Risk · R-105 · Data Protection & Privacy
Inference of sensitive attributes
A system infers health, beliefs, sexual orientation or other special category information about a person from data that does not state it.
- Severity
- High
- Audience
- Heads-of · Practitioner
- Frameworks
- Section 13 · Lawfulness and transparency · Article 5 · Articles 22A to 22D
- Responsible AI domain
- Lifecycle stage
- Reviewed
- 2026-11-11
What it is
Models find correlations, and some correlations reveal what people have not disclosed: a pregnancy from shopping patterns, a health condition from typing or travel, religion from names and locations. The inference may be the system's purpose or a by-product nobody intended. Either way, an inference about an identifiable person is personal data, and an inference about a special category is treated with the same seriousness as a statement of it, whether or not it is correct. Intentional categorisation of people by biometric data is a prohibited practice, covered in biometric categorisation of sensitive attributes.
How it happens
- A model is given many ordinary attributes that together stand for a sensitive one.
- A segmentation or recommendation model groups people in ways that track a protected characteristic.
- A general-purpose model is asked to guess attributes from text or images.
- Inferred labels are stored and reused by other systems.
Why it matters
Processing special category data needs a specific condition in law, and an organisation that did not know it was inferring such data will not have one. Decisions that then differ by the inferred attribute may be discriminatory. People experience such inferences as intrusive, and a wrong one can follow them.
Warning signs
- Segments or scores correlate strongly with a protected characteristic.
- Nobody assessed what the model's features might reveal in combination.
- Inferred attributes appear as fields in downstream systems.
- Marketing or pricing differs in ways that track health or family status.
Who owns it
The business owner of the system is accountable. The data protection officer advises on whether special category data is being processed. Data science tests what the model has learned.
In practice
An insurer's marketing model learns that customers who buy certain products at certain times are likely to be undergoing fertility treatment, and adjusts the messages they receive. No one designed it to do so. The same pattern generalises to any model with rich behavioural data and freedom to find what predicts.
Mitigating controls
Primary controls
- Data minimisation and lawful basis review Directive
- Requires a basis for inferences as for collected data. Before personal data is used to train, fine-tune, ground or run an AI system, a documented review confirms the lawful basis, the compatibility of the purpose and that only necessary data is used.
- Data protection impact assessment for AI Directive
- Identifies inferred special category data as processing that needs justification. A data protection impact assessment is completed before AI processing that is likely to result in high risk to individuals begins, and is kept current as the system changes.
Supporting controls
- Bias testing and fairness monitoring Detective
- Testing reveals when outputs track protected characteristics. A system's performance and outcomes are measured separately for each relevant group before release and on a set cycle in production, against thresholds that trigger investigation.
- Intended-use definition and use restrictions Directive
- Excludes sensitive inference from the approved purpose. The approved purpose, users, population and conditions of use of each AI system are stated precisely, and any use outside them requires reassessment before it begins.
Framework and clause references
| Framework | Clause | Title |
|---|---|---|
| Equality Act 2010 | Section 13 | Direct discrimination Inferred characteristics can ground discrimination. |
| Guidance on AI and data protection | Lawfulness and transparency | Lawfulness, fairness and transparency Lawfulness of inferences. |
| UK General Data Protection Regulation | Article 5 | Principles relating to processing of personal data Fairness and purpose limitation. |
| UK General Data Protection Regulation | Articles 22A to 22D | Automated processing and significant decisions Tighter limits where special category data is used. |
Sources
- Guidance on AI and data protection UK ICO · guidance
Related terms
How to cite
Axiom Verity (2026). Inference of sensitive attributes. Reference Library, R-105. https://axiom-verity.com/risks/inference-of-sensitive-attributes. Reviewed 2026-11-11.
Found an error or a better source? Report a correction.