Toward Decentralized Intelligence: A Systematic Literature Review of Blockchain-Enabled AI Systems

Abstract

The convergence of decentralized artificial intelligence (DAI), blockchain technology, and smart contracts is reshaping the design and governance of intelligent systems. As these technologies rapidly evolve, addressing privacy within their architecture, usage models, and associated risks has become increasingly critical. This systematic literature review examines architectural patterns, governance frameworks, real-world applications, and persistent challenges in DAI systems. It identifies prevailing designs such as federated learning integrated with consensus protocols, smart contract-based incentive mechanisms, and decentralized verification methods. Drawing from a diverse body of recent literature, the review highlights implementations across sectors, including healthcare, finance, IoT, autonomous systems, and intelligent infrastructure, each demonstrating significant contributions to privacy, security, and collaborative innovation. Despite these advancements, DAI systems face ongoing obstacles such as scalability limitations, privacy trade-offs, and difficulties with regulatory compliance. The review emphasizes the need for integrative governance approaches that balance transparency, accountability, incentive alignment, and ethical oversight. These elements are proposed as co-evolving pillars essential to establishing trustworthiness in decentralized AI ecosystems. This work offers a comprehensive review for understanding the current landscape and guiding the development of responsible and effective DAI systems in the Web3 era.

1. Introduction

Decentralized Artificial Intelligence (DAI) refers to the development and deployment of AI models on decentralized technologies, such as blockchain, to mitigate risks inherent to centralized systems. These risks include a lack of data transparency, the emergence of data monopolies, and vulnerability to single points of failure [1,2]. As a rapidly emerging paradigm, DAI combines distributed machine learning, multi-agent systems, edge intelligence, and decentralized governance to enable AI systems that operate without centralized control.

AI systems are increasingly pervasive in critical sectors such as healthcare, finance, and urban infrastructure. In healthcare, for example, AI has been employed to enhance disease prediction capabilities and to manage health records and medical imagery [3,4]. However, centralized medical AI systems often suffer from inadequate data transparency, raising concerns among individuals about how and where their data is being utilized and shared [5]. This lack of transparency may erode public trust and heighten fears regarding data misuse [5]. These challenges highlight the urgent need for a more trustless, transparent, and collaborative AI ecosystem. This literature review aims to address the identified problem by systematically examining technical gaps, unresolved challenges, and emerging trends. In doing so, it seeks to provide a comprehensive research roadmap intended to guide future efforts toward the development of scalable, trustworthy, and autonomous AI infrastructures.

DAI models have emerged as a promising paradigm to address these concerns by storing data locally and training models at the source, while using distributed ledger technologies (DLT) such as blockchain and smart contracts to securely store, automate, and validate the models. The immutable and transparent nature of blockchain facilitates verifiability of data usage and access, thereby enhancing trust [6,7,8]. If healthcare institutions each deployed localized AI models and interconnected them via a shared blockchain, the collective system could benefit from cumulative learning across institutions without compromising patient privacy, as only the models, not the underlying data, would be exchanged [5,6,9]. Furthermore, centralized AI systems rely on a single coordinating entity to manage the learning process, thereby introducing a single point of failure. This issue can be effectively addressed through decentralized infrastructures supported by blockchain, which eliminate reliance on any one system [1,10]. This review systematically examines the existing literature on DAI systems, with particular emphasis on the role of blockchain and smart contracts in enabling privacy-preserving, scalable, and economically viable DAI frameworks.

The remainder of this paper is structured as follows. Section 2 provides an overview of key concepts in DAI systems and discusses the various implementations of DAI, highlighting how these models differ from other approaches and identifying the domains in which they are particularly applicable. Section 3 outlines the application of the PRISMA methodology used to synthesize peer-reviewed articles, as well as the core research questions that guided the literature review. Section 4 addresses the first research question by presenting a taxonomy of emerging trends, architectural patterns, and commonly adopted technologies, along with the rationale for their use. Section 5 responds to the second research question by examining how smart contracts facilitate the implementation of incentive mechanisms and governance structures within DAI systems, as well as the associated challenges. Section 6 discusses the third research question by exploring the various applications of DAI and the specific use cases identified in the literature. Section 7 engages with the fourth research question, focusing on unresolved challenges, existing technical gaps, and the ethical and legal issues inherent in decentralized models. Section 8 discusses the limitations of the study, and Section 9 outlines potential future research directions that warrant further investigation. Finally, Section 10 presents the conclusion.

2. Background and Related Work

Recent advancements in decentralized technologies have shifted how AI systems are built, governed, and deployed. This section provides an overview of key concepts and prior research that inform the development of decentralized AI architectures, including blockchain and smart contracts, federated learning, multi-agent coordination, tokenized marketplaces, and on-chain governance. Understanding these components is essential for situating our work within the broader landscape of decentralized intelligent systems.

2.1. Blockchain and Smart Contracts in Decentralized AI

As security and privacy become major concerns in this new era of rapid technological evolution, centralized approaches pose serious risks of privacy exposure and sensitive data leakage [11,12]. These vulnerabilities have pushed researchers and developers to rethink the foundation of intelligent systems, leading to a growing shift toward decentralized architectures. Decentralization relies on a distributed network of nodes, where each participant holds a complete copy of the ledger. This setup removes the dependence on a central trusted authority, enhancing system security, improving resilience, and eliminating single points of failure [13,14]. A Blockchain is a shared, tamper-proof ledger that lets multiple parties securely record and verify transactions without a central authority [14].

