TL;DR
This paper argues that blockchain technology is the missing architectural piece needed to achieve "Responsible AI." By shifting from centralized "Rule by Code" to decentralized "Rule of Code," we can use immutable ledgers and cryptographic proofs to bake transparency, accountability, and user rights directly into the AI lifecycle.
What Problem Does This Paper Address?
As AI models are integrated into critical sectors, their inherent risks—opacity, bias, unreliability, and the potential for manipulation—become existential threats. Current AI development is largely centralized, meaning a single entity controls the "code" and the data, leading to a "trust deficit."
The author, Aggelos Kiayias, addresses the challenge of building "Responsible AI" (RAI). He explores how to move beyond treating ethics as an afterthought and instead use decentralized infrastructure to ensure that AI systems are fair, transparent, and governed by the collective rather than a single corporation.
Methods
The paper is a conceptual framework and position piece that outlines a "Rights by Design" approach. Rather than proposing a single algorithm, it identifies four key use cases where blockchain primitives solve RAI challenges:
- Decentralized Identity (DID): Providing "Robot Identification" so humans know when they are interacting with AI, complete with a provable audit trail and "fingerprints" of specific model versions.
- Blockchain-Based Governance: Applying "digital constitutionalism" to AI, where stakeholders use immutable ledgers and private voting protocols to collectively decide how a model evolves.
- Privacy-Enhanced Data Mining: Using Secure Multiparty Computation (MPC) and Federated Learning with a blockchain registry to train models on private data without the data ever leaving the owner's control.
- Decentralized Machine Learning: Creating tokenized economies where computational power and data contributions are rewarded with digital currency, allowing for an AI lifecycle maintained without a central authority.
The paper also incorporates summaries of three specific technical papers that implement these ideas, covering blockchain-based AI auditing, tokenized licensing for creators, and new oracle protocols for verifying internet data.
Datasets
Not discussed in the paper. As a position paper and session overview, it focuses on architectural frameworks rather than training specific models on a particular dataset.
Key Insights and Ideas
- Rule by Code vs. Rule of Code: In centralized systems, the entity that writes the code holds all the power ("Rule by Code"). Blockchain enables the "Rule of Code," where the rules are transparent, immutable, and collectively enforced by a network.
- Rights by Design: Inspired by "Security by Design," this principle argues that fundamental rights (privacy, data portability, the right to be informed) must be the starting point of AI architecture, not a layer added later.
- Zero-Knowledge Proofs (ZKPs) for Fairness: The paper suggests using ZKPs to allow developers to prove their model is accurate or unbiased without having to reveal proprietary weights or sensitive training data.
- The "Neutral" Substrate: Blockchain can act as a neutral ground for AI, removing single points of failure and providing a "ledger of record" for every decision or inference an AI makes.
Key Results
While the paper is primarily theoretical, it highlights several qualitative "wins" for the decentralization approach:
- Auditability: The framework transforms AI from a "Black Box" to a "Glass Box" by recording data sources, model parameters, and algorithmic updates on a tamper-proof ledger.
- Creator Empowerment: Through the ORA (Ownership, Rights, Attribution) framework, creators can use tokenized licensing to set granular, machine-readable permissions for how GenAI models use their work.
- Fair Data Exchange: The paper highlights a new protocol ("encrypt-TLS-keys-to-future") that allows for the first decentralized fair exchange between Web2 data (like fiat currency or private records) and Web3 assets without needing a trusted intermediary.
Limitations and Future Work
The author and the featured papers acknowledge several hurdles:
- Technical Complexity: The barrier to entry for Distributed Ledger Technology (DLT) remains high, and the usability of these systems needs significant improvement.
- Identity Risks: While DIDs help with attribution, robust systems are still needed to prevent "false attribution" or malicious actors spoofing AI identities.
- Over-Restriction: There is a risk that bespoke, tokenized licenses could become so restrictive that they stifle innovation or derivative creative works.
- Interdisciplinary Gap: Future work must bridge the gap between technical protocols, legal governance, and market models to build a truly sustainable creative ecosystem in the age of AI.
Summary updated