Most people think of their personal data as the profile fields they fill out: name, email address, date of birth. That picture is wrong by several orders of magnitude. The data economy runs on a far deeper archive, one that the people who generate it almost never see and almost never control. Dr. Patrick Fisher named this reality in The Invisible Data, Volume 6 of The Invisible Series, and the argument has only grown more urgent as large-scale AI training has moved from research labs into production infrastructure that touches billions of people every day.
This article unpacks what invisible data is, how AI training pipelines exploit the structural gap between data subjects and data holders, and what the Personal Data Asset Origination System (PDAOS) offers as a technical correction rather than a regulatory hope.
What Invisible Data Actually Means
The phrase "invisible data" is not a metaphor. It refers to data that is generated by a person, about a person, without that person having any meaningful record that the data exists. Dr. Fisher's framing in The Invisible Data draws a sharp line between data that users consciously produce, such as a post or a search query, and data that platforms infer, derive or synthesize from behavioral signals the user was never told were being captured.
This second category is the one that matters most to AI systems. Dwell time on a paragraph, scroll velocity, mouse hover patterns, keystroke cadence, the sequence in which a user abandons a form: none of these appear in any privacy policy in legible form, yet all of them feed into models that predict intent, emotion and purchasing behavior.
The invisibility is not accidental. It is an architectural feature of the attention economy. Platforms that disclose the full depth of their signal collection lose a competitive advantage. So the disclosure stops at the surface layer, and the deeper archive remains invisible to the person who generated it.
The Layers of Data Harvested Without User Awareness
To reason clearly about personal data ownership, it helps to map the actual layers of data collection. They run deeper than most privacy discourse acknowledges.
Layer 1: Declared data. Name, age, location, account preferences. Users know this exists.
Layer 2: Observed behavioral data. Clickstreams, session duration, device fingerprints, navigation paths. Partially disclosed in cookie banners. Rarely explained in terms of what is actually recorded.
Layer 3: Inferred data. Emotional state predictions, income estimates, political affiliation scoring, health condition inference. Generated entirely by platform models. Never shown to the user. Often more accurate than self-reported equivalents.
Layer 4: Synthetic derived data. Embeddings, latent space representations, vector encodings of user identity that persist inside model weights even after the source record is deleted. This is the layer that data rights frameworks like GDPR's right to erasure have the hardest time addressing, because the data no longer exists as a row in a database. It exists as learned weights.
That fourth layer is where the AI training era introduces a genuinely new problem, not just an amplified version of an old one.
How the AI Training Era Intensifies the Asymmetry
Pre-AI data harvesting was extractive but relatively contained. A platform collected signals, built a profile, served targeted content. The data subject could, in principle, request a copy of that profile under GDPR Article 15 or CCPA. The profile had edges. It lived in tables that could be queried and partially audited.
Large language model training breaks that containment entirely. When a model trains on a corpus that includes user-generated text, behavioral logs or interaction histories, the user's data does not enter the model as a retrievable record. It dissolves into parameter space. The model learns statistical patterns from it, carries those patterns forward into every inference it makes, and cannot unlearn them in any technically verified sense even when a retraction or deletion is requested.
The scale compounds the problem. A frontier model trained on web-scale data may have processed the writing, questions, complaints and confessions of hundreds of millions of people, none of whom consented to that specific use, none of whom were compensated, and none of whom can verify whether their data contributed to a particular output.
This is the asymmetry Dr. Fisher identifies as structurally distinct from earlier privacy violations. It is not that a company knows things about you that you would prefer they did not. It is that your cognitive and behavioral patterns have been absorbed into a system that now generates economic value continuously, at scale, without attribution, without compensation and without any mechanism for you to audit your contribution or revoke it.
Research from the machine unlearning literature demonstrates how difficult verified data removal from trained models actually is. Proposed techniques including gradient-based unlearning and SISA training offer partial solutions, but none provide cryptographic guarantees of the kind that a genuine data ownership framework requires.
The Provenance Gap at the Heart of the Problem
Every serious technical discussion of data ownership eventually arrives at the same bottleneck: provenance. Who created a piece of data, when, under what conditions, and for what stated purpose?
Current data infrastructure does not answer those questions from the subject's perspective. Provenance records exist, but they exist inside platform systems. The data holder controls the ledger. The data subject has no independent record that a court, a regulator or an automated consent system could verify.
This is not a regulatory gap that better laws can fill on their own. Laws require verifiable facts to enforce. When the provenance record is held exclusively by the party with the most incentive to obscure it, enforcement is structurally compromised before it begins.
The W3C Data Privacy Vocabulary Community Group has worked toward standardized provenance annotation through the Data Privacy Vocabulary (DPV), which provides a taxonomy for expressing purposes, legal bases and data categories in machine-readable form. This is useful work. It does not, by itself, solve the provenance gap, because annotation standards only function when both parties use them and when neither party can unilaterally rewrite the record.