2.2. Federated and Distributed Learning Approaches

It is important to look at different approaches used in centralized and decentralized learning to perform machine learning without exposing raw data and to protect user privacy. Federated Learning (FL) is a privacy-preserving method where participants train a shared model by sharing updates, not raw data, keeping data decentralized [15]. In decentralized AI, FL is integrated with blockchain to ensure traceability, verifiability, and fairness in training processes. Smart contracts, which a self-executing programs stored on a blockchain, may be used to coordinate training rounds, aggregate gradients, and distribute rewards or penalties [11,14]. The role that smart contracts play in federated learning will be discussed more in depth in Section 5.1. Additionally, hybrid architectures such as off-chain training with on-chain verification strike a balance between efficiency and trust [13]. In more dynamic environments, peer-to-peer (P2P) distributed learning approaches are emerging, supporting fluid topologies and localized adaptation without the need for central parameter servers [14]. These methods are critical in domains like healthcare and autonomous mobility, where data sensitivity and low-latency decisions are paramount [16].

2.3. Multi-Agent Systems and Decentralized Coordination

Multi-Agent Systems (MAS) enable decentralized coordination by allowing multiple intelligent agents to operate autonomously while collaborating toward shared objectives. These systems often leverage DLT, including blockchain and smart contracts, to coordinate economic decisions and enhance trust. A notable example is the use of Autonomous Economic Agents (AEAs), which interact via peer-to-peer protocols and operate without centralized control [17]. In decentralized federated learning (DFL), MAS principles help coordinate learning without central servers, using peer-to-peer or gossip-based communication. This approach enhances scalability, fault tolerance, and privacy [18]. Applications of MAS include decentralized marketplaces, computation offloading, energy management, and decentralized autonomous organization (DAO) governance, where smart contracts facilitate coordination and decision-making [19,20]. Privacy-preserving techniques like Secure Multiparty Computation (SMPC) are integrated to ensure secure coordination without exposing individual data [21].

2.4. Token-Based AI Marketplaces

Token-based AI marketplaces represent a growing paradigm where data, models, and computing resources are treated as digital assets that can be exchanged, rented, or licensed [19]. Here, “token-based” refers to the use of cryptographic digital tokens; typically implemented on a blockchain as units of value, access, or governance within the platform. These platforms aim to democratize access to AI by enabling stakeholders to contribute and monetize resources within a decentralized economic framework [22]. Tokens serve not only as units of exchange but also as tools for incentivization, governance, and quality control. Smart contracts in such marketplaces often implement reward distribution, stake-based validation, and dispute resolution [15,23]. Participants are typically incentivized to act honestly through mechanisms like reputation scores, token staking, or performance-based payouts. Examples include platforms where individuals can contribute private datasets for model training, or edge nodes can offer surplus computing power in return for tokens. The integration of blockchain ensures transaction traceability, while federated learning preserves data privacy [16]. These systems are increasingly being explored for applications in personalized medicine, decentralized finance (DeFi), and smart infrastructure [24].

2.5. On-Chain Governance in AI Systems

Emerging implementations also explore hybrid governance models, where on-chain rules trigger off-chain computations or moderation processes, striking a balance between efficiency and transparency [1,7]. These models frequently utilize smart contracts to automate key decisions on-chain, while relying on oracles to incorporate off-chain data or initiate external actions such as moderation or analysis by human participants [25]. However, token-based voting structures, particularly the “one-token, one-vote” approach, concentrate decision-making power among large stakeholders, creating governance issues like ‘whale dominance’ or collusion [26]. Additionally, smart contracts enhance trust and tamper-resistance but are costly and complex to modify, limiting adaptability for evolving AI governance [27,28]. While on-chain governance enhances trust and autonomy, it introduces challenges related to voter apathy, majority capture, and rigidity in evolving AI workflows [29,30]. Nevertheless, as AI systems increasingly operate in open, distributed environments, on-chain governance is becoming a foundational pillar for ensuring equitable and secure participation [31].

3. Methodology

This section outlines the systematic approach undertaken to investigate the evolving landscape of DAI, with particular attention to blockchain integration, smart contract governance, and privacy-preserving coordination. Following the PRISMA 2020 guidelines [32,33], the methodology is structured around clearly defined research questions, a rigorous search and screening strategy, and a systematic framework for data extraction and synthesis. The objective is to ensure transparency, reproducibility, and analytical depth in evaluating the state of the field and identifying emergent patterns, gaps, and challenges. The Preferred Reporting Items for Systematic reviews and Meta-Analyses extension for Scoping Reviews (PRISMA-ScR) Checklist is provided in the Supplementary Materials.

3.1. Research Goal and Questions

The primary objective of this review is to synthesize the current state of DAI, with a focus on blockchain-enabled architectures, smart contract-based governance, and privacy-preserving coordination mechanisms. As this work is designed as a scoping review, the focus is on systematically mapping and synthesizing existing literature rather than presenting new experimental data or empirical testing. This approach ensures transparency and breadth of coverage while highlighting current research directions and knowledge gaps. The paper examines how decentralized AI is designed, governed, and deployed across real-world domains, identifying key innovations and persistent challenges in the field. The study is guided by four research questions:

•

RQ1: What are the dominant architectural patterns and technologies in DAI leveraging blockchain and smart contracts?

