Algorithmic bias occurs when an artificial intelligence system learns, reproduces, or even amplifies the prejudices contained in the data with which it was trained. Since models learn from historical examples, if those data reflect social inequalities or stereotypes, the system will tend to perpetuate them in its predictions and decisions, often in a way that is invisible to the user.
It matters because these systems are used in sensitive areas where an unfair decision has real consequences. Some common examples:
- Personnel selection that penalizes certain profiles based on gender or age.
- Credit granting that discriminates by zip code or ethnicity.
- Facial recognition with higher error rates on dark skin tones.
An important nuance is that bias does not always stem from the data: it can also be introduced in the model design, in the choice of variables, or in how the results are interpreted. Therefore, mitigating it requires audits, representative data, and continuous human supervision, not just occasional technical adjustments.