What is needed is a provenance infrastructure that is cryptographically anchored, subject-controlled and interoperable across the data lifecycle. That is precisely what PDAOS is designed to provide.
How PDAOS Infrastructure Corrects the Structural Imbalance
The Personal Data Asset Origination System (PDAOS) is the technical infrastructure developed by Own Your Data Inc to make personal data ownership enforceable at the cryptographic level rather than the policy level. The distinction is critical. Policy-level ownership means a document says you own your data. Cryptographic-level ownership means no party can use your data without a verifiable authorization token that you issued and that you can revoke.
PDAOS approaches this through three interlocking mechanisms.
Data origination anchoring. When a personal data asset is created, PDAOS records a cryptographic commitment to its content and provenance on a structure that the data subject controls. This is not a copy of the data stored somewhere else. It is a hash-based proof that a specific piece of data existed, was created by a specific subject and was released under specific terms at a specific time. The data holder cannot alter that record without breaking the cryptographic proof.
Decentralized identifiers (DIDs). PDAOS uses the W3C DID specification to anchor data subject identity without relying on a platform-controlled identifier. A DID is self-sovereign: it is created by the subject, controlled by the subject and can be used to sign consent authorizations without requiring a trusted intermediary to validate it.
Verifiable credentials for data authorization. Rather than a checkbox consent event that disappears into a platform's server logs, PDAOS issues W3C Verifiable Credentials as consent artifacts. These credentials are cryptographically signed by the data subject, specify the exact purpose and duration of authorization and can be presented and verified by any system in the data supply chain without requiring the original platform to confirm them.
Applied to AI training specifically, PDAOS infrastructure means that a model trainer would need to hold a verifiable credential from each contributing data subject before including that subject's data in a training corpus. The credential specifies the permitted use. If the trainer uses the data outside that scope, the breach is verifiable against a cryptographic record the subject controls.
Consent Receipts and Selective Disclosure as Technical Controls
Two specific capabilities within the PDAOS architecture deserve focused attention: consent receipts and selective disclosure.
A consent receipt, as defined in the Kantara Initiative's Consent Receipt Specification and further developed in PDAOS implementation, is a machine-readable record of a consent transaction that is returned to the data subject at the moment of consent. It is not a log entry inside a platform. It is a signed artifact the subject holds. When a regulator or auditor asks whether consent was obtained for a specific data use, the subject can produce that receipt as independently verifiable evidence.
Selective disclosure addresses a different problem: the fact that verifying identity or eligibility often requires exposing far more data than the verification actually requires. A system that needs to confirm a user is over 18 does not need their exact birthdate, their name or their address. It needs a cryptographic proof that an attribute is true.
Zero-knowledge proofs provide this capability. Using ZK constructions such as zk-SNARKs, a data subject can prove a claim about their data without revealing the underlying data to the verifier. PDAOS integrates ZK-based selective disclosure so that participation in AI training datasets, health research or financial services can be conditioned on verified attributes without those attributes being transmitted to the counterparty in raw form.
The NIST Privacy Framework and its companion guidance on privacy engineering explicitly identify data minimization and use limitation as core privacy outcomes. Selective disclosure is the technical mechanism that makes those outcomes achievable at runtime rather than aspirational at design time. NIST's Privacy Framework documentation is publicly available and worth reading alongside the PDAOS technical specification for practitioners aligning implementations with federal guidance.
A Technical Path Toward Data Sovereignty
The argument in The Invisible Data is not that data collection is inherently wrong. It is that collection without provenance, without subject-controlled consent artifacts and without attribution violates a basic principle of fairness that the data economy has never been forced to honor. AI training makes the violation more consequential because the data does not just inform a product decision. It becomes the product, encoded in weights that generate economic value indefinitely.
MyDataKey, the product companion to Own Your Data Inc's research mission, implements PDAOS primitives for individual users who want to begin exercising data sovereignty today. The underlying philosophy is documented in The Invisible Data and made technically actionable through the PDAOS specification that Own Your Data Inc continues to develop and publish as open research.
The path forward is not a patchwork of stronger privacy laws applied to infrastructure that was designed to evade them. It is infrastructure redesign: cryptographic provenance at origination, subject-controlled consent credentials, selective disclosure for minimum-necessary verification and attribution chains that travel with data through its full lifecycle, including inside AI training pipelines.
That redesign is technically achievable. The W3C standards, ZK proof systems and DID infrastructure to support it exist today. What has been missing is the architectural vision and the institutional commitment to build systems that treat the data subject as the origin point of the data supply chain rather than as an input to be processed.
That is the problem The Invisible Data names. PDAOS is the answer Own Your Data Inc is building.