Purpose limitation in health AI: why secondary use consent is not a catch-all
Secondary use consent does not cover purpose limitation. A patient who consents to data sharing for treatment does not consent to that same data training a sepsis prediction algorithm, feeding a payer risk model, or populating a pharma commercial dataset. Purpose limitation is a distinct legal and ethical constraint that broad consent language cannot absorb.
Broad secondary use consent is treated as a blank check in health AI. It is not.
A patient signs a consent form agreeing to the secondary use of their health data for "research purposes" or "quality improvement." Six months later, that data trains a commercial risk stratification model sold to a payer. Twelve months later, it feeds a pharmaceutical company's real-world evidence submission to the FDA. The patient consented to secondary use. They did not consent to any of those specific purposes.
This is the purpose limitation problem, and it is one of the most consequential gaps in health AI governance today.
What purpose limitation actually means
Purpose limitation is a principle embedded in nearly every major data protection framework. GDPR Article 5(1)(b) states that personal data must be collected for "specified, explicit, and legitimate purposes" and not further processed in a manner incompatible with those purposes. The Common Rule in the United States requires that human subjects research adhere to the scope defined in the informed consent document.
In health AI, purpose limitation means that the specific use of a data record must match the specific consent given for that record. Not the broad category. Not the general domain. The specific use.
A consent form that says "your data may be used for research" does not specify whether that research is academic, commercial, or regulatory. It does not specify whether the output will be a published paper, a licensed algorithm, or a payer decision engine that determines coverage. Each of those is a different purpose. Each requires separate consideration under purpose limitation.
How artificial intelligence affects informed consent in healthcare
Traditional informed consent was designed for a world where a clinician explained a procedure, a patient understood the risks, and both parties agreed to a defined course of action. AI breaks every assumption in that model.
First, the uses of data in AI are not fixed at the time of collection. A dataset collected for one clinical study can be repurposed, recombined, and retrained across dozens of models over years. The patient who consented in 2019 has no visibility into what their data is doing in 2025.
Second, AI outputs are emergent. A model trained on 500,000 records produces predictions that no single record holder anticipated. The informed consent doctrine assumes the patient can understand what they are consenting to. When data flows into opaque model architectures that produce probabilistic outputs used by third parties the patient has never interacted with, the concept of "informed" collapses.
Third, AI creates derivative value that was not contemplated at the time of consent. A radiology image consented for diagnostic use generates training signal that becomes a component of a licensed commercial product. The patient consented to the diagnosis. They did not consent to becoming an unlabeled contributor to a commercial asset.
Research published in BMC Medical Ethics found that fewer than 30% of patients understood how their data would be used in AI contexts, even after reading consent documents. The gap between what consent forms say and what patients believe they are agreeing to is not a minor discrepancy. It is structural.
In what situation is informed consent not needed
There are legitimate exceptions to informed consent requirements. HIPAA permits use of protected health information without patient authorization for treatment, payment, and healthcare operations. The Common Rule provides exemptions for research using de-identified data or data from publicly available sources. Public health surveillance, legally mandated reporting, and certain quality improvement activities can proceed without individual consent.
But these exceptions do not eliminate purpose limitation. They narrow the consent requirement. They do not erase the obligation to use data only for the purpose that justified the exception.
A hospital that uses patient data for internal quality improvement under a HIPAA operations exception cannot then share that same data with a commercial AI vendor for model training and claim the operations exception still applies. The purpose changed. The exception does not travel with the data to new purposes.
This distinction matters enormously in practice. Many health systems treat de-identification as a purpose limitation override: if the data is de-identified, they assume it can be used for anything. But de-identification addresses the identity risk. It does not address the purpose constraint. A de-identified dataset collected under a research consent still carries the purpose limitation of that original consent, even if no individual can be re-identified from it.
Why secondary use consent is not a catch-all
Secondary use consent typically appears as a single clause in a multi-page consent document. It uses language broad enough to cover nearly any future use. That breadth is the problem.
The European Data Protection Board has stated that consent must be "specific" to be valid under GDPR. Blanket consent for undefined future uses does not meet this standard. The Office for Civil Rights in the United States has similarly noted that HIPAA authorizations must describe the purpose of the use or disclosure "with sufficient specificity" for the individual to understand what they are agreeing to.
When a consent form says "your data may be used for future research," the following uses are all technically covered by that language:
These are five fundamentally different purposes with different risk profiles, different beneficiaries, and different commercial implications. Collapsing them into a single consent checkbox is not informed consent. It is consent theater.
The trust cost of purpose drift
Purpose drift occurs when data collected for one purpose migrates to another without explicit re-consent or governance review. It is pervasive in health AI because the technical infrastructure makes it easy and the consent infrastructure does nothing to prevent it.
A 2023 survey by the Pew Research Center found that 67% of Americans say they understand little to nothing about what companies do with their personal data. In healthcare, where the data is more sensitive and the stakes are higher, this understanding gap translates directly into trust erosion.
When patients discover that their health data was used for a purpose they did not anticipate, the damage is not hypothetical. The backlash against the UK's care.data program, which attempted to centralize NHS patient records for secondary use, led to 1.2 million patients opting out and the eventual shutdown of the program. The Royal Free Hospital's data-sharing arrangement with Google DeepMind resulted in a formal finding by the UK Information Commissioner that patient data had been used inappropriately.
These failures did not occur because consent was absent. They occurred because the purpose of use exceeded what patients believed they had agreed to. The consent existed. The trust did not.