•

RQ2: How are incentive mechanisms and governance achieved in DAI systems using blockchain?

•

RQ3: In which real-world domains are blockchain-enabled DAI systems implemented, and what are their impacts?

•

RQ4: What technical, organizational, and ethical challenges remain in blockchain-enabled DAI, and how are they addressed?

3.2. Literature Search and Screening Strategy

A comprehensive literature search was conducted across five major scholarly databases, including ACM Digital Library, IEEE Xplore, SpringerLink, Scopus, and Web of Science, to capture the breadth of research at the intersection of DAI, blockchain, and smart contracts. The review focused on publications from 2016 to 2025, a period that reflects the rise of blockchain-enabled decentralized intelligence. A Boolean Search query was applied using the keywords mentioned in the paper.

The inclusion criteria for the paper selection were as follows:

• Peer-reviewed articles proposing or analyzing decentralized AI architectures.
• Explicit use of blockchain, smart contracts, or decentralized governance in AI systems.
• Real-world implementation or domain-specific case studies.

Exclusion Criteria included:

• Studies focused exclusively on centralized AI systems.
• Theoretical frameworks that did not implement blockchain-based decentralization or any decentralized approach.
• Non-peer-reviewed sources, such as non-peer-reviewed material, blogs, or preprints.
• Any duplication of the sources.

3.3. Data Extraction, Categorization, and Synthesis

To facilitate organization and synthesis, each article was reviewed and categorized based on key thematic attributes extracted from the abstract and full text. These attributes included the type of decentralization (data, computation, or governance), computation architecture (on-chain, off-chain, or hybrid), blockchain platform used (e.g., Ethereum, Hyperledger), learning algorithm (e.g., federated learning, swarm intelligence), and the role of smart contracts in system coordination. Articles were further grouped based on incentive mechanisms (e.g., token-based, reputation-based), governance models (centralized, DAO, hybrid), and privacy-preserving techniques. For example, differential privacy adds statistical ‘noise’ to datasets or model updates to protect sensitive individual information while still allowing useful analysis, whereas zero-knowledge proofs are cryptographic methods that allow one party to prove possession of information (such as a model’s validity) without revealing the underlying data itself. Application domains such as healthcare, finance, smart infrastructure, and autonomous mobility were also identified. For each study, real-world implementation status was noted, as well as the presence of key challenges, including incentive design issues, scalability constraints, privacy and security risks, governance complexities, interoperability barriers, and legal or ethical concerns. The extracted data were synthesized to reveal common architectural patterns and recurring governance models. A thematic analysis helped group technologies, identify research clusters, and highlight underexplored areas. Additionally, co-citation analysis was employed to uncover scholarly linkages and trace the intellectual development of the field. These analytical strategies provided a comprehensive foundation for discussing the current state of decentralized AI systems and identifying future research directions.

3.4. Screening Process

The initial search yielded 2702 records. After deduplication, 2321 unique records remained. During the first screening phase, papers were filtered based on title relevance, resulting in the retention of 900 articles. In the second phase, abstract screening was conducted, narrowing the pool to 459 articles. Inter-rater agreement during title and abstract screening was assessed using Cohen’s Kappa statistic on a random subset of 50 records, yielding a score of 0.70, which was deemed acceptable. Discrepancies were resolved through discussion. Following full-text review, 288 articles were deemed eligible for further consideration. Of the total corpus reviewed, 92 articles were selected for detailed analysis based on their relevance to the study’s thematic focus and the level of the analytical depth required. While the remaining 196 articles met the initial inclusion criteria, they were not included in the final synthesis due to limited methodological detail or to preserve a focused and coherent analytical scope. A comprehensive list of all 288 reviewed articles, including those not included in the final synthesis, is available in the publicly accessible GitHub repositoryat commit a85802e7a5a8d0d768c7cdaf3b2a196f8eaebe54 on the main branch [34]. Figure 1 presents a visual demonstration of the article identification, screening, exclusion, and inclusion process for our final review.

4. Thematic Analysis of Decentralized AI System

This section addresses RQ1 by synthesizing the architectural and technological patterns observed across DAI systems that incorporate blockchain and smart contracts. The findings are organized under three key themes: system dimensions, technology stack, and use cases. These themes reflect common design approaches, coordination architectures, and implementation contexts within the reviewed literature.

4.1. System Dimensions

DAI systems exhibit diverse architectural properties that influence how they manage computation, ownership of data, and governance. These foundation aspects are collectively referred to as system dimensions that define the structural backbone of a DAI framework. This section categorizes and analyzes these dimensions into three subthemes: computation architecture, data ownership models, and governance structure.

4.1.1. Computation Architecture

DAI systems employ various computational designs, which determine how and where model training, aggregation, and decision-making occur. The primary computation architectures identified in the review are on-chain, off-chain, and edge AI approaches. On-chain architecture executes computation and coordination directly on the blockchain, ensuring transparency and immutability [35,36,37]. However, they are limited by performance constraints and transaction costs [7,38]. Several studies adopt on-chain learning or verification to enhance trustworthiness in sensitive applications [39,40]. Whereas off-chain computation shifts heavy tasks like training, optimization, and aggregation off the blockchain, improving efficiency while using on-chain anchors for auditability [30,41]. The other approach is that edge AI architectures extend learning to decentralized edge devices, reducing latency and preserving data locality. This architecture is especially relevant in the Internet of Things (IoT) and autonomous systems, where edge participation is crucial for scalability and responsiveness [5,42]. In practice, hybrid computation architectures combining off-chain training and on-chain verification or coordination are the most prevalent [39,43]. These systems strike a balance between trust, performance, and scalability [7,41,42].

