The modern internet is built on a structural asymmetry: individuals generate data, but institutions own, store, and monetize it. Personal data is fragmented across corporate silos, accessed through brittle APIs, duplicated endlessly, and governed by policies users neither see nor control.
This document outlines an alternative architecture: personal data as a first鈥慶lass, user鈥慶ontrolled system, accessed peer鈥憈o鈥憄eer and mediated by AI rather than centralized platforms. The goal is not merely better privacy, but a simpler, more composable, and more human鈥慳ligned data ecosystem.
At its core, the system treats personal data the way we treat people socially: you do not hand over your entire life to every party you interact with. Instead, you selectively reveal insight, summaries, or claims, often mediated by trust, context, and interpretation.
- Every individual controls their own data store.
- Data may be self鈥慼osted or hosted by a third鈥憄arty provider acting as a custodian, not an owner.
- Custodians are interchangeable; portability is a baseline requirement.
- Data access is fundamentally connection鈥慴ased, not globally addressable.
- To learn about someone, you must be connected to them (directly or via consented mediation).
- There is no universal database to query; there are only peers.
- Raw data is rarely shared.
- Insights are computed in situ via AI mediators.
- The requesting party receives answers, summaries, or proofs鈥攏ot the underlying records.
- Computation moves to the data, not the other way around.
- Queries are evaluated under strict disclosure constraints.
- The system generalizes ideas from privacy computing, zero鈥慿nowledge proofs, and secure enclaves, but prioritizes practical expressiveness over academic purity.
- All personal data conforms to a shared global schema.
- Each individual鈥檚 database is separate, private, and sovereign.
- Uniform shape enables universal tooling without centralization.
Each individual operates (or delegates) a Personal Data Node:
- Encrypted storage of all personal data
- Schema鈥憊alidated data model
- Policy engine for access control
- AI mediator runtime
- Network interface for peer connections
The PDN is the unit of identity in the system.
AI mediators act as interpreters between raw data and external queries.
Responsibilities:
- Interpret incoming requests
- Apply disclosure policies
- Generate bounded responses (summaries, scores, explanations)
- Prevent data exfiltration through inference attacks
Crucially, mediators are not neutral鈥攖hey are explicitly aligned to the data owner鈥檚 interests.
- Connections are explicit, consented, and revocable.
- Institutions (schools, banks, employers) are peers, not superusers.
- Trust is contextual, not absolute.
This mirrors social graphs more than client鈥憇erver models.
Requests fall into three broad categories:
-
Direct Disclosure
- Explicitly shared fields (e.g., name, email)
-
Derived Insight
- AI鈥慻enerated summaries or classifications
-
Verified Claims
- Assertions with proofs (e.g., "income above X", "credit risk below Y")
The system intentionally mimics how information spreads among humans:
- You may know that something is true without knowing why
- You may receive a broad outline without intimate details
- Confidence levels and uncertainty are explicit
This is a feature, not a limitation.
Instead of uploading bank statements, tax returns, and transaction histories:
-
The lender submits a query: "Assess creditworthiness under policy P"
-
The AI mediator evaluates local data
-
The lender receives:
- A risk score
- Supporting explanations
- Optional cryptographic attestations
Raw financial data never leaves the PDN.
Personal data is surprisingly uniform:
- Identity
- Relationships
- Finances
- Health
- Education
- Activity logs
Standardizing shape enables:
- Universal AI models
- Reusable queries
- Shared tooling
- Lower cognitive overhead
- Extensible but opinionated
- Versioned and backward鈥慶ompatible
- Semantically rich, not just syntactic
- Shape is global and shared
- Content is local and private
This is the key inversion compared to today鈥檚 platforms.
- Data lives in corporate silos
- Access via proprietary APIs
- Users grant blanket permissions
- Data is copied, cached, and resold
- Revocation is largely illusory
- Data lives with the individual
- Access via peer connections
- Fine鈥慻rained, query鈥憀evel disclosure
- No bulk data transfer
- Revocation is immediate and real
Today:
- Simple for institutions
- Complex and opaque for users
Proposed:
- Slightly more complex for institutions
- Radically simpler and clearer for users
- Institutions are curious but rational
- Attackers may attempt inference, correlation, or replay attacks
- AI models are fallible and must be constrained
- Query rate limiting
- Differential disclosure
- Output entropy controls
- Auditable mediator decisions
- Cryptographic attestations where needed
Trust is earned per interaction, not granted once.
- Standardization governance for the global schema
- Marketplaces for third鈥憄arty mediators
- Liability models for incorrect AI disclosures
- UX for explaining mediated answers to humans
- Interoperability with legacy systems
This system reframes personal data as something closer to personhood than property. It replaces bulk transfer with conversation, APIs with interpretation, and centralized trust with peer relationships.
The ambition is not to eliminate institutions or analytics, but to realign power: insight without exposure, access without ownership, and intelligence without surrender.
If successful, the internet becomes less like a set of data warehouses鈥攁nd more like a network of people again.