What is AI Bias?
AI bias is a systematic skew or prejudice in the outputs of an artificial intelligence system. It occurs when an algorithm produces results that are consistently unfair, inaccurate, or discriminatory toward specific groups of people or concepts.
How does it work?
A common misconception is that AI bias comes solely from flawed training data. While biased data (e.g., a hiring dataset containing mostly male resumes) is a massive contributor, bias can also arise from:
- Labelling: Humans providing subjective or prejudiced labels during supervised learning.
- Model design: Developers optimizing the algorithm for metrics that inadvertently harm a demographic.
- Evaluation: Testing the model on a benchmark that lacks diversity, giving a false sense of fairness.
- User interaction: Users deliberately prompting the model to reinforce their own biases.
What is a simple example?
If a facial recognition system is trained primarily on images of people with lighter skin tones, the resulting model will perform significantly worse when attempting to recognize people with darker skin tones. This is a direct consequence of biased training data representation.
Why does it matter?
AI models are increasingly used to make critical decisions about loan approvals, criminal justice sentencing, and medical diagnoses. If these systems are biased, they act as scalable engines of discrimination, automating and amplifying societal inequalities.