Age Bias in AI Systems

Age bias is not one technical problem. It may arise from protected-age rules in a decision workflow, from stereotypes in language, or from demographic differences in biometric performance. In 2022, the U.S. Equal Employment Opportunity Commission sued iTutorGroup, alleging its application software automatically rejected female tutor applicants age 55 or older and male applicants age 60 or older. The case was resolved by a consent decree in 2023. This is a concrete employment example, but it does not establish that all automated hiring tools discriminate by age. NIST’s Face Recognition Vendor Test evaluated algorithms submitted by developers and found demographic differentials, including by age, in many tested algorithms. NIST cautions against broad claims across systems: results depend on algorithm, application and data. A separate 2024 NIST evaluation of age-estimation software found sensitivity to image quality, gender, region of birth and age; accuracy of age estimates is different from recognizing a person’s identity. In LLMs, a 2024 study tested six models using value-orientation prompts and reported age-related patterns, while a 2026 AgeismSet study created a benchmark for measuring ageism in model outputs. These are prompt-based research evaluations, not proof of how all older users are treated in products. Age groups are internally diverse. Chronological age, life stage, disability, health, language, digital access and experience may affect a task differently. A fairness review should test the actual use, compare specific age bands, and distinguish an adverse employment decision from a face-matching error or stereotyped text generation. Avoid building an “older user” persona from assumptions; include older adults in research and give people accessible ways to correct outputs.