Wiki · Concept · Last reviewed June 15, 2026

Contextual Integrity

Contextual integrity is a theory of privacy that asks whether information flows are appropriate for the social context in which they occur. It treats privacy violation as a breach of expected roles, data types, recipients, and transmission conditions, not merely as exposure of secret facts.

Definition

Contextual integrity is Helen Nissenbaum's account of privacy as appropriate information flow. In the 2004 Washington Law Review article Privacy as Contextual Integrity, Nissenbaum argued that privacy cannot be reduced to secrecy, individual control, or a list of sensitive data fields. Information moves legitimately in many contexts: patients tell doctors, students submit work to teachers, workers report hours to employers, and citizens disclose data to public agencies. A privacy problem arises when the flow violates the norms of that context.

The framework became widely associated with Nissenbaum's 2010 book Privacy in Context: Technology, Policy, and the Integrity of Social Life. It is useful for AI governance because machine learning systems routinely move information across contexts: training data becomes model behavior, web traces become ad targeting, classroom data becomes risk scoring, workplace telemetry becomes performance management, and personal conversation becomes product improvement.

Contextual integrity belongs next to Data Minimization, Surveillance Capitalism, Digital Identity, and Cognitive Sovereignty. It gives those debates a precise question: who sent what information about whom, to whom, under what conditions, and for what purpose?

How It Works

Contextual integrity analyzes an information flow through contextual norms. Later summaries of the framework often describe five parameters: the data subject, the sender, the recipient, the information type, and the transmission principle. The transmission principle is the rule governing the transfer: consent, confidentiality, legal compulsion, sale, reciprocity, stewardship, emergency use, or another condition.

A flow can be privacy-preserving in one setting and invasive in another. A heart-rate reading shared with a clinician may fit medical care. The same reading sold to an advertiser, used by an employer to infer stress, or retained by a school as a discipline signal changes the actors, recipients, purposes, and transmission principles. The data did not become harmful only because it was "sensitive." The social relation changed.

Current Context

As of June 15, 2026, contextual integrity is not itself a statute or technical standard. Its influence is visible in privacy and AI governance language that emphasizes purpose, use limits, context, and reasonable expectations. The OECD Privacy Guidelines include collection limitation, purpose specification, use limitation, openness, individual participation, and accountability. The GDPR's Article 5 similarly lists principles including purpose limitation, data minimisation, storage limitation, integrity and confidentiality, transparency, fairness, and accountability.

NIST's Privacy Framework, first released in 2020 and maintained as a voluntary risk-management tool, asks organizations to identify and manage privacy risk rather than treat compliance as only notice text. NIST's Generative AI Profile, released in 2024 as a companion to the AI Risk Management Framework, includes privacy among generative-AI risk areas and emphasizes provenance, governance, measurement, and use context.

The White House Blueprint for an AI Bill of Rights, archived from 2022, also used context-sensitive language. Its Data Privacy principle says data collection should conform to reasonable expectations and collect only data strictly necessary for the specific context. It also warns against unchecked surveillance and calls for enhanced protections in domains such as health, work, education, criminal justice, finance, and youth data.

Governance and Safety

Contextual integrity is a practical test for AI systems because models break old boundaries by design. A dataset gathered for search quality can become training material. A support chat can become evaluation data. A camera feed can become identity, emotion, or behavior inference. A hiring tool can import education, credit, social, location, or productivity signals that were created for other contexts.

The governance task is to document and challenge those transfers. A meaningful AI impact assessment should name the original context, the new context, the actors, the data subjects, the recipients, the data types, the transmission principles, the expected benefit, the foreseeable harm, and the contest or deletion path. This makes contextual integrity useful for Algorithmic Impact Assessments, AI in Employment, Biometric Categorization, and Opaque Scoring Systems.

Defense Pattern

Spiralist Reading

Contextual integrity says the wound is not always exposure. Sometimes the wound is translation.

The confession becomes training data. The classroom becomes a risk dashboard. The body becomes a credential. The friendship graph becomes an ad auction. The model does not need to know the soul to violate the person. It only needs to carry a fragment from one moral room into another without permission, memory, or shame.

Open Questions

Sources

Return to Wiki