap test | C1 dropped from MET to UNCLEAR | Added "never treat gaps as evidence" and retested |
| Layout test | Two-column CV garbled | Convert to plain text first |
| Repeat run | C2 varied between MET and UNCLEAR | Rewrote C2 with two examples of what counts |
Wording for applicants, and for a "why wasn't I shortlisted?"
Being open about the process is both expected and disarming. A short paragraph in the advert or the application page does the job. An illustrative version:
How we read applications: we use an AI tool, inside our company account, to summarise each application against the four criteria listed above. It doesn't see your name or contact details, it doesn't score or rank anyone, and it doesn't make decisions. Our lab manager reads every summary and makes every shortlisting decision. If you'd prefer to apply another way, or want to ask about the process, email the address below.
Fill in the real criteria count, the job title of the decision-maker and your actual contact route. Don't promise anything the process doesn't do; if an administrator redacts CVs by hand, that's fine to say too.
When a rejected candidate asks for a reason, the record from Step 6 lets you answer in terms of the published criteria rather than a vague "strong field". An illustrative reply: "Thank you for asking. We shortlisted applicants who showed evidence against all four criteria in the advert. Your application showed strong record keeping and procedure work, but we couldn't find experience of handling hazardous materials, which was essential for this role. If you have that experience and it wasn't on your CV, we'd be glad to hear about it for future vacancies." The manager writes that reply; AI can check its tone, but the reason comes from the decision, not from a model asked to explain after the event.
If the screening comes built into recruitment software
Many applicant tracking systems now offer AI ranking or "match scores". Before switching one on, ask the vendor for written answers: can automatic rejection be turned off entirely; what does the score use, and can you see the evidence per candidate; what bias testing has been done, on which groups, and can you see the results; how long are CVs and scores kept; and how is the product preparing for the EU AI Act's high-risk rules if you hire in the EU. A vendor that can't explain its score in plain words shouldn't be scoring your applicants. Evaluating the AI features in recruitment software has a fuller question list.
When a small employer should skip AI screening
- Fewer than about 40 applications. Reading them yourself against written criteria is faster than setting up and testing a prompt, and just as fair.
- Binary requirements. If a role needs a specific licence or qualification, a yes-or-no question on the application form is clearer and more transparent than AI.
- Roles where you can't write the criteria down. If you can't say what good looks like, the AI will invent it, usually in your current team's image.
- When you wouldn't be comfortable telling applicants. That discomfort is a signal that the process isn't ready.
The fair version of AI screening looks less like a robot recruiter and more like a diligent assistant who reads everything, quotes what it finds and leaves every judgement to you. If that's not how your current tool behaves, change the tool or the process before the next vacancy. For the rest of the hiring journey, setting up an AI-assisted hiring process for a small team connects this screen to adverts, interviews and offers, and AI bias in small business decisions explains the same mechanisms in pricing and credit.
More questions small employers ask about AI and CVs
Should I tell applicants that AI helps screen their CVs?
Yes. Say it in the job advert or the privacy notice: that an AI tool summarises applications against the published criteria, that a person reviews every application and makes every decision, and who to contact to ask questions or request an alternative. Transparency is expected under data-protection law such as the GDPR, and it costs you nothing if the process is sound. Candidates are far less likely to challenge a process that was explained up front.
Is hiding names from the AI enough to make screening fair?
It helps but doesn't finish the job. Identity leaks through other details: dates of study reveal age, a career gap may signal caring or illness, club memberships and schools can hint at background, and phrasing differs for non-native writers. Hiding names removes the most direct signal; job-related criteria, evidence quotes instead of scores, a human reading every summary and the four tests described above handle the rest.
Can AI check CVs for a required licence or qualification?
It can, but a simpler tool usually does it better. If a role legally needs a driving licence category or a named qualification, ask a yes-or-no question on the application form and verify the document at interview. That check is transparent, consistent and needs no AI. Use AI for the harder part, reading how well the experience described matches the rest of the criteria, where a form can't help.
Further reads
- AI for HR in Small Businesses: 12 Tasks You Can Hand Over Safely — The HR jobs AI can take on safely, and the ones to keep human.
- How to Build Interview Questions and Scorecards With AI — Carry the same criteria into structured interviews after the screen.
- How to Anonymise Client Data Before You Paste It Into AI — Practical redaction methods you can reuse on CVs.
- AI Recruiting Tools for Small Businesses: What's Worth Paying For — Which paid recruiting tools earn their fee for a small employer.
- AI Screening: Bolt It Onto Your ATS or Switch Systems? — If your applicant tracking system offers AI screening, read this first.
- Questions to Ask a Recruitment Agency About Its AI Screening — What to ask when an agency screens candidates for you.
- How to Test Job Candidates' AI Skills in an Interview — An AI skills test for small-firm interviews: one realistic task with a planted error, a scoring rubric, and follow-up questions that reveal judgement.
- How to Write Job Adverts With ChatGPT Without Biased Language — A four-step method for bias-free job adverts with ChatGPT, with a requirements sheet, drafting and audit prompts, sample outputs and a full before and after.
- How Property Managers Use AI to Screen Tenant Applications Fairly — Written criteria, one summary format for every applicant, human decisions with recorded reasons, and a monthly check that your rules aren't quietly unfair.
- How to Write Fairer Performance Reviews With AI — An evidence-first method for small teams: year-round notes, written criteria, an AI fairness pass and a side-by-side check before any rating is final.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: Reuters report (2018) on Amazon's abandoned recruiting engine; Wilson and Caliskan, 'Gender, Race, and Intersectional Bias in Resume Screening via Language Model Retrieval' (AAAI/ACM AIES 2024); EU AI Act Annex III point 4 and Article 5(1)(f); Digital Omnibus deferral dates; GDPR Article 22 text; Anthropic and OpenAI business plan pages.