Blockchain for AI

Over the last year, we have been seeing an explosive growth in the capabilities and applications of AI. This includes a dramatic increase in the maturity of text-to-image models and large language models and their applications to many business use cases from search and recommendations to facilitating software development and big data analytics. It has been predicted that Generative AI alone will be a $1.3 trillion market by 2032, with a Compound Annual Growth Rate (CAGR) of 42% over the coming decade.

When we look at the intersection of blockchain and AI, there are two main aspects to consider. The first is the use of AI/ML models and methods to enhance blockchain platforms, decentralized applications, and trading of digital on-chain assets. The second, which is the primary focus of this document, is to leverage blockchain capabilities to provide value to developers and users of AI/ML products and services.

Fundamental blockchain capabilities

There are a number of fundamental blockchain capabilities that can be leveraged for AI model development, deployment, and operation. These capabilities span cryptographic primitives, blockchain protocols, and smart contracts. They enable the following fundamental properties, that are all highly relevant for AI use-cases, as we discuss in the following sections:

• Data security: helping to store data in a tamper-proof and immutable manner, with high availability using decentralized servers that are less susceptible to attack, manipulation, censorship and denial of service.

• Data provenance, traceability, auditability: recording transactions and assets in an immutable and transparent manner, facilitating tracking of the origin, ownership, provenance of data and agreements with digital signatures and time-stamping. This also provides an auditable and verifiable trail. 

• Decentralized decision making: enabling decisions to be made by several entities or directly between two parties in settings where they do not have a prior existing trust-relationship among themselves or with a central entity.

• Autonomous and transparent code execution: enabling the execution of programs as smart contracts that are transparent to all concerned parties and operate autonomously without requiring trusted and centralized intermediaries. 

• Decentralized identities: providing mechanisms for secure digital identities that allow users to interact with services without compromising privacy.

• Micropayments: providing a secure and lightweight way to make payments with low friction.

1. How blockchain can benefit AI

As the above figure shows, the AI ecosystem has diverse stakeholders who interact with each other over data, models, as well as compute infrastructure. Due to the administrative and economic boundaries dividing potential stakeholders, these interactions have requirements of trust as well as payments — blockchain can help with both of these.

We can categorize the potential benefits of blockchain for AI and corresponding products and services that can be developed into four main classes, shown in the figure below. In the following, we discuss each of these in turn.

2.1. Data and model integrity: Blockchain can be used to develop solutions that help users and developers ensure that the data and models have not been modified without their knowledge. For example, an API-based service could allow data-owners and AI developers to record time-stamped hashes of datasets and models to ensure their integrity and log the entire process of model development and the datasets used to track the entire lifecycle, in a way that could be made available to third party auditors or regulators. The system could even be directly integrated into ML development tools such as Pytorch. This could help improve the integrity and trustworthiness of models, by making their development process more transparent and secure. It may also be possible to log relevant proofs of the “unlearning” of particular data from models onchain to demonstrate to the satisfaction of regulators that a certain provider's data have been removed from a given model. Logging hashes for data and model outputs onchain can also help to combat deep fakes - for example, applications may be able to ensure the authenticity of the data used by checking digital signatures associated with the source of the data onchain, or a decentralized version of “Snopes.com” could be designed and implemented onchain to flag deep fakes.

2.2. Data and model usage and access rights management: A non-fungible token can demonstrate one’s ownership of any given digital content or data. Depending on the use case the content in question could be a model input, such as a prompt to a Generative AI tool, it could be data used to train a model, it could be parameters of a model, or it could be the output of a model. The NFT would allow the user or developer to assert their ownership and further to transfer ownership of the corresponding digital asset to another. It is also possible to envision a blockchain-based access control mechanism for data and models, for example, a smart contract to allow/limit access based on a given list of user addresses. Or it could be integrated with a decentralized identity solution (possibly using state of the art cryptographic techniques such as zero-knowledge proofs) to allow access based on certain demonstrated attributes (e.g. allowing access based on demonstrating the user is old enough, or only from certain geographical locations) while preserving the privacy of the user.

2.3. Incentives and payments for data, models and compute resources: Blockchain can facilitate low-fee micropayments for use of a generative AI model using a stable coin. A smart contract could allow revenue sharing across multiple co-owners of a model in a decentralized fashion. Such co-ownership models, effectively a “decentralized Hugging Face”, may allow small-medium size model developers to join forces and compete against larger firms in the space. It could also be used to incentivize data providers, data annotators, model developers or human feedback providers anywhere in the world to join a new decentralized project to develop a new generative AI model or solution, with appropriate mechanisms to track contributions so that incentives can be fairly allocated. Blockchain could also be used to create a decentralized data/model/compute marketplace that makes it easy for compute providers, training data providers, model developers and users to search for and be matched to each other, to offer incentives, make payments, and enter into contracts and agreements. A blockchain-based decentralized review system implemented using smart contracts could incorporate both automated and human-based reviewers in a system that incentivizes high throughput, thorough, high quality review of data and models.

2.4. Deploying AI onchain: This category pertains to running certain AI models directly on the blockchain to benefit from greater transparency and trust. The AI models may be directly serving some inference or generative use-case for end users, giving them peace of mind that only the model they intended to provide the input to received said input and generated the output they are seeing without any manipulation or falsification or censorship. Or, the AI models may be deployed to help a smart contract adapt and optimize its own parameters in response to user transactions. The AI model may also be a smart contract that uses historical and current transaction data onchain to make buy/sell/trade decisions on digital assets to make a profit. These models may be deployed on layer 1 chains as smart contracts or via layer 2 systems such as zk-rollups. The models may be owned by private parties or they may be decentralized in the form of DAOs that allow multiple individuals and entities to own “shares” in a given onchain model. In the long run, for such applications, given the high data and compute requirements of AI applications, there may be interest to research and develop a new platform that is designed from the ground up to support AI workflows.

3. How Coinbase can help

Coinbase is on a mission to increase economic freedom for more than 1 billion people. As crypto’s uses grow, we’re focused on building the most trusted, compliant products and services, and supporting other builders.We can see blockchain for AI as building on this strategy pillar – enabling individuals and organizations that are part of the emerging generative AI ecosystem (that is today based almost entirely on centralized and less transparent Web2 frameworks) to gain benefits from blockchain and crypto-based onchain solutions, broadly defined.

Coinbase is particularly well-situated to be a leading contributor in this space because a) it has significant visibility and mindshare from both retail and institutional users of blockchain-based assets and services, b) a proven track record of helping to bridge the gap between the new emerging world of Web3 and the existing Web2 systems, and c) has a strong ML developer team within that has a good understanding of the growing generative AI ecosystem and the needs of developers and other stakeholders.

Coinbase is interested in exploring partnerships and integrations with like-minded companies with expertise in AI solutions and ecosystem services to bring some of these ideas to reality.

Acknowledgements

We would like to thank Dr. Bhaskar Krishnamachari of USC for their contributions to this piece. Dr. Krishnamachari is a paid consultant for Coinbase and has assisted on this article in this capacity.

--

Related work

• Coinbase Institute Whitepaper: Blockchain and Artificial Intelligence (AI): Complementary Technologies That Can Make Each Other Better, Fall 2023.

• David Duong, At the Intersection of AI and Crypto: What are the common issues in generative AI that could be addressed by blockchain technology?, Coinbase Research, May 2023.

• Use Cases of AI in Blockchain, Chainlink Blog, May 2023.

• Steve Vassallo, AI x Blockchain: The Next Level, Forbes Digital Assets, June 2023.

• Salah et al., Blockchain for AI: Review and Challenges, IEEE Access, 2019.

• Tian et al., Blockchain for AI: A Disruptive Integration, IEEE CSCWD 2022.

• Karger et al., Blockchain for AI Data – State of the Art and Open Research, ICIS 2021.