Attribution-Based Control: The Missing Element in AI

The Information Market Problem

Today’s AI ecosystem suffers from a fundamental market failure:

For Data Owners:

You can’t select specific uses for your data—only provide it entirely or withhold it completely
Once your data is used for AI training, you lose control over how it influences outputs
You have no mechanism to prevent privacy violations, copyright infringement, or IP theft
You face an economic bundling problem: sell all possible uses AND misuses—or none

For AI Users:

You can’t choose which data sources inform the AI predictions you receive
You have no way to verify if outputs come from trustworthy or accurate sources
You can’t distinguish between fact-based conclusions and AI hallucinations
You face the same bundling problem: trust all sources in the system—or none at all

Attribution-Based Control: The Core Solution

Attribution-Based Control (ABC) addresses this fundamental challenge by answering a critical question:

How can AI users choose which data sources they rely upon for each prediction, while data owners choose which AI predictions to support or abstain from supporting?

ABC provides:

For data owners: Granular control over how their data influences AI outputs
For AI users: The ability to select trusted sources for specific predictions
For society: A mechanism to align AI systems with human values and preferences

The Market Consequences of Missing ABC

Without Attribution-Based Control:

The 0.0001% Problem

Less than 0.0001% of the world's data is shared for AI training, and this small fraction suffers widespread misuse

The 99.999% Problem

Over 99.999% of valuable data remains inaccessible for AI development, preventing solutions to critical problems like early disease detection

The Geopolitical Challenge

Democratic nations face a dilemma—either accept rampant data misuse or fall behind command economies that can centralize data more effectively

Implementing Attribution-Based Control through Network-Source AI

To solve the ABC challenge, we’re developing Network-Source AI (NSAI)—a technical implementation that relies on two key technologies:

Model Partitioning

Divides AI model weights according to their training data sources — Implemented through Mixture of Experts, Model Ensembling, RAG, RETRO/ATLAS, and Model Merging

Structured Transparency

Enables data owners to collaborate without exposing raw inputs — Powered by cryptography and distributed systems technologies

This approach creates an infrastructure where:

Every prediction can be traced to its contributing sources
Data owners maintain control over how their data influences outputs
AI users can select trusted sources for specific predictions

Syft: Building ABC into the Information Infrastructure

To bring Attribution-Based Control to life, we’re creating Syft—an open, public network for non-public information:

For data owners

Offer predictive capabilities via API without surrendering control of data or weights

For AI users

Select and combine predictions from specific trusted sources

For both

Leverage cryptographic tools to maintain privacy and security while enabling collaboration

Syft represents a new paradigm—a public network designed specifically for non-public information, with attribution at its core.

OpenMined: Implementing Attribution-Based Control in Global Information Markets

OpenMined is a global non-profit dedicated to solving the Attribution-Based Control challenge in information markets. We believe that by enabling source-specific consent and selection:

Data owners gain control over how their information is used
AI users gain transparency into the sources behind predictions
The world unlocks the potential of the 99.999% of data currently unavailable for AI

Our work spans both quantitative data (“remote data science”) and qualitative data (“network-source AI”), already unlocking insights at organizations like Microsoft’s LinkedIn, X/Twitter, Reddit, the US Census Bureau, the UN Statistics Division, and the UK’s AI Security Institute.

Stay Informed on Attribution-Based Control & Network-Source AI

Join our newsletter to receive updates on our progress implementing Attribution-Based Control, upcoming events, and opportunities to help build a more accountable AI future.