Stockertim Dictionary · Human resources · Technology
Proxy discrimination
Also called discrimination by proxy
In short
Unfair treatment that arises when a model relies on data, such as postcode, school or career gaps, that stands in for a protected characteristic.
- Advanced
- Concept
- Checked
In detail
Leaving protected characteristics out of a model does not keep them out of its decisions. Many ordinary inputs are correlated with them. Postcode can track ethnicity, a gap in employment can track sex or disability, graduation year tracks age, and membership of certain clubs or the wording of a CV can track social background. A model that learns from such inputs can reproduce the pattern it would have produced with the characteristic itself.
Proxy discrimination is especially likely when a model is trained to imitate past decisions that were themselves biased. The model finds whatever signals best predict those decisions, and proxies often do. In Great Britain the result can amount to indirect discrimination under the Equality Act 2010 if a group is put at a particular disadvantage without justification; the 2024 audits of AI recruitment tools by the UK data protection regulator found tools that inferred characteristics such as gender and ethnicity from names.
Detection relies on testing outcomes by group, examining which inputs drive results and checking whether each input is genuinely job-related. Inputs without a clear job link are removed even if they help prediction.
In practice
Model reviews list every input with its business reason, test whether each input correlates with protected characteristics, and remove or constrain inputs that act as proxies without a strong job-related justification.
Watch out
- A model can find a proxy in a combination of harmless-looking inputs.
- Dropping a proxy can lower measured accuracy; that is often the price of fairness.
Learn it properly
More lessons about proxy discrimination
Official sources
- Equality Act 2010, section 19: Indirect discrimination · legislation.gov.uk (The National Archives)
- Intervention into AI recruitment tools leads to better data protection for job seekers · Information Commission (UK)
- Towards a Standard for Identifying and Managing Bias in Artificial Intelligence, NIST SP 1270 · National Institute of Standards and Technology
Facts last checked on 3 November 2026. Rules and figures can change: check the official source before you rely on them.
Education, not advice: the Stockertim Dictionary explains what terms mean. It never recommends buying or selling anything.