4.1.2. Data Ownership

DAI systems adopt diverse data ownership models depending on privacy goals and trust assumptions. Three dominant ownership models are observed: federated, peer-to-peer (P2P), and shared data pools. Federated ownership remains the most commonly adopted model, enabling clients to retain local data while only sharing model updates. It is predominant in privacy-sensitive domains [1,36,44]. In contrast, peer-to-peer (P2P) models emphasize fully decentralized control, where clients collaboratively participate without centralized intermediaries [38]. Shared pool models, with data contributed to a common repository, are rare but used in exploratory collaborative training systems [36,45].

4.1.3. Governance Model

Governance in DAI systems defines how decisions are made, rules are enforced, and conflicts are resolved. The three main types observed are centralized governance, DAO-based governance, and hybrid models. DAO-based governance is increasingly popular due to its ability to encode policies as immutable smart contracts [46]. Systems implement DAOs for managing participation, reward distribution, and dispute resolution [45]. Centralized governance is still used in several research prototypes or semi-decentralized implementations, where a central coordinator or admin node retains veto power or initiates aggregation [1,47]. However, hybrid governance attempts to merge the strengths of centralized oversight with decentralized consensus. For example, the governance layer where key parameters are determined collectively by participants, but fallback control resides with a regulatory node for legal compliance [48].

4.2. Technolgy Stack

The foundation technology stack of DAI systems integrates a diverse array of blockchain platforms, smart contract functionalities, machine learning protocols, and privacy-preserving mechanisms. This stack underpins the technical architecture and coordination mechanisms across various domains.

4.2.1. Blockchain Platforms

Ethereum is the most adopted platform, valued for its mature ecosystem and smart contract capabilities, with over 15 articles citing its use in DAI. It supports programmable logic, essential for deploying decentralized AI services [38,46]. Hyperledger is also employed in privacy-sensitive contexts, especially in permissioned settings where organizational control is required [3,49]. Other platforms like Polkadot and Cosmos appear in designs aiming for cross-chain interoperability and scalability [7,50].

4.2.2. Smart Contract Functionalities

Smart contracts in DAI systems are leveraged for a range of coordination functions [1,51]. Incentive logic is the most prevalent use case, enabling reward distribution based on contribution [52]. Model verification and aggregation mechanisms are also embedded in contracts to ensure tamper-resistant validation and secure result consolidation [29,35]. Some platforms use structured smart contract interfaces with state loading, block validation, and InterPlanetary File System (IPFS) persistence to ensure secure execution and prevent malicious behavior [35,38].

4.2.3. Learning Protocols

Federated learning (FL) is the dominant learning paradigm across DAI systems [9,53]. Recent blockchain-enabled federated learning (BCFL) frameworks decentralize training and mining at the client level, enabling peer-based aggregation, personalized privacy, and eliminating third-party intermediaries, similar to swarm learning [13]. Several systems integrate multi-agent reinforcement learning or hybrid knowledge-based AI architectures for task-specific intelligence [5,25,53,54].

4.2.4. Privacy-Preserving Techniques

Privacy is a major concern in decentralized environments, leading to the adoption of cryptographic tools. Zero-knowledge proofs (ZKPs) are used to validate model contributions without revealing raw data [2,55,56]. Differential privacy and SMPC are also implemented to protect against inference and poisoning attacks [6,53,57,58]. Together, these technologies form the operational core of DAI ecosystems, enabling trusted coordination, secure learning, and privacy-aware computation across diverse, decentralized environments.

4.3. Use Cases and Implementation Contexts

DAI systems underpinned by blockchain technologies are increasingly finding real-world applicability across a diverse array of domains, each leveraging the decentralized, privacy-preserving, and incentive-aligned properties of these architectures. Figure 2 presents a high-level taxonomy that categorizes the fundamental components of DAI systems, including system dimensions (computation, data ownership, governance), technology stack (blockchain layers, smart contracts, protocol coordination), and real-world use cases (application domains, actors, and incentive mechanisms).

4.3.1. Use Case Domains

Healthcare emerges as a prominent case where decentralized federated learning has been employed to enhance predictive modeling without compromising patient privacy [1,4]. Several studies implemented blockchain-backed federated learning frameworks for disease prediction and personalized treatment while preserving data locality and auditability [6,9,15]. For example, a privacy-preserving framework for medical image classification across hospitals and another one that focuses on brain tumor classification using a decentralized network of edge devices [16]. In the financial sector, DAI architecture is used to coordinate predictive modeling while ensuring transactional transparency and integrity [59]. A decentralized data-sharing platform was presented to facilitate financial fraud detection without centralized oversight [60]. Autonomous systems and smart cities also benefit from decentralized intelligence and coordination [5]. A decentralized vehicle-to-vehicle (V2V) learning system that uses blockchain to synchronize updates across vehicles, enhancing road safety while maintaining privacy [1,40,42,45]. Similarly, studies [45,61] deploy DAI to optimize energy management and traffic flow in urban settings by coordinating distributed agents through consensus protocols.