Why shouldn't we use AI in healthcare
The question is not whether AI should be used in healthcare. The question is whether the data governance infrastructure supporting health AI is adequate to protect patients and maintain institutional trust.
The strongest argument against current health AI deployment is not about the algorithms. It is about the data supply chain. When training data lacks provenance, when consent does not match purpose, when patients cannot verify how their records were used, the resulting AI system inherits every governance failure in the data pipeline.
A sepsis prediction model trained on EHR data that was consented for clinical care but never explicitly consented for predictive model training is not just an ethical problem. It is a legal liability. It is a regulatory risk under emerging FDA guidance on AI/ML-based medical devices. And it is a trust problem that compounds over time as patients, regulators, and payers ask questions that health systems cannot answer.
The right response is not to avoid AI. It is to build the consent and purpose governance layer that makes AI use defensible. That means scoring every record for consent integrity before it enters a training pipeline. That means tracking purpose at the record level, not the dataset level. That means giving patients visibility into how their data is actually used, not just how it might be used.
Key statistics
How purpose limitation changes the consent architecture
Purpose limitation requires consent systems to track not just whether a patient consented, but what they consented to and whether the current use matches that scope. This is fundamentally different from a binary consent model.
A binary model asks: did the patient consent to secondary use? Yes or no. A purpose-aware model asks: did the patient consent to this specific secondary use? And if the use changes, does the system flag the mismatch, block the record, or trigger a re-consent workflow?
Most health data platforms do not support purpose-level consent tracking. They store a consent flag. They do not store the purpose the consent was tied to. This means that when a dataset moves from research to commercial use, or from internal quality improvement to external model training, there is no system-level check that catches the purpose mismatch.
ConsentOS, SuperTruth's five-tier consent governance architecture, addresses this by binding consent to purpose at the record level. Each record carries metadata specifying the consented purposes, the consent timestamp, the consent source, and the re-consent requirements triggered by purpose changes. When a query requests records for a purpose that exceeds the original consent scope, the system enforces the restriction before the data moves.
This is not optional governance. Under GDPR, under the emerging FDA AI framework, and under state-level health data privacy laws like the Washington My Health My Data Act, purpose limitation is becoming a legal requirement, not a best practice.
The DTI consent dimension and purpose scoring
The Data Trust Index scores consent as 20% of the total trust score for every health data record. Within that 20%, purpose alignment is a core factor. A record with valid consent but a purpose mismatch between the original consent and the current intended use receives a lower consent score than a record with purpose-matched consent.
This means that a dataset with 100% consent coverage but 40% purpose mismatch will score materially lower than a dataset with 80% consent coverage but 100% purpose alignment. The math reflects reality: consent without purpose alignment is not meaningful consent.
For organizations building health AI, this scoring approach changes how they evaluate training data. Instead of asking "do we have consent for these records," they ask "do we have consent for this specific use of these records." The distinction is the difference between a defensible AI system and a liability.
What this means for health systems, payers, and pharma
Health systems deploying clinical AI need to audit their consent infrastructure for purpose granularity. If their consent forms use broad secondary use language without specifying permissible purposes, every downstream AI use case carries unquantified legal risk.
Payers building risk models from claims and clinical data need to verify that the data they receive was consented for risk modeling, not just for clinical care. The HIPAA payment exception covers certain payer activities, but it does not cover building commercial AI products from clinical data.
Pharma companies using real-world evidence for regulatory submissions face the highest scrutiny. The FDA's emerging framework for AI/ML-based devices includes expectations around data provenance and consent integrity. An RWE submission built on records with purpose-mismatched consent is a submission waiting to be challenged.
In each case, the fix is the same: purpose-level consent tracking, enforced at the data layer, scored before the data reaches a model.
From consent theater to consent governance
The industry's current approach to secondary use consent is largely performative. Consent forms exist. Checkboxes are checked. But the connection between the consent and the actual use of the data is unmonitored, unenforced, and unscored.
Purpose limitation is the mechanism that converts consent from a checkbox into a governance function. It requires specificity in what is consented to, tracking of how data is actually used, enforcement when use exceeds scope, and re-consent workflows when purposes change.
This is not a theoretical framework. It is a practical requirement for any organization that wants its health AI to survive regulatory scrutiny, maintain patient trust, and produce outputs that are legally defensible.
The organizations that build purpose-aware consent infrastructure now will have a structural advantage as regulation tightens. The organizations that rely on broad secondary use consent as a catch-all will discover, under audit or in court, that their catch-all caught nothing.
The DTI Engine scores every health data record 0-100 across 8 trust dimensions, including purpose-level consent alignment, before your AI model sees it. ConsentOS enforces purpose limitation at the record level, blocking data movement when use exceeds consented scope. If your team is evaluating data for training, compliance, or clinical use and needs consent governance that goes beyond checkboxes, contact Louis Simeonidis at louis@supertruth.ai or (215) 918-4140.
Further reading:
Jason Alan Snyder
Co-founder of SuperTruth and Artists & Robots, and an inventor on the Data Trust Index patents. Twenty-plus years building technology inside Interpublic Group. He writes here nearly every day on data trust, provenance, and what AI should be allowed to act on, and publishes essays on his Substack.
About SuperTruth · LinkedIn · Substack · jasonalansnyder.com
See it in practice
DTI scores the record, not the patient.
8 dimensions. 0–100. Travels with every record permanently.