AI Recruitment: screening CVs without screening people out

AI recruitment tools score and rank CVs against patterns in past hires. In South Africa that makes them an employment practice under the Employment Equity Act, which covers applicants, not only staff. A screen that filters on a proxy for a listed ground is unfair discrimination the employer must justify, and POPIA section 71 forbids a purely automated rejection.

Most guides to AI recruitment are written for a human resources department with a budget line for software and a lawyer down the corridor. That is not the reader here. The reader here is an owner who gets two hundred applications for one role, has a week to fill it, and has just been shown a tool that promises to hand back the top ten. The promise is real. So is the exposure, and almost none of the vendor pages mention it.

This post reads the exposure from the primary documents: the Employment Equity Act 55 of 1998 as amended, the Protection of Personal Information Act 4 of 2013, and the one large, well documented case of an AI recruiting engine going wrong, as Reuters reported it in October 2018. All three were read directly on 12 September 2026. The point is not that you should avoid AI recruitment. It is that the law already tells you what a screen must look like, and a screen built that way also hires better.

What does an AI recruitment tool actually do to a CV?

Strip away the interface and most AI recruitment software does one of two things. The older kind matches keywords: the job advert says "five years experience" and "CA(SA)", and the tool counts how many applicants say the same words. The newer kind is trained on past outcomes: it is shown the CVs of people who were hired and people who were not, and it learns which features predict a hire. It then gives each new applicant a score.

The second kind is the one that sells, because it needs no rules written by you. It is also the one the law has the most to say about, because its rules are not written by anyone. They are inferred from your history, and your history is where the bias lives.

What happened when Amazon built one?

The clearest public account of this failure is still the one Reuters published on 10 October 2018, from five people who had worked on the project. Amazon's team had been building programs since 2014 to review applicants' CVs and score them from one to five stars. By 2015 they found the system was not rating candidates for technical roles in a gender-neutral way.

The reason is the whole lesson. The models were trained on CVs submitted to the company over a ten-year period, and most of those came from men. The system taught itself that male candidates were preferable. It penalised CVs that included the word "women's", as in "women's chess club captain", and it downgraded graduates of two all-women's colleges. Amazon edited the programs to be neutral to those particular terms, but, in the words of the people Reuters spoke to, that was no guarantee the machines would not devise other ways of sorting candidates that could prove discriminatory. The models also favoured verbs more common on male engineers' CVs, such as "executed" and "captured".

Two details matter for a small employer. First, nobody wrote a rule that said "prefer men". The rule emerged from the data, and it expressed itself through proxies, a word, a college, a verb. Second, gender was not the only problem. The people familiar with the project said unqualified candidates were often recommended for all manner of jobs, with results almost at random, and that is why the project was shut down. A biased screen and a useless screen turned out to be the same screen.

Why is a biased filter a legal problem in South Africa, not just a bad hire?

Because the Employment Equity Act says so in plain words, and it says so about applicants. Four sections do the work.

Section 6(1): the prohibition, and the phrase most vendors have never read

No person may unfairly discriminate, directly or indirectly, against an employee, in any employment policy or practice, on one or more grounds, including race, gender, sex, pregnancy, marital status, family responsibility, ethnic or social origin, colour, sexual orientation, age, disability, religion, HIV status, conscience, belief, political opinion, culture, language, birth, or on any other arbitrary ground. The words "or on any other arbitrary ground" were added by the 2013 amendment. A filter does not have to name a listed ground to be caught. "Indirectly" covers the proxy, the Amazon word, the Amazon college, and "any other arbitrary ground" covers a filter that is simply irrational.

Section 9: applicants count

For purposes of sections 6, 7 and 8, "employee" includes an applicant for employment. A CV screen is an employment practice applied to applicants, so the prohibition reaches it before anyone is hired. This is the section that turns a screening tool from a private efficiency into a regulated act.

Section 11: who has to prove what

If discrimination is alleged on a listed ground, the employer must prove, on a balance of probabilities, that it did not take place as alleged, or that it is rational and not unfair, or is otherwise justifiable. Read that against a trained model. To prove a screen was rational you need to be able to say what it screened on. If the honest answer is "the vendor's model scored them and we took the top ten", you cannot discharge the burden, and the burden is yours. On an arbitrary ground the complainant carries it instead, which is some comfort, but a gender or age proxy is a listed ground.