4.3.2. Implementation Context and Actors

DAI systems are implemented across diverse domains through a range of decentralized actors. Edge devices (e.g., phones, IoT sensors, or medical instruments) act as training nodes and data holders [36]. Validators on blockchain networks verify model updates and enforce consensus rules [1,39]. In more advanced settings, DAOs govern contributions, rewards, and model updates, as seen in AI marketplace platforms [49,62]. Deployment maturity varies: some systems have been field-tested, like the hospital network [48] and the V2V communication scenarios [42], while others remain at the simulation or pilot stage [1,21]. These real-world applications illustrate how DAI technologies are adapted to domain-specific needs around trust, data governance, and coordination [36,42,48].

5. Smart Contracts and Protocol Design

Smart contracts are digital agreements that automatically execute when predefined conditions are met [6,63]. Primarily associated with blockchain technology, smart contracts gained significant mainstream adoption through platforms like Ethereum, which serves as a prominent example [30,63]. In the context of DAI, smart contracts play a crucial role in enabling decentralization by replacing the need for a central server. They coordinate learning processes and facilitate model averaging to produce a new global model [1,10,35]. Specifically, smart contracts manage nodes and model parameters while storing metadata on each node, allowing for the tracking of both the model being used and the data on which that model operates [50,64,65]. Collectively, these capabilities underscore smart contracts’ pivotal role in securing trust, enforcing incentives, and maintaining decentralized governance in blockchain-enabled DAI architectures, directly addressing RQ2.

5.1. Roles of Smart Contracts in Decentralized AI

Aggregation smart contracts are employed within DAI systems to merge individual models from edge nodes into a unified, improved global model, which is then redistributed across the blockchain network [4,61,66,67]. These contracts serve as a decentralized alternative to centralized training servers, while retaining several of the key advantages offered by centralized approaches, including distributed training and enhanced privacy preservation [11,42]. Moreover, aggregation smart contracts address several critical security vulnerabilities associated with centralized systems. These include the elimination of a single point of failure, enhanced resistance to transaction fraud due to the transparency and immutability of the blockchain ledger, and increased robustness against data poisoning attacks through the deployment of model verifiers [39,61]. In the DAI context, smart verification contracts are utilized to authenticate submitted AI models and to detect poisoned or malicious nodes that may compromise the global model’s integrity [61,68]. According to [61], when participants upload meaningless models or when malicious nodes launch poisoning attacks, the global model may deviate significantly or even fail to converge. To mitigate these risks, governance smart contracts are implemented to provide access control, auditability, and the application of incentives or penalties based on model performance [35,39]. Some governance models adopt hierarchical structures, wherein designated nodes possess auditing capabilities to identify malicious behavior in other nodes, potentially leading to their exclusion from the network upon detection [61]. Incentive mechanisms within these systems can include rewards, registration fees, and gamification elements, which are designed to enhance participation and model quality [39,45,50].

5.2. Protocols Analyzed Include

Semi-decentralized Federated Learning (SDFL) represents a hybrid approach that integrates both centralized and decentralized elements of federated learning, typically through the automation of various protocol components using smart contracts [6]. The first type of smart contract utilized in this architecture is the system initialization smart contract, which disseminates upcoming federated learning tasks to all nodes within the blockchain network. This contract provides key task specifications, including the learning objective, the nature of the data involved, a base model from which to initiate training, and the corresponding incentive mechanisms [10,39].

Model training smart contracts function differently based on whether training occurs on-chain or off-chain. In on-chain scenarios, smart contracts are responsible for directly handling model updates and data processing tasks [50,65,69]. Conversely, in off-chain training settings, smart contracts perform combinations of model aggregation, verification, and filtering of poisoned models, depending on the model-specific implementation [4,13,35]. The aggregation and verification mechanisms employed in this context closely resemble those outlined in Section 5.1.

In addition to coordination and validation, smart contracts also manage the incentive structures of SDFL [50]. The automation of incentive logic benefits from the inherent security and privacy-preserving properties of blockchain networks [5,70]. Within these networks, dedicated smart contracts can evaluate the performance and accuracy of submitted models and subsequently reward or penalize the associated nodes based on their contributions [35,45,61,71]. Rewards may be distributed directly to trainers or facilitated through tokenization mechanisms [9,30], whereas penalties typically involve reductions in node reputation or forfeiture of deposited collateral, where applicable [39,50].

Tokenized Artificial Intelligence (AI) leverages blockchain technology and smart contracts to enable the exchange and monetization of data and AI models [22,49]. Like Semi-decentralized AI, Tokenized AI utilizes versions of the smart contracts, which serve comparable functions but differ in their specific implementations to accommodate the distinct nature of AI models in this context [1,4,7,35]. During this review, it was observed that most tokenized AI models discussed in the literature employ a combination of registration mechanisms and fraud prevention smart contracts to bolster trust, accountability, and system integrity [6,50].

The registration smart contract facilitates the onboarding of new peers into the system, which can be accomplished through the payment of a security deposit or by providing identifiable information, such as an Ethereum address [30,47]. These contracts may also serve to verify the identity and account of peers through methods such as digital signatures or asymmetric encryption [13,50,70].

