AI Glossary / Governance, ethics and safety / Intermediate
Sensitive attribute
Definition
A sensitive attribute is a characteristic of a person that deserves particular care because of legal, ethical, social, or personal considerations. In AI, such attributes may be relevant to privacy, discrimination, and fairness assessment. Sensitivity depends on context, and a technical label does not by itself determine the legal rules governing the information.
What it means
Examples can include disability, ethnicity, religious belief, health information, or other characteristics whose use or disclosure could affect people significantly. Different applications and jurisdictions draw different boundaries. Information that is not part of a particular legal category can still be sensitive in an everyday or ethical sense.
In fairness tools, sensitive attributes often define the groups whose outcomes or errors are compared. This use does not mean that the attribute must be an input to the prediction model. Data used for controlled evaluation may be separated from data used to make individual predictions, with appropriate access and purpose restrictions.
How it works
Identify why an attribute matters in the specific system: privacy, potential discrimination, unequal performance, or another concern. Determine how the information was obtained, whether it is accurate, and what uses are permitted. Avoid assuming that a field name captures a person's identity or circumstances perfectly.
For fairness assessment, define relevant groups and metrics, examine data completeness, and compare outcomes with attention to uncertainty and context. For protection, consider collection limits, access controls, retention, and disclosure risks. These goals can create practical tensions: missing group information can limit evaluation, while collecting it creates responsibilities. The solution requires justified governance, not an automatic instruction to collect everything.
Three practical examples
- An accessibility evaluation compares an assistant's performance for participants with different communication needs using appropriately governed study data.
- A service audit checks whether error rates differ across relevant age groups, while restricting who can access the underlying records.
- A privacy review examines whether free-text customer notes reveal health information even though the database has no field named health status.
Think of it this way
Think of sensitive attributes as information that may require special handling in a library of records. The label prompts attention to purpose and access, but does not tell you every rule. The analogy is limited because sensitivity can arise from context, inference, and combinations of data, not only from a clearly marked document.
Worked example
A fictional assessment contains 300 cases with an appropriately collected group attribute. Group A has 200 cases and 20 incorrect predictions, giving a 10% error rate. Group B has 100 cases and 20 errors, giving a 20% error rate. The observed difference is 10 percentage points, even though each group contributes the same number of errors.
There are also 60 cases with no group value. The reported group comparison does not describe how errors are distributed within those missing records. Analysts should investigate missingness and uncertainty rather than assigning people to groups based on guesses. The difference also does not, by itself, establish a causal explanation or a legal conclusion. It identifies a performance pattern requiring contextual investigation.
Illustrative case study
At fictional Aspen Learning, a team evaluates a speech-based study assistant. Participants can provide relevant information under the study's governance arrangements, but some choose not to answer certain questions. Analysts report performance for observed groups and clearly state the limits caused by missing data.
They keep the evaluation records separate from routine prediction inputs and review access to detailed results. The fictional case illustrates balancing assessment and protection. It does not establish a universal legal basis for collecting sensitive data, nor imply that a fairness purpose permits unrestricted inference of personal characteristics.
Common misconceptions and mistakes
Sensitive attribute is not a universal synonym for protected attribute or special-category personal data. Those expressions can have distinct legal or disciplinary meanings. For example, UK data-protection guidance distinguishes its defined special categories from other information that people may consider sensitive.
Removing an attribute from model inputs does not guarantee fairness because other variables can act as proxies. Conversely, including an attribute in a protected audit dataset does not automatically mean it is being used for individual decisions. Do not treat model-inferred attributes as unquestionably true or assume that a person's group membership is fixed, binary, or fully captured by a convenient coding scheme.
Business use
Careful handling supports trustworthy evaluation and reduces the risk of inappropriate use or disclosure. Organizations should record the purpose, permitted access, data quality, group definitions, and retention arrangements. Reports should distinguish measurement from decision-making and avoid publishing small-group details that expose individuals. The business objective is useful evidence with proportionate protection, not indiscriminate removal or accumulation of personal information.
People also ask
Must sensitive attributes be fed into the model to assess fairness?
Not necessarily. They may be used separately to compare outcomes or errors.
Does deleting them eliminate discrimination risk?
No. Proxies, historical patterns, and workflow decisions can still produce harmful differences.
Are all sensitive attributes legally classified the same way?
No. Legal categories and requirements vary by jurisdiction and purpose, while ethical sensitivity can extend beyond a formal list.
Check your understanding
An audit finds equal error counts for two groups of different sizes. Does that show equal performance?
Answer: No. Rates need appropriate denominators. In the example, 20 errors among 200 cases is 10%, while 20 among 100 is 20%. Missing group data and uncertainty also affect interpretation.
Related terms
- Protected attribute
- Privacy
- Missing data
- special-category data
- proxy variable
- fairness assessment
- group disparity
Learn next
For one hypothetical evaluation, separate the data needed for prediction from the data needed for assessment. Explain why each sensitive attribute is relevant, how missing values will be handled, and who can access it. Then identify which conclusions the resulting data can and cannot support.
More in Governance, ethics and safety
- Selective Adherence Bias - Selective Adherence Bias is the tendency to accept algorithmic advice more readily when it agrees with pre-existing beliefs or stereotypes, while...
- Self-healing system - A self-healing system detects certain faults and automatically takes predefined or policy-governed actions to restore acceptable operation.
- Semantic mapping - Semantic mapping connects items according to their meaning, rather than relying only on matching words or formats.
- Sensitivity analysis - Sensitivity analysis examines how a model's outputs or a decision's conclusions change when inputs, parameters, or assumptions are varied.
- Shallow learning - Shallow learning is an informal term for machine-learning approaches that do not rely on a long hierarchy of learned transformations.
- Sibyl attack - A Sybil attack occurs when one controlling entity presents multiple identities to gain influence that a system intends to distribute among independent...
Sources
Sources: Google Machine Learning Glossary, sensitive attribute: https://developers.google.com/machine-learning/glossary#sensitive-attribute . Fairlearn, fairness assessment: https://fairlearn.org/v0.12/user_guide/assessment/perform_fairness_assessment.html . ICO, What is special category data?: https://ico.org.uk/for-organisations/uk-gdpr-guidance-and-resources/lawful-basis/special-category-data/what-is-special-category-data/ . Aalto Dictionary of Machine Learning: https://aaltodictionaryofml.github.io/ADictML_Main.pdf .
Previous: Semivowel Next: Sensitivity analysis