Underrepresentation occurs when some groups, perspectives, languages, or experiences are sparsely represented or absent from data used by language technologies and decision systems. What is rarely recorded leaves little statistical trace, so absence can distort how groups and experiences appear. section 7.3.4
Large language models often draw on high-resource languages and internet-accessible discourse. When prompts concern less represented contexts, models may rely on dominant narratives, flatten distinctions, or omit locally grounded knowledge. Historical training data also limits coverage of later events and developments. section 7.3.4
Sparse or uneven representation can contribute to unequal system outcomes:
- A healthcare algorithm used cost as a proxy for medical need, underestimating Black patients’ needs against a history of less care for the same level of illness. section 7.3.5
- Commercial speech-recognition systems had higher word error rates for Black speakers; dialect-insensitive hate-speech systems risked over-moderating African American English. section 7.3.5
- Gender-classification systems performed worse on darker-skinned women, while predictive-policing feedback loops could concentrate surveillance in already over-policed neighborhoods. section 7.3.5