In addition, fraud prevention smart contracts utilize the blockchain to establish a secure ledger, ensuring data integrity and preventing fraudulent activities. This is achieved by enabling fraud detection mechanisms that flag accounts exhibiting suspicious behavior. The blockchain’s immutability ensures that the data remains accurate and auditable, thus enhancing transparency and security [9,72,73]. Some Tokenized AI systems, such as the one developed by [50], use pre-existing blockchain networks like Ethereum to implement fraud prevention protocols, leveraging the security features inherent in cryptocurrency blockchains.

DAO is a self-governing entity that operates autonomously through collective control facilitated by smart contracts [23,46]. Governance smart contracts are fundamental to the functioning of a DAO, forming an integral part of its adaptive control mechanism that evolves in response to changing conditions [25]. These contracts are responsible for establishing and enforcing rules, proposing actions for member voting, updating models, and implementing corrective measures for non-compliant models [20,25,35].

DAOs often adhere to established token management standards, such as ERC-20 for fungible tokens and ERC-721 for non-fungible tokens (NFTs), particularly when integrated within a DeFi system [6,59,74]. Within such systems, smart contracts can automate essential financial functions, including lending, borrowing, trading, and interest rate calculations [24,25,59].

In addition to these governance and financial functions, DAOs utilize inference verification smart contracts, which enable the on-chain validation of computational results generated off-chain. This process is facilitated through zero-knowledge proofs, ensuring that sensitive data or model information remains confidential while simultaneously verifying the validity of the model [8,68].

5.3. Challenges

While the use of smart contracts in federated learning offers numerous potential benefits, there are also notable drawbacks associated with their implementation. The primary disadvantage is the resource demand required to execute smart contracts [8,10]. Smart contracts, particularly on blockchain networks like Ethereum, incur high gas costs, which are significantly elevated when running complex algorithms, such as linear regression, on-chain compared to off-chain execution [75]. Additionally, the consensus process in blockchain systems introduces substantial computational and communication overhead, leading to increased latency, higher energy consumption, and greater resource demands [9,40].

Another critical limitation is the inability of current smart contracts to handle the substantial computational loads associated with collaborative learning processes [13,45]. Furthermore, the technical expertise required to develop and maintain such systems is considerable. Implementing cutting-edge decentralized AI applications, such as a decentralized AI banking system, necessitates a deep understanding of both AI and DeFi, fields that currently lack sufficient numbers of professionals proficient in both domains [59]. For individuals without expertise in these areas, working on these systems can be challenging, as the integration of blockchain with federated learning (FL) via smart contracts adds significant maintenance complexity, particularly due to the lack of interoperability between various blockchain platforms and AI frameworks [8,72,73,76]. Moreover, challenges related to scalability, privacy, security, and governance also exist, which will be further explored in Section 7.

6. Application and Use Case

The practical deployment of DAI systems reveals a rapidly expanding range of real-world applications across multiple sectors. These implementations reflect the versatility and resilience of blockchain-enabled architectures, smart contract coordination, and privacy-preserving learning techniques. In line with RQ3, this section explores how DAI frameworks are applied in key industry verticals, the roles and interactions of system actors, the design of incentive mechanisms, and the emergence of novel use cases in scientific research, logistics, and decentralized governance. By analyzing these applications, we demonstrate the transformative potential and adaptive capacity of DAI technologies in addressing complex, data-driven challenges across decentralized environments.

6.1. Industry Verticals

6.1.1. Healthcare

Privacy-preserving AI models are increasingly adopted in healthcare settings where data sensitivity is paramount. The challenge lies in training accurate models without centralizing patient records due to strict legal and ethical regulations [48]. Federated Learning (FL) offers a solution by enabling collaborative training of diagnostic or predictive models across hospitals, without moving data from local environments [31]. Research [21] demonstrates a privacy-aware DAI architecture utilizing SMPC to prevent leakage and resist inference and poisoning attacks. Similarly, study [4] outlines a decentralized framework where patient data remains local while the shared model benefits from aggregated updates, addressing Health Insurance Portability and Accountability Act (HIPAA) compliance concerns in the United States. Further applications in decentralized healthcare AI include the deployment of Swarm Learning, which enables hospitals to train models for tasks such as cancer detection collaboratively and COVID-19 CT scans, while keeping data on local servers. DFL also supports use cases like Federated Tumor Segmentation, where U-Net architectures are trained across institutions to segment brain tumors without centralizing patient data [77]. Similarly, the MELLODDY project (Machine Learning Ledger Orchestration for Drug Discovery) allows pharmaceutical companies to collaboratively train models on sensitive compound data using federated learning, without disclosing proprietary information [78].

6.1.2. Finance

The financial sector faces stringent compliance rules and privacy mandates, hindering data pooling across institutions. To overcome this, decentralized learning models are applied to detect fraud, assess creditworthiness, and conduct anti-money laundering (AML) analysis. Articles [21,48] describe secure aggregation and differential privacy approaches to facilitate collaborative model training across banks. Blockchain enhances auditability and ensures transactional transparency, while tokenized incentives motivate institutions to contribute data securely, as shown in [52].

6.1.3. Smart Cities

Urban infrastructures generate large volumes of distributed data from transportation, utilities, and safety systems [19]. Traditional centralized systems are inefficient for handling this at scale [5]. Decentralized AI, especially when integrated with Edge AI and blockchain coordination, supports real-time traffic management, pollution control, and emergency response [2,4,12,73]. Study [54] presents a multi-agent system coordinated via a DAO for energy optimization, while [25] illustrates on-chain automation of smart contracts to trigger urban actions based on environmental sensor data.

