Skip to main content

AI in hiring decides about people. Someone has to answer for it.

Screening, ranking and assessing candidates is one of the clearest cases of a decision about a person made by profiling. Most of it runs on tools bought from suppliers, and the employer still answers for every outcome.

AIC.Karoo Lending (Demo)

Decision log: CV screening

Sample data

  • Applicant 4471Held for review
  • Applicant 4472No review needed
  • Applicant 4473Held for review
  • Applicant 4474No review needed

For

  • Heads of HR and talent
  • Recruitment agencies
  • HR technology companies
  • Employment equity committees

Where AI decides, and what goes wrong

  • Decides about: Applicants

    Screening CVs before a person reads them

    Qualified people filtered out on proxies such as a postcode, a gap in employment or the university attended.

  • Decides about: Candidates

    Ranking a shortlist

    The ranking learns who was hired before rather than who can do the job.

  • Decides about: Candidates

    Scoring video interviews and online assessments

    An accent, a disability or a poor connection is marked down as poor performance.

  • Decides about: Employees

    Flagging staff for performance or attrition risk

    A label follows someone into decisions about pay and promotion without their knowing.

A rejected candidate asks why

The moment that tests whether anyone really answers for an automated decision.

  1. 1

    The question arrives

    A candidate rejected minutes after applying writes to ask who decided and on what basis. Under POPIA section 71 they may make representations about an automated decision.

  2. 2

    You find the decision

    The decision log shows the screener's outcome, the reason it gave, and whether a person looked at it before the rejection went out.

  3. 3

    A named person answers

    The register names who answers for the screener. They review the rejection and record whether it stands, with their reason.

  4. 4

    The record holds

    The review, the reason and the person stay on the record, ready for an assessor or for the day the decision is challenged.

Rules that already apply

A summary, not legal advice about your organisation.

POPIA section 71

Restricts decisions based solely on automated processing that profile a person and have legal or substantial effect. Rejecting an application is such an effect. Read the guide

What it asks of youA person able to intervene, a chance for the candidate to make representations, and enough about the logic to make that meaningful.

Employment Equity Act, sections 6 and 8

Section 6 prohibits unfair discrimination in employment policy and practice. Section 8 allows psychological testing and similar assessments only if they are scientifically valid and reliable, applied fairly and not biased against any employee or group.

What it asks of youEvidence that an AI assessment is valid and fair for the people it is used on, whoever built it.

EU AI Act

Lists AI used for recruitment and selection as high-risk, for any company whose AI's output is used in the EU. Obligations apply from 2 December 2027. Read the guide

What it asks of youHuman oversight, logging, and a risk management system for the screening tools in scope.

Questions

Is our CV screener making decisions based solely on automated processing?

If no person meaningfully reviews an application before it is rejected, it very likely is. A person who approves rejections in bulk without looking does not change that. This is general information, not legal advice about your process.

We buy the screening tool. Are we still responsible?

Yes. Under POPIA the employer is the responsible party for the decisions it makes about applicants, and the supplier processes information for you as an operator under a written agreement. Asking the supplier how its AI behaves is part of answering for it.

Do we have to stop using AI in recruitment?

No. The law restricts decisions made solely by automated processing. A named person who can intervene, and a route for the candidate to be heard, are the safeguards it expects.