Section 6(2)(b): the only defence that works

It is not unfair discrimination to distinguish, exclude or prefer any person on the basis of an inherent requirement of a job. This is the line a lawful screen is built on. A filter for a driver's licence on a driving job is an inherent requirement. A filter for an unbroken employment history is not, and it will land hardest on women who took parental leave and on anyone who was retrenched in a bad year. The test for every criterion in your screen is: could I stand up and call this an inherent requirement of this job?

What does section 8 require of any assessment?

Section 8 of the Act prohibits psychometric testing and other similar assessments of an employee unless the test or assessment has been scientifically shown to be valid and reliable, can be applied fairly to employees, and is not biased against any employee or group. Whether a CV scoring model is a "similar assessment" in the sense of section 8 is not something we have found a South African court decide either way, and we are not going to pretend it has. But the three tests are exactly the right standard to hold a vendor to, because they are the three things Amazon's engine failed: it was not reliable, it could not be applied fairly, and it was biased against a group. Ask a vendor for the validity evidence. If they cannot produce it, the law has already told you what to think.

What does the Act say "suitably qualified" means?

This is the section nobody building a filter reads, and it is the most useful one. Section 20(3) says a person may be suitably qualified for a job as a result of any one of, or any combination of, four things: formal qualifications, prior learning, relevant experience, or the capacity to acquire, within a reasonable time, the ability to do the job. Section 20(4) says an employer determining suitability must review all four. Section 20(5) says an employer may not unfairly discriminate against a person solely on the grounds of that person's lack of relevant experience.

Section 20 sits in the chapter on affirmative action, which binds designated employers, so a business with fewer than fifty employees is not directly bound by it. It is still the Act's own definition of a suitably qualified person, and it reads like a specification for a screen. A keyword filter that rejects every CV without "five years experience" is rejecting on the one criterion the Act says cannot be the sole ground. A filter that cannot see prior learning or capacity to learn, because those do not appear as keywords, is blind to half the definition. Build the screen on the four legs and you are both closer to the law and closer to the candidate who will actually do the job.

What does POPIA add when the screen is automated?

Section 71(1) of POPIA says a data subject may not be subject to a decision which results in legal consequences for them, or which affects them to a substantial degree, which is based solely on the automated processing of personal information intended to provide a profile of that person, and the section names performance at work as one of the things such a profile might cover. A rejection from a job you applied for affects you to a substantial degree. If the rejection was produced by the tool alone, section 71(1) is engaged.

Section 71(2) gives the way out. The prohibition does not apply if the decision is taken in connection with the conclusion or execution of a contract and appropriate measures have been taken to protect the data subject's legitimate interests. Section 71(3) then says what appropriate measures are: an opportunity for the data subject to make representations about the decision, and sufficient information about the underlying logic of the automated processing to enable them to do so. In practice that means three things. A human makes the final call, so the decision is not based solely on the automation. Applicants are told a tool is used and how to query its result. And you can explain the logic, which rules out any tool whose vendor cannot explain it to you. We covered the same section for customer service bots in what to hand to a bot and what must stay human; the reasoning is identical for applicants.

How does the screening work actually get reduced?

None of the above says do it by hand. Two hundred CVs read by a tired owner on a Sunday is not a fairer process, it is a less documented one. The law is telling you what to automate, and in what order. This is what that looks like built properly.

Capture once: a structured application, not a CV parser

The unfairness in a CV screen starts with the CV, a free-text document where the same fact is expressed a hundred ways and where a chess club can cost you the job. Replace it with an application form that asks every applicant the same questions, in the same order, with answers that land as fields: the licence, the qualification, the specific tasks done before, the date available. The tool never has to infer anything, because nothing is free text. The same record later becomes the employment contract and the onboarding file, so the applicant's details are typed once, by the applicant.

Reuse: the inherent requirements, written once, scored in the open

Before the advert goes out, write the inherent requirements of the job as a short list, and mark each one as a must or a score. That list is the entire screening rule. Every applicant is scored against those items and nothing else, and the rule is readable by a person in thirty seconds, which is what section 11 will ask of you. A language model is useful here for one narrow job: reading an applicant's free-text answer to "describe a time you did X" and checking whether it answers the question. It is not useful for ranking people against each other on a signal nobody can name.

Stop re-keying: the shortlist is a view, not a spreadsheet

When the answers are fields, the shortlist is a filter on the same records, not a new document. Interview notes attach to the record. The offer is generated from it. The equity data you will need for an employment equity report, if you grow into being a designated employer, is already on the file because the applicant declared it, in a field, separately from the scored answers, so it never touches the screen.

Where the reminder comes from: the notice and the human step are steps, not memory

The section 71 notice goes out automatically when the application is received. The human review is a task on the workflow that has to be ticked before a rejection can send. The request for representations is a reply-to address that opens the record. Nobody has to remember the law on a busy week, because the law is a step in the sequence.

What the automation must not do is the same thing Amazon's could not do: decide alone. Your AI policy should say that in one sentence, and your vendor checklist should make the vendor say how their tool supports it.

Five questions to ask any AI recruitment vendor

1. What does the model score on? If the answer is "patterns in successful hires", ask which patterns, and ask what happens when the successful hires were mostly one group.
2. Can you show validity and reliability evidence for this role type, the section 8 standard, whether or not a court has applied it to you yet?
3. Can a human at my business read the reason for any individual score? If not, section 11 cannot be met and section 71(3) cannot be satisfied.
4. Where is the applicant data processed and stored, and for how long after the role closes? The POPIA checklist covers the rest of that conversation.
5. Does the tool support a mandatory human decision before any rejection is sent, and a candidate query route? If it does not, it is built for a jurisdiction that does not have section 71.

Frequently Asked Questions

Is it legal to use AI recruitment software in South Africa?

Yes, with conditions. The Employment Equity Act prohibits unfair discrimination against applicants in any employment practice, and POPIA section 71 prohibits a decision based solely on automated profiling unless a human can be asked to reconsider and the logic can be explained. Use the tool to organise, not to decide alone.

Does the Employment Equity Act apply to job applicants or only to employees?

Section 9 says that for purposes of the anti-discrimination sections, "employee" includes an applicant for employment. A CV screen is therefore covered from the first application, regardless of whether the person is ever hired.

Can I filter CVs on years of experience?

You can score it. Section 20(5) of the Act says an employer may not unfairly discriminate solely on a lack of relevant experience, and section 20(3) lists experience as one of four routes to being suitably qualified, alongside formal qualifications, prior learning and the capacity to learn the job in a reasonable time.

Must I tell applicants that AI is screening them?

If the tool contributes to a decision that affects them substantially, POPIA section 71(3) requires that they can make representations and are given enough information about the underlying logic to do so. Practically, that means a notice at application and a route to query the outcome.

What went wrong with Amazon's AI recruiting tool?

According to Reuters in October 2018, it was trained on ten years of CVs that were mostly from men, learned to penalise the word "women's" and two all-women's colleges, and also recommended unqualified candidates almost at random. Amazon edited the terms, found that was no guarantee, and disbanded the project.

The short version

An AI recruitment tool is an employment practice applied to applicants, so the Employment Equity Act reaches it from the first CV. Build the screen on inherent requirements you can name, score against the four legs of "suitably qualified", keep a human decision before any rejection, and be able to explain the logic to the person who asks. Capture applications as fields rather than free text and most of the screening effort disappears without a model ever ranking a person. If you want to see where hiring, onboarding and records in your business currently get re-typed and where a tool is safe to put, that is what a Business Autopsy is for, and a discovery call is where it starts.

Sources, all read directly on 12 September 2026: Employment Equity Act 55 of 1998 as amended, sections 6, 8, 9, 11 and 20, from a consolidated text carrying the amendments up to the Employment Equity Amendment Act 4 of 2022, cross-checked against the Department of Employment and Labour's own summary of the Act (form EEA3). Protection of Personal Information Act 4 of 2013, section 71, from the Government Gazette text. Reuters, "Amazon scraps secret AI recruiting tool that showed bias against women", Jeffrey Dastin, 10 October 2018, re-read today under our rule that any account older than twelve months is re-fetched from its source before it ships. No figures from vendor marketing material are used.