6.1.4. AI Marketplaces

To counteract the monopolization of AI assets, decentralized marketplaces enable the tokenized exchange of datasets, trained models, and computer resources [2]. These platforms provide transparency and fair access by leveraging smart contracts for licensing and incentives [56]. For instance, Ocean Protocol, which is a decentralized exchange of data and compute, facilitates the publication and exchange of data using tokenized assets, including NFTs for ownership and data tokens for access control [79]. Its Compute-to-Data feature allows AI models to be trained on private data without relocating it, preserving confidentiality while enabling federated learning across multiple silos. While Ocean Protocol focuses on data exchange, Singularity NET offers a peer-to-peer marketplace for AI agents and services, where developers can share, monetize, and consume AI models across a decentralized infrastructure [73]. Numeraire, by contrast, is a decentralized platform that supports crowd-sourced financial modeling using staked predictions, incentivizing contributors to submit accurate models in exchange for rewards [80]. Study [19] proposes a decentralized brokerage system where contributors are rewarded in tokens for sharing valuable AI assets, while [72] elaborates a reputation-based model to rank contributions and enforce service-level agreements in marketplaces.

6.2. System Actors and Interaction

The operational dynamics of DAI ecosystems depend on the coordinated roles of multiple actor types, each contributing distinct functions in training, governance, and enforcement [12,75]. Figure 3 illustrates the step-by-step lifecycle of DAI model development, structured into six stages: data collection, model training, model aggregation, validation and verification, deployment, and incentivization.

• Data Collection: Data is sourced in a distributed manner from edge devices and local nodes. This decentralized approach enhances data privacy, reduces latency, and ensures a more diverse and representative data foundation for training.
• Model Training: Training occurs locally using federated learning or swarm learning techniques. This preserves data sovereignty by ensuring that raw data remains on local devices while models are updated collaboratively.
• Model Aggregation: Smart contracts are utilized to facilitate the secure and autonomous aggregation of local model updates. This step is critical for combining individual model contributions into a global model without central oversight.
• Validation and Verification: The integrity and authenticity of the aggregated models are ensured through blockchain-based validation mechanisms. This immutable verification layer enhances trust, accountability, and reproducibility in the model development process.
• Deployment: The validated global model is deployed across the decentralized network. This distributed deployment architecture supports resilience, scalability, and operational continuity without reliance on a single point of control.
• Incentivization: Participants are rewarded through token-based mechanisms that align stakeholder incentives with network objectives. These incentives drive continued participation, data sharing, and computational contribution.

6.2.1. Edge Devices

These include IoT sensors, smartphones, and embedded systems that generate and process raw data locally [81]. Their primary role in decentralized AI systems is to perform initial computation, pre-training, or inference near the data source, which reduces latency and enhances privacy. Study [53] presents a decentralized learning system in which edge devices train models locally and share only encrypted parameters using SMPC, thereby mitigating risks of adversarial inference and poisoning attacks. Similarly, research [76] introduces an edge-centric protocol that employs homomorphic encryption for secure update aggregation, ensuring data confidentiality during federated learning coordination.

6.2.2. Participants

Participants are users or institutional clients who contribute to model training by submitting updates, often using federated or peer-to-peer mechanisms [11,43,82]. Their behavior is incentivized or penalized depending on their contributions or honesty [83,84]. In article [85], participants submit encrypted gradients that are verified and aggregated on-chain using smart contracts. In [2], training participants were also given governance rights, allowing them to vote on protocol upgrades through a DAO framework.

6.2.3. Validators

Validators assess the quality of model contributions. They may perform this evaluation via consensus algorithms, reputation systems, or cryptographic proofs [21,40,68]. In article [4], validators verify uploaded models’ performance through a challenge-response protocol using synthetic datasets. Article [10,39] details a reputation-based validator mechanism where participants stake tokens and validate others’ work, with slashing penalties for dishonest evaluations.

6.2.4. Smart Contracts

Smart contracts automate essential tasks such as model aggregation, reward distribution, and dispute resolution. Payouts can be triggered only when submitted model updates satisfy predefined validation criteria [36,62]. They can also enforce time-bound aggregation and coordinate contributors asynchronously without relying on a central authority [10,47].

6.2.5. DAOs

DAOs serve as decentralized governance bodies that handle tasks such as protocol upgrades, incentive distribution, and dispute resolution [5,20,23]. One system manages resource allocation and enforces compliance policies within a decentralized AI training marketplace through blockchain-based coordination [86]. Another example outlines a city-scale DAO coordinating multi-agent decisions for traffic and energy optimization, promoting transparency and democratic governance [25].

Figure 4 illustrates the key components of decentralized AI systems, including end devices, edge devices, validators, DAOs, and smart contracts, and their interactions across multiple application domains such as healthcare, finance, smart cities, and AI marketplaces.

6.3. Industry Incentive Structure

In DAI systems, incentive mechanisms are critical for ensuring active, honest, and sustained participation from diverse actors [39,45,52]. These structures not only motivate data sharing and model training but also govern quality assurance, penalize malicious behaviors, and drive protocol sustainability. The four dominant incentive schemes observed are token-based rewards, reputation systems, staking mechanisms, and hybrid models [39,42,45,47].

