Bias and fairness in recruitment

Indonesia's Law No. 13 of 2003 on Manpower contains two principles that bear directly on hiring. Article 5 says every worker has an equal opportunity, without discrimination, to obtain work. Article 6 says every worker has the right to equal treatment, without discrimination, from the employer. The law was amended by the Job Creation law now in force as Law No. 6 of 2023, and Constitutional Court decision No. 168/PUU-XXI/2023 changed several other employment provisions. So always check the latest official text before quoting it. This lesson is a general picture, not legal advice.

For everyday work we translate that principle into one rule: age, sex, religion, ethnicity or race, marital status, pregnancy, disability, place of residence or origin, photo or appearance, and health conditions not relevant to the job do not become screening criteria or rubric points, unless there is a lawful and genuine job requirement. A lawful requirement is usually tied directly to the task or to another rule, such as the minimum working age. Such requirements are checked with HR and legal, and the reason is written down. Otherwise, criteria judge skills and work results.

AI adds two kinds of risk. First, a model repeats the bias you give it. If a criterion says at most 30 years old or attractive appearance, the model will apply it diligently to every CV, and an unfairness that was one sentence now falls on hundreds of applicants. Second, model bias can appear unasked. A model learns from human text, including its prejudices. It may describe a leader with male pronouns, rate certain names or schools higher, or treat a career gap as a weakness when the gap may have been maternity leave or caring for family.

What is often missed are stand-ins: data that does not name a protected trait but carries it. Graduation year implies age. A photo implies appearance, sex and sometimes religion. An address can imply ethnicity or economic class. A career gap can imply pregnancy. A rubric that gives points for graduating no more than five years ago is in fact screening on age. So a bias check does not stop at a list of banned words; for each criterion it asks whether this really measures the job.

Test items can be biased too. A reasoning item that uses examples from a particular sport, regional terms or situations familiar only to one group tests familiarity, not reasoning. An item with more than one right answer punishes the candidate who reads more carefully. An interview rubric that gives points for confidence or a convincing way of speaking often favours a particular speaking style. When AI helps write items or rubrics, ask it to check for these, then check yourself: solve each item from scratch, and read each rubric point while picturing candidates from different backgrounds.

One small habit helps: keep a bias-check note for each opening. It lists the criteria removed or changed, the reason, and who checked. The note is useful when the same opening is posted again, because old requirements most often come back through copy-paste from the previous post.