Definition
The features that may not be used as the basis for decisions, such as race, religion, national origin, gender, marital status, age, and socioeconomic status.
What are protected variables
Protected variables are personal characteristics that should not be used in ways that lead to unfair discrimination. In many countries, these characteristics are already recognised in equality laws. When we talk about AI, the same idea carries over into how systems are built and tested. Common variables considered as protected variables are, for example, age, sex or gender, ethnicity or race, religion or belief, disability, or sexual orientation
These are called protected because people should not be treated worse simply because they belong to one of these groups.
Why this matters in AI
AI systems learn patterns from data. If the data reflects past biases in society, the system can copy those patterns without anyone meaning it to. For example, if historically fewer people from a certain group were approved for loans, an AI trained on that data might continue to reject them more often.
That can happen even if the system is not explicitly told someone’s race or gender. Other pieces of information, such as postcode or education history, can act as stand ins and indirectly reveal protected traits.
Therefore, considering protected variables is fundamentally about trust. When an algorithm plays a part in a decision about our lives, we want to know:
- We are not being judged on traits we cannot change
- The system has been checked for bias
- There is some form of accountability if things go wrong
Laws, standards and technical practices around protected variables are part of building that trust.
How do you "protect" variables
At a technical level, AI systems use input data, often called features, to make predictions. A protected variable is simply one of these features that relates to a sensitive personal trait. There are a few key ways these variables are handled in practice:
- Feature exclusion
The protected variable, such as gender, is removed from the set of inputs the model uses to make its prediction.
- Proxy detection
Developers look for other features that strongly correlate with a protected variable. For example, a postcode might be closely linked to ethnicity in some areas. Statistical tests and analysis help spot these hidden links.
- Fairness metrics
After the model is trained, its results are compared across groups. Teams might check whether approval rates, error rates or false rejections are much higher for one group than another.
- Constraint or adjustment methods
Some models are trained with extra mathematical rules that aim to keep differences between groups within certain limits. This can slightly change how the model weighs different inputs.
- Separate use for auditing
Even if a protected variable is not used to make decisions, it may be kept in a separate dataset so the system can be tested and audited for bias.
Key takeaways
- Protected variables are sensitive personal traits that can lead to unfair treatment if misused
- AI models treat them as data features, which can be removed, analysed or used for auditing
- Other data can act as proxies and indirectly reveal protected traits
- Technical tools like fairness metrics and model constraints help reduce bias
- Careful handling of these variables supports fairer and more trustworthy AI systems