AI systems can be affected by several sources and types of bias, and these can create significant challenges for HR managers:
Perhaps the most common source of bias is data bias, which occurs when the training data used to build an AI model doesn’t accurately represent the real world. In such circumstances, the model will likely perpetuate and even amplify existing societal biases. There are several types of data bias:
- Historical Bias. Data reflecting past prejudices and societal inequalities (e.g., hiring practices from decades ago under-representing women in certain fields).
- Representation Bias. Certain groups are underrepresented or overrepresented in the dataset. (e.g., a facial recognition system trained primarily on images of one ethnicity might perform poorly on others).
- Measurement Bias. The way data is collected or labelled introduces bias (e.g., using biased search terms when collecting training data for a language model).
Even with unbiased data, algorithms themselves can introduce bias due to design choices. The features chosen to train the model can disproportionately affect certain groups, introducing bias.
If the optimisation function used to train the model isn’t carefully considered, it can lead to biased outcomes. Prioritising overall accuracy over fairness metrics can lead to disparate performance across groups.
The biases of the people involved in designing, developing, and deploying AI systems can inadvertently creep into the process. This includes biases in data labelling, feature engineering, and model evaluation.
Other types of human biases that can impact on the model design. Cognitive Bias occurs when developers have personal bias that can impact dataset or model behaviours. Confirmation bias is caused by relying too much on pre-existing beliefs or trends in the data. Exclusion bias derives from leaving out important data, which may derive from not knowing what one doesn’t know.
In the context of HR management, some types of bias are more significant than others
Recruitment algorithms may favour candidates who fit a certain type, ignoring equally qualified people with different backgrounds. This bias can lead to a workforce that lacks diversity and limits opportunities for underrepresented groups.
AI-powered applicant tracking systems often favour candidates whose CVs resemble those of past recruits, unintentionally sidelining those from nontraditional backgrounds
Automated performance evaluations can deepen workplace inequities, as HR AI interprets productivity metrics without accounting for structural biases in assignments or opportunities.
AI systems tend to perpetuate societal biases, and it is difficult for human decision-makers to correct these biases later.