Oxford’s AI Researcher Samuele Marro On Decentralized AI And Blockchain: When Integration Adds Value—But Limits Innovation

Decentralized AI projects are increasingly integrating blockchain infrastructure to access funding and ecosystem support, even when such integration may not be technically necessary. According to Samuele Marro, Head of the Institute for Decentralized AI and a DPhil student at Oxford University’s AIMS CDT, this trend raises an important question for builders and investors: does a blockchain-first approach strengthen decentralized AI, or does it risk constraining it?

In a conversation with MPost, Samuele Marro discussed when blockchain adds value to decentralized AI systems and when it may introduce additional cost and latency. He also addressed why incentive design can be more critical than default chain integration, and how selective tokenization can support—rather than distort—the development of decentralized AI networks.

How do you distinguish between “decentralized AI,” “crypto-integrated AI,” and “Web3 AI”?

Decentralized AI refers to any AI system where data, compute, or stakeholders are distributed. For example, free data learning counts as decentralized AI. Web3 AI also counts as decentralized AI, but different types of AI that the Web3 community would consider decentralized are actually centralized. Web3 AI is more about using cypher principles—strong commitments to anti-censorship, privacy, and resisting centralized control. Finally, crypto AI, or blockchain AI, is any project at the intersection of AI and blockchain. It can be centralized or decentralized, Web3 or not. Here, the emphasis is on technology.

Why do decentralized AI projects feel pressure to integrate blockchain?

The pressure comes from perception: people often equate decentralization and Web3 with blockchain. Projects feel they are not truly decentralized unless they issue a token or create a tokenized project. Sometimes this leads to building a new Layer 1 blockchain for tasks that could be handled with simpler distributed systems, like databases or peer-to-peer networks.

Nevertheless, people sometimes need blockchain integration in their projects. It enables transactions between entities without legal identities, such as AI agents. It also allows contracts to be enforced in a trusted manner and provides public verifiability. In general, it is one tool among many for enabling trust and coordination, but it is not always necessary.

Why does incentive design matter more than default blockchain integration?

Chain integration makes sense when a project wants access to an existing ecosystem, like Ethereum or Solana—that is why they choose them. Human participants tend to commit to one ecosystem, which creates network effects. However, AI systems can now manage interactions across ecosystems. Therefore, incentive design is often more important.

Can you share examples of incentive designs that successfully coordinated contributors or sustained funding for decentralized AI projects?

Bittensor illustrates this well. The protocol design is very good—for example, Yuma on Bittensor—their design encourages competition between subnets, allocating resources based on community-assessed contributions. This mechanism is decentralized yet flexible, allowing fine-tuning for specific use cases. Similar approaches apply to Torus and other projects that emerge from the same philosophy.

How can selective tokenization support decentralized AI networks?

Tokenization enables funding, which is crucial for large-scale AI projects requiring significant capital for pretraining or fine-tuning. Tokens allow these projects to be funded in a decentralized way.

At the same time, tokens enable a variety of incentive systems. You can experiment with these incentives to achieve the goals you want, for good or for bad.

What are the main risks when projects tokenize components of an AI stack, and how can those risks be mitigated?

Tokens tie a project’s success to the token’s market value. This can lead to prioritizing token price over the project’s long-term goals—features may be added to support token holders rather than improve the system.

This makes sense from a business perspective, but it can jeopardize the project if keeping the token price high becomes the primary goal at all costs. Clear incentive design and separating token utility from core project goals are necessary to mitigate these risks.

How should developers decide when blockchain integration is justified in an AI project?

A concrete example of when you definitely do not want blockchain is agent economies. These involve point-to-point interactions where one part of the network communicates with another. Using blockchain constrains the number of interactions due to bandwidth limits, which limits scalability.

Most blockchain use cases are about proving that something happened publicly—for example, sending a certain amount of USD. If you want private contracts or interactions where public visibility is not required, blockchain is often unsuitable. The strong incentive not to use blockchain in these cases is scalability.

No matter how well-designed a blockchain is, there is always a bandwidth limit. Increasing bandwidth too much reduces the number of participants who can contribute. On one side is the bandwidth constraint, on the other is the network. Tying your system to a blockchain forces you to fit as many interactions as possible inside a single channel. This is a losing game.

Anything more complex than contracts and payments, like dense agent economies, cannot rely on blockchain because it caps the size of your network.

What is needed to support decentralized AI projects that choose not to use blockchain?

There is a lot of “cargo cult thinking” in the Web3 ecosystem about what a project needs. The required technologies vary over time.

Culturally, there has been a feeling that if you do not integrate blockchain, you are not a real project. This is not top-down; it persists due to cultural inertia.

To incentivize participants, decentralized AI founders, community members, and researchers need to understand what actually makes a project work.

This understanding can develop naturally. For example, ERC-8004, an Ethereum standard for agent reputation and interaction, comes from the Web3 ecosystem but does not strictly require blockchain. Many AI researchers are reaching the same conclusion: much of the technology developed for decentralized AI does not require blockchain.

I imagine a scenario where initially, everyone believes blockchain is required, but then the community realizes scaling is better without it. The projects willing to invest in funding, research, building, and community awareness around non-blockchain solutions will likely succeed in this shift.

The infrastructure depends on the project’s needs but should support decentralized funding, research, and community engagement. Effective decentralized AI coordination can happen without blockchain, as standards like ERC-8004 for agent reputation demonstrate. Researchers increasingly recognize that much decentralized AI technology does not require blockchain. Projects that invest in building non-blockchain solutions may gain an advantage.

From your perspective, how would the future of the intersection of blockchain and decentralized AI evolve?

Even if some projects abandon blockchain, it will remain valuable for two main use cases: payments and smart contract enforcement. Payments are easy to implement on-chain, have been optimized by the community over a decade, and do not require legal entities—fitting any decentralized AI economy.

Smart contract enforcement allows agents, AI systems, mechanical systems, or humans to form contracts executed automatically, without lawyers or judges. This can scale significantly.

There is untapped potential for what an agent can do with another agent using blockchain as the execution environment. Low-cost, fully automated smart contracts that can be developed, deployed, and executed in minutes will be highly valuable for all types of decentralized AI systems.

The post Oxford’s AI Researcher Samuele Marro On Decentralized AI And Blockchain: When Integration Adds Value—But Limits Innovation appeared first on Metaverse Post.