Most AI Tokens Are Not Decentralized AI, Research Finds

Decentralized AI promises something profound.

It imagines a future where intelligence is not controlled by a handful of technology firms, where users retain ownership over data and models, and where AI systems operate as shared public infrastructure rather than proprietary assets.

Over the past few years, this vision has been bundled into a fast-growing category of crypto assets known as AI-based tokens..

Momentum accelerated after the release of ChatGPT in late 2022. A familiar narrative took hold: crypto would decentralize AI in the same way it was supposed to decentralize finance.

The research paper AI-Based Crypto Tokens: The Illusion of Decentralized AI? asks the necessary (and uncomfortable) question beneath that narrative:

How much of this decentralization is real, and how much exists only in name?

Decentralized in branding, centralized in practice

At first glance, AI-token projects appear decentralized. They use blockchains. They issue governance tokens. They advertise open participation and community ownership.

The paper shows that this surface-level decentralization masks a deeply centralized reality.

Across nearly all major projects, the most important computational tasks—model training, inference, and data hosting—occur off-chain on servers controlled by a small group of actors. The blockchain functions primarily as a coordination and payment layer, not as the execution environment for intelligence itself.

This architectural choice is understandable. Modern AI workloads are expensive, data-intensive, and poorly suited to on-chain execution. But it creates a fundamental contradiction: the trustless guarantees of blockchain end precisely where AI computation begins.

Why AI tokens fail to outperform centralized AI

The research systematically compares AI-token platforms with centralized AI services such as cloud APIs and model marketplaces.

The conclusion is not subtle.

Centralized providers outperform AI-token platforms on nearly every practical dimension: performance, cost, reliability, and user experience. They scale quickly, deploy massive GPU clusters, and offer simple interfaces. AI-token platforms introduce additional friction, wallets, tokens, governance votes, and settlement delays, without delivering commensurate benefits.

In many cases, the underlying business model is nearly identical to centralized AI:

• credit cards are replaced with tokens

• APIs are wrapped in smart contracts

• infrastructure control remains concentrated

The result is not decentralization, but indirection.

The result is indirection.

The Core Misconception: Tokens ≠ Decentralization

The paper’s most important insight is conceptual.

Tokenization and decentralization are not the same thing.

The paper calls this the illusion of decentralized AI.

The illusion persists because decentralization is often inferred from design elements rather than measured in practice. The relevant questions are operational:

• Who runs the models?

• Who controls updates?

• Who can meaningfully influence outcomes?

In most AI-token systems, the answers still point to a small core team.

This mirrors a broader Web3 pattern: decentralization assumed by architecture, undermined by incentives, coordination costs, and technical constraints.

What the Research Actually Finds

1. Decentralization Rarely Extends to the AI Core

Across nearly all surveyed projects, decentralization stops at coordination.

The paper finds that:

• Training is centralized

• Inference is centralized

• Data pipelines are centrally curated

Blockchains are used primarily for:

• payments

• access control

• token distribution

In other words, the blockchain coordinates economic activity, but does not execute intelligence.

The authors describe this pattern precisely as “coordination decentralization without execution decentralization.” The system appears decentralized at the surface level, while the computational core remains tightly controlled.

This distinction is critical. Control over training, inference, and data determines what models can do, how they evolve, and whose interests they serve. When those functions remain centralized, decentralization becomes cosmetic rather than structural.

2. Governance Tokens Do Not Alter Control Paths

The paper closely examines on-chain governance across AI-token platforms, with a focus on what token holders can actually influence.

The findings are consistent:

• Voting rarely affects core model decisions

• Infrastructure upgrades remain team-driven

• Token holders influence parameters, not architecture

Governance mechanisms tend to operate at the margins—adjusting fees, rewards, or access rules—while decisions about model design, training regimes, and deployment pipelines remain centralized.

As a result, governance tokens function less as instruments of control and more as:

• signaling tools

• fundraising instruments

They convey participation without conferring authority.

The paper is explicit on this point: decentralized governance cannot compensate for centralized execution. When the technical core is not governable, voting becomes symbolic.

3. Verification Is the Missing Primitive

The most repeated technical conclusion in the paper is stark:

Without verifiable computation, decentralization collapses at the boundary between blockchain and AI.

Blockchains cannot currently verify:

• whether inference was executed correctly

• whether training followed protocol

• whether datasets were manipulated

Because these steps occur off-chain, trust is reintroduced precisely where decentralization claims are strongest.

Some projects attempt mitigation through staking, reputation systems, or peer evaluation. The paper acknowledges these efforts but finds them insufficient. They reduce risk at the margins without eliminating the trust gap.

For the authors, this is the central unsolved problem. Until computation itself becomes verifiable, decentralized AI remains structurally incomplete.

4. Economic Incentives Favor Centralization

Even in cases where decentralization is technically feasible, the paper finds that economic incentives consistently push systems back toward centralization.

Documented advantages of centralized infrastructure include:

• cheaper GPU access through scale

• lower latency via centralized coordination

• faster iteration and better user experience

Over time, these pressures lead teams to recentralize rationally rather than ideologically.

The paper explicitly connects this dynamic to earlier Web3 patterns:

• DeFi protocols with centralized frontends

• Layer-2 networks with single sequencers

• Cross-chain bridges with trusted operators

In each case, decentralization erodes unless actively defended against economic gravity.

AI-token systems follow the same trajectory.

5. Speculation Dominates Measured Usage

Finally, the paper examines usage data rather than narratives.

Across AI-token ecosystems, empirical indicators show:

• low transaction-to-market-cap ratios

• limited recurring usage

• shallow developer adoption

Token velocity is driven primarily by trading, not by demand for AI computation.

The authors are careful not to frame this as a moral failure. Instead, they describe it as a structural outcome. When tokens are freely tradable and product-market fit is weak, speculative activity overwhelms functional use.

AI tokens, the paper argues, have largely replicated this pattern.

Examples of Decentralized AI

Bittensor

Bittensor attempts decentralization at the model contribution and incentive layer.

• Participants contribute models rather than raw compute

• Models are scored by peers based on usefulness

• Rewards are allocated algorithmically, not by a central operator

What’s still centralized

• Core protocol development

• Network parameters

• Heavy reliance on off-chain execution

Why it matters
Bittensor shows that model-level decentralization is possible, even if infrastructure-level decentralization is not yet solved.

Gensyn

Gensyn targets the compute verification problem directly.

• Focuses on decentralized ML training

• Uses cryptographic techniques to verify off-chain computation

• Explicitly acknowledges blockchains cannot run AI directly

What’s still missing

• Production-scale adoption

• End-user applications

• Fully trustless verification at scale

Why it matters
The paper highlights Gensyn as an example of research-aligned design: it tackles the actual bottleneck, not the narrative layer.

Fetch.ai

Fetch.ai focuses on agent coordination, not foundation models.

• Autonomous agents interact on-chain

• AI logic often runs locally or off-chain

• Blockchain handles discovery, identity, and settlement

Trade-off

• Intelligence is distributed

• Training and inference are not

Why it matters
The paper categorizes this as partial decentralization: useful, but not equivalent to decentralized AI infrastructure.

Off-Chain Computation Is the Real Bottleneck

One finding appears repeatedly throughout the research: heavy reliance on off-chain computation.

Blockchains cannot efficiently run modern AI models. As a result, decentralized AI platforms depend on:

• external servers

• GPU providers

• specialized compute nodes

The blockchain records coordination and payments, but cannot verify whether computation was executed correctly.

This creates a structural trust gap.

Users must trust off-chain providers to behave honestly. Some projects attempt mitigation through staking, reputation systems, or peer evaluation. These mechanisms help at the margins, but they remain incomplete.

Until AI computation becomes verifiable, decentralization remains partial by design.

Speculation overwhelms utility, again

The paper places AI tokens within a familiar historical pattern.

Across multiple crypto cycles, utility tokens have often failed to achieve sustained usage even as speculation flourished. AI tokens follow this trajectory closely. Prices respond strongly to AI narratives, while actual platform usage remains limited.

In many ecosystems, token trading dominates token utility.

This outcome is not necessarily driven by bad faith. It is structural. When tokens are freely tradable, financial incentives often overwhelm functional ones. Without strong product-market fit, tokens become speculative assets first and coordination tools second.

The research suggests most AI-token ecosystems have not yet crossed that threshold.

What real decentralized AI would require

The paper does not reject decentralized AI outright. Instead, it outlines the conditions under which it could become real.

Promising directions include verifiable off-chain computation using zero-knowledge proofs, trusted execution environments, or AI oracles that allow blockchains to verify AI outputs. Federated learning coordinated on-chain could enable collaborative model training without centralized data control. Modular blockchain architectures could support AI-specific execution layers rather than forcing AI onto general-purpose chains.

Most importantly, decentralized AI must offer capabilities centralized AI cannot—or will not—provide. Privacy-preserving training on sensitive data. Collective ownership of foundational models. Censorship-resistant deployment.

Without such differentiation, decentralization remains ideological rather than practical.

The deeper structural lesson

The broader lesson of this research extends beyond AI tokens.

Decentralization is not a branding choice. It is an emergent property of architecture, incentives, and governance. When systems rely on centralized computation, centralized coordination, and centralized upgrades, tokens alone cannot make them decentralized.

This explains why many AI-token projects feel simultaneously ambitious and underwhelming. The vision points forward. The infrastructure pulls backward.

Conclusion: decentralized AI is still research, not product

The paper’s conclusion is measured but firm.

AI-based crypto tokens today are better understood as experiments rather than solutions.

They explore important ideas and surface real constraints, but they have not yet delivered decentralized AI at scale.

Decentralized AI needs:

• verifiable off-chain computation

• cryptographic proofs of inference

• incentive-compatible data contribution

• governance that reaches execution, not just policy

None of these are solved problems.

Decentralized AI will not emerge by “adding crypto” to AI.

It will emerge when AI systems are redesigned around verification, coordination, and collective ownership from first principles.