The AI Recruitment Arms Race: When Algorithms Assess Algorithms

Overview 

• AI on both sides: Candidates use AI to write, tailor and mass-submit applications, while employers use AI to screen them. The screener can end up rewarding tool use instead of job capability.
• Weaker signals: Keyword matches, polished writing and fluent interview answers no longer show reliably what a candidate can do.
• Legal exposure: POPIA section 71 restricts decisions based solely on automated profiling, and the Employment Equity Act puts the burden on employers to show their assessments are valid and unbiased.
• Our position: AI should rank candidates, not reject them on its own. A named person approves every shortlist, and skills are tested through work the candidate does live.

What happens when AI applicants meet AI screeners?

When candidates use AI to apply and employers use AI to screen, the process starts measuring how well someone uses AI tools rather than whether they can do the job. Neither side intends this. It happens because both systems optimise for the same surface signals: keywords, structure and fluent answers.

For South African employers, the cost is more than a noisy shortlist. A hiring decision you cannot explain is one you will struggle to defend under POPIA or the Employment Equity Act. Below, we cover how the loop forms, where the legal risk sits, and what a defensible use of AI in recruitment looks like.

How are candidates using AI to apply for jobs?

Candidates now use AI at almost every stage of applying. A Gartner survey of 3,290 job candidates found that 39% used AI during the application process, mostly to write CV text, cover letters and assessment answers. In a separate Gartner survey, 6% of candidates admitted to interview fraud, and Gartner predicts that one in four candidate profiles worldwide will be fake by 2028.

Three types of tool drive most of this:

• Application agents scan job boards, tailor a CV to each advert's keywords and submit at scale. A recruiter looking at 400 applications may be looking at far fewer people who actually read the advert.
• CV optimisers rewrite experience to mirror the job description, so the CV scores well against keyword filters but says less about the person.
• Live interview assistants listen to questions during a video interview and put suggested answers on screen while the candidate speaks.

Trust is low in both directions. The same Gartner research found only 26% of candidates trust AI to evaluate them fairly. For practical ways to test for genuine intent, see Combating the Apply-Bot Surge.

Why do AI screening tools struggle with AI-assisted applications?

Most AI screening tools were calibrated on human applicant behaviour, so AI-assisted applications break the assumptions they rely on. An optimised application hits keyword targets and avoids the small irregularities that sometimes mark a genuine one as weak. The screener can end up advancing bot-assisted applications over authentic ones.

Candidates also change how they present themselves when they know AI is assessing them. A 2025 study in PNAS found that people assessed by AI emphasised analytical traits and played down intuitive or emotional ones, because they believed AI rewards analytical thinking. The researchers showed this shift can change who gets selected.

The feedback loop makes it worse. A model trained on past outcomes learns from whoever got through, so if AI-assisted applications keep advancing, it learns to reward those signals and the skew grows with each hiring round. Our article on algorithmic bias in AI-assisted recruitment explains how this kind of drift starts.

What does South African law say about AI hiring decisions?

South African law does not ban AI screening, but it makes fully automated rejection hard to defend. Three provisions matter most.

1. POPIA section 71. Section 71 says a person may not be subject to a decision with legal consequences or substantial effect that is based solely on automated profiling, including of their work performance, reliability or conduct. Rejecting a job applicant arguably meets that threshold. The exceptions come with safeguards, such as letting the person make representations and giving them enough information about the underlying logic to do so.
2. Employment Equity Act section 8. Psychological tests and "other similar assessments" must be scientifically valid and reliable, fair to everyone assessed and free of bias, and section 9 applies this to job applicants. AI tools that score personality, tone or facial expression arguably fall within this. The HPCSA certification requirement fell away on 1 January 2025, but the other three tests still apply.
3. Employment Equity Act section 11. When unfair discrimination is alleged on a listed ground, the burden of proof sits with the employer. If your screening logic is a black box, you cannot show the decision was rational and fair.

Buying a vendor's tool does not transfer the liability to the vendor. The organisation that deployed it answers for its decisions. For a wider view of these obligations, see our recruitment compliance management solution and our breakdown of the pros and cons of recruitment automation.

This is a summary, not legal advice. Check your process with a labour law specialist.

Graylink's view: AI should rank candidates, not reject them

Our position is that AI should narrow and order the pool, and a named person should make every shortlisting and rejection decision. That keeps the speed benefit of automation while leaving a human who can explain each outcome.

We also think trying to out-detect AI-written applications is a losing race, because writing tools improve as fast as detectors do. Changing what you measure lasts longer than trying to catch the tool. If the only thing between a candidate and your shortlist is a model scoring text, you are effectively hiring for text.

For the same reason, we would advise against video tools that score facial expressions or vocal patterns. They are hard to validate under section 8, and live coaching tools have made the signals they read unreliable. Fraud checks still belong in the workflow, but as agents that flag identity gaps or scripted answers for a recruiter to review, not as automatic rejections.

How can employers keep AI screening fair and defensible?

Use AI for volume and sorting, and put human judgement wherever a decision affects a candidate. Five safeguards cover most of the risk:

1. Human sign-off on every shortlist. A recruiter reviews and approves the shortlist before it moves forward, and the reason is recorded.
2. Skills tested live. Work samples, timed practical tasks and structured problem-solving are much harder to fake with a coaching tool than a written answer. Our guide to recruitment assessment tools covers the options.
3. Criteria you can explain. If you cannot explain your screening criteria to a candidate who asks, you will struggle to defend them in a dispute.
4. Verification without paranoia. Short checkpoints that confirm identity and intent, such as a live conversation through a recruitment chatbot like txthr, filter out bots without punishing genuine candidates whose CVs happen to be polished.
5. An audit trail that outlasts any tool. Whatever the candidate used to apply and whatever assisted your screening, the record of who decided what, and why, should sit in your applicant tracking system and feed your EE and compliance reporting.

A quick test for your current process

You can check whether your screening is defensible by picking one recently rejected applicant and answering four questions:

1. Which criteria screened them out, and were those criteria set before the vacancy opened?
2. Did a named person review the decision, and is that review recorded?
3. Could you explain the reason to the candidate in plain language?
4. If a tool scored personality or behaviour, can the vendor show validity and bias testing on South African candidates?

Any answer of "no" or "we would have to ask the vendor" is the first gap to close.

Final takeaway

AI on both sides of hiring is already normal, and it will keep getting better at writing applications and at reading them. The employers who hire well will be clear about what they are measuring and will keep a named person accountable for each decision.

Tools will change every year. Under POPIA and the Employment Equity Act, the accountability stays with the employer.

Frequently asked questions

Can South African employers reject job applicants using AI alone?

It carries real risk. POPIA section 71 restricts decisions based solely on automated profiling that substantially affect a person, and the Employment Equity Act puts the burden on employers to show screening was fair. A person should review and approve rejections.

How can recruiters tell if an application was written with AI?

Often they cannot, and AI detection tools are unreliable. It is more effective to test what AI cannot easily supply, such as live skills tasks, short real-time conversations and structured interviews that probe the candidate's own examples.

Do AI video interview tools fall under the Employment Equity Act?

Tools that score personality, tone or facial expression arguably count as "other similar assessments" under section 8. If so, the employer must be able to show they are valid, reliable, fair and unbiased for every applicant assessed.