In a 2025 IEEE S&P study, 30 professional developers implemented three programming tasks designed around GDPR. 79 of the 90 solutions were not compliant, and participants named purpose limitation, consent, and data minimization as the requirements they struggled with. A 2026 PoPETs interview study with 15 developers describes the same gap in organizations: a privacy policy that looks compliant from outside, decoupled to varying degrees from what the software does. The authors call this symbolic compliance.
Privacy maturity models are one way organizations measure progress on gaps like these. The NIST Privacy Framework tiers and the AICPA/CICA Privacy Maturity Model both score the privacy program. An organization can reach a high tier while its pipelines copy restricted data into unlabeled tables. The five levels below score the data platform: what it checks, where, and when.
What existing models measure
NIST Privacy Framework 1.0 defines four Implementation Tiers: Partial, Risk Informed, Repeatable, and Adaptive. The 1.1 draft keeps them. Each tier is described by four elements: the risk management process, the integrated risk management program, relationships in the data processing ecosystem, and the workforce. NIST calls the tiers "a progression, albeit not a compulsory one" and ties success to the outcomes in a Target Profile rather than to a tier.
The AICPA/CICA Privacy Maturity Model (2011) scores 73 GAPP criteria on five levels borrowed from the Capability Maturity Model: ad hoc, repeatable, defined, managed, and optimized. Its levels describe how formal, documented, and reviewed each practice is.
The engineering outcomes are in the NIST Core, as three subcategories that the tier descriptions never reference:
- ID.IM-P5. The purposes for the data actions are inventoried.
- CT.DM-P7. Mechanisms for transmitting processing permissions and related data values with data elements are established and in place.
- CT.DM-P9. Technical measures implemented to manage data processing are tested and assessed.
The five levels turn these outcomes into checks, ordered so that each level closes the gap left by the one beneath it.
The five levels
Level 1, Documented. Purposes exist in the privacy policy and in the records of processing that GDPR Article 30 requires. Engineers map purposes to tables in design review. The platform has no record of purposes, so violations surface in audits, after the data has been used.
Level 2, Declared. Purposes are machine-readable labels on systems and datasets. In the open-source Fides project, a system's privacy_declarations list data categories, a data use, and data subjects, and a CLI evaluates them against policy rules such as "no user.contact data for advertising". The evaluation reads only the declarations, and nothing observes the running jobs. When the code changes and the declaration does not, the two diverge without any check failing, which is the decoupling the PoPETs study describes.
Level 3, Checked at read. Each job declares a purpose at submission, and the query engine compares it with the labels on every table and column the job reads. Disallowed reads fail, and per-row consent is applied by filtering against the consent store. The check stops at the read. A job allowed to read a column can write its values into a new table with no label, and every reader of that table is unrestricted.
Level 4, Enforced across flows. Labels propagate through joins, writes, and training jobs. Each output carries the union of its inputs' restrictions, which is the join operation in the label lattice of information flow control. Meta's public description of Policy Zones shows one implementation: a SQL parser extracts the inputs and outputs of each query, a lattice-based evaluator checks each flow and whether a consent check ran, and training outputs inherit the workflow's annotation. A new policy runs in logging mode until its violations are fixed, then switches to enforcement mode, where unremediated flows are denied. The recorded flows also form the lineage graph from the previous post in this series. Enforcement still happens at run time, so a code change that violates a policy, or a policy change that blocks a legitimate flow, is found only when the job runs in production.
Level 5, Verified before deployment. Changes are evaluated before they ship. For a code change, the new queries are parsed and evaluated against current labels before merge. For a policy change, the evaluator replays recently recorded flows under the new policy and diffs the allow and deny results. The diff reports two failure classes: flows that would violate the policy, and allowed data that the new rule would drop. Meta's Policy Zones Manager does this for proposed annotations by simulating enforcement on downstream flows.
Figure 1 shows the check that each level adds. At Levels 1 to 3, nothing checks the path from a source table to a derived table and then to a model.
Figure 1.
Scoring a pipeline
A pipeline is at Level N when the checks for Levels 1 through N all pass. A check that passes above a failed level does not count. In the example below, labels propagate on write, but jobs run without a declared purpose, so the labels have nothing to be compared with. The pipeline stays at Level 2, and each comment points to the evidence for its check.
pipeline: churn_features_daily checks: L1_purposes_inventoried: pass # records of processing, entry 14 L2_datasets_labeled: pass # catalog tags on all 6 inputs L3_job_purpose_checked: fail # Spark jobs submit no purpose L4_labels_propagate: pass # output tables inherit labels L5_changes_tested_predeploy: fail # no policy check in CI level: 2 # highest N with L1..LN all passing next: require a purpose at job submission Mapping to legal requirements
No law prescribes a level. Each requirement below asks for a record, and the level shown is the first that produces it from the platform instead of a document.
- GDPR Article 30 and NIST ID.IM-P5 ask for an inventory of purposes. Level 1 produces it.
- CCPA risk-assessment regulations (in effect since January 1, 2026) state that the purpose of processing must not be described in generic terms, and make training automated decisionmaking technology for significant decisions a trigger for an assessment. First submissions to the CPPA are due April 1, 2028. A purpose declared per job and checked at read (Level 3) ties the purpose in the assessment to what the job is allowed to read.
- EU AI Act Article 10(2)(b) requires data governance for high-risk training data to cover the origin of the data and, for personal data, the original purpose of collection. At Level 4, lineage records the origin and the propagated labels record the purpose. The Digital Omnibus moved the Annex III high-risk obligations to December 2, 2027.
- FTC algorithmic-disgorgement orders, such as Everalbum (2021) and Rite Aid (2023), required deletion of models and algorithms derived from the data in question. Finding those models requires lineage from source tables to models (Level 4).
Choosing a target level
As with the NIST tiers, not every pipeline needs the top level. The mapping above sets a floor of Level 4 for pipelines that feed model training or join sensitive categories, because the FTC orders and, for high-risk systems, the AI Act both depend on lineage into models. Level 2 or 3 is enough for the rest. Levels 4 and 5 need query-engine integration, label governance, and a fix for label creep, where one incidental read restricts every dataset derived from it. Meta's 2025 post describes a fix through governed reclassification. With the checklist kept per pipeline, the gap between each pipeline's level and its target becomes a list of failed checks that can be assigned and tracked.
References
- Horstmann, S. A., et al. "Sorry for bugging you so much." Exploring Developers' Behavior Towards Privacy-Compliant Implementation. IEEE Symposium on Security and Privacy, 2025.
- Saha, R. P., et al. From Lines of Code to Lines of Policy? Exploring Software Developers' Perceptions of Their Privacy Policy-Related Activities. Proceedings on Privacy Enhancing Technologies 2026(3).
- NIST. Privacy Framework Version 1.0. January 16, 2020.
- CPA Canada and AICPA. Privacy Maturity Model. 2011.
- Ethyca. Fideslang: Policy resource and System resource.
- Meta Engineering. How Meta enforces purpose limitation via Privacy Aware Infrastructure at scale. August 27, 2024.
- Meta Engineering. Policy Zones: How Meta enforces purpose limitation at scale in batch processing systems. July 23, 2025.
- California Privacy Protection Agency. Notice of approval: CCPA updates, cybersecurity audits, risk assessments, and ADMT regulations. September 2025.
- European Commission AI Act Service Desk. Regulation (EU) 2024/1689, Article 10.
- Jones Walker. Yes, August 2 Still Matters: The EU Approved a High-Risk AI Delay, but Most Transparency Obligations Remain.
- Regulation (EU) 2016/679 (GDPR), Article 30.