6.3.1. Token-Based Rewards

Many DAI platforms utilize tokens to reward participants for contributing data, computation, or model updates. Article [39] describes a system where contributors earn utility tokens based on the accuracy of their updates and timely participation. These tokens are typically distributed automatically through smart contracts once model updates are validated, ensuring transparency and trust in the reward process [36]. To maintain the integrity of contributions, many systems also require participants to stake tokens in advance, which discourages spam or low-quality submissions by introducing a potential penalty for malicious or careless behavior [50]. Similarly, in [86], tokens are used as micro-payments for autonomous systems that execute tasks like traffic control or energy load balancing within a city-scale DAO.

6.3.2. Reputation Systems

Reputation scores track the historical behavior and reliability of participants, guiding trust and future rewards. Study [39] outlines a decentralized reputation framework that dynamically scores users based on their validation accuracy and submission integrity. In study [83], validators accumulate reputation by consistently verifying accurate models, which increases their weight in aggregation and voting mechanisms.

6.3.3. Staking and Slashing

Participants often stake or lock tokens as collateral, which can be partially or fully slashed if they act dishonestly [62]. Research [39] incorporates a slashing protocol that penalizes dishonest participants for submitting poisoned updates or attempting inference attacks. Similarly, [10] introduces a staking mechanism that ensures validators and contributors remain accountable; slashing occurs when their output is flagged as low-quality by consensus.

6.3.4. Hybrid Incentives

Several systems combine token rewards, reputation scores, and staking to balance motivation and security. In [45], a multi-layered system integrates token incentives with DAO-governed staking policies. Participants with higher reputation scores earn bonus tokens and are entrusted with voting rights in governance decisions. Study [36] implements a hybrid model where both reputation and financial incentives regulate collaboration and verification tasks, providing redundancy and robustness.

6.4. Emerging and Niche Applications

Beyond well-established sectors like healthcare and finance, DAI systems are also gaining traction in emerging domains that benefit from collaborative intelligence and privacy-preserving data sharing. These applications showcase the adaptability of DAI frameworks and highlight future research directions.

6.4.1. Decentralized Scientific Collaboration

Several platforms facilitate cross-institutional training of AI models for scientific research. Study [16] introduces a federated learning framework tailored for biomedical research, where multiple academic institutions train diagnosis models without pooling raw data. This approach mitigates privacy risks while fostering open science and reproducibility. Similarly, research [62] explores collaborative genomics research using DAI infrastructure, emphasizing distributed learning across institutional silos.

6.4.2. Supply Chain Intelligence

Decentralized AI is applied to optimize routing, demand prediction, and compliance within complex supply chains. Study [63] presents an architecture where IoT-enabled edge nodes collect logistics data, which is then processed using federated learning to predict shipment delays. Smart contracts ensure provenance verification and compliance with regulatory standards. In [18], AI agents operate across supplier nodes to adjust production rates and shipping schedules in real time, guided by demand forecasts.

6.4.3. Decentralized Content Moderation

DAI frameworks are also being tested in decentralized social platforms to enable community-driven content moderation. Study [36] describes a reputation-based moderation system where contributors train classifiers for toxic content, and validators assess model performance via smart contracts. Similarly, [45] employs a voting-based governance layer that uses federated learning to train their moderation models. These platforms attempt to decentralize control over content decisions traditionally monopolized by centralized tech platforms.

7. Challenges and Open Problems

Despite significant advancements in DAI powered by blockchain technologies, the integration of these systems remains constrained by several critical challenges. These challenges span technical, organizational, and ethical domains, posing obstacles to performance, adoption, and trust. In line with RQ4, this section synthesizes the most prominent issues identified across the literature, highlighting both foundational and emerging issues that must be addressed to support the effective deployment and long-term sustainability of DAI systems.

7.1. Scalability and Performance Bottlenecks

Over 50% of the articles filtered in our literature review identify scalability as a significant concern in the context of blockchain technology. The consensus among these articles is that scalability challenges primarily stem from the computational overhead introduced by the execution of smart contracts and the consensus mechanisms inherent in blockchain operations [6,73,87,88]. These challenges are compounded by the difficulty of training large-scale models (e.g., LLMs) across heterogeneous, low-power nodes. Several articles propose potential solutions to address these issues. Some suggest leveraging off-chain nodes to handle resource-intensive computations or to store the models [4,8]. However, the tradeoffs between off-chain and on-chain coordination complicate deployment strategies. Other studies advocate for modifications to the underlying blockchain algorithms, such as transitioning from proof of work to more efficient consensus mechanisms [10,13], or updating existing algorithms to newer, more efficient versions [8,35,39]. A critical open problem is how modular or Layer-2 solutions, such as rollups and directed acyclic graph (DAG)-based structures, can be effectively integrated to enhance throughput and reduce operational costs in decentralized learning systems [8,19,40,59].

7.2. Privacy and Security

The primary motivation for the transition from a centralized server to a DAI system is the privacy and security concerns associated with a single point of failure [1,89]. A breach of the central server would result in the loss not only of the entire AI model but also of all associated user data [1,35,90]. In the context of DAI, the key privacy and security challenges are data privacy and confidentiality, as well as the risk of malicious attacks [1,2,14,91]. Directly sharing raw data poses significant privacy risks; therefore, to mitigate these concerns, only model parameters, updates, and gradients are exchanged between nodes, with each node training on its own dataset [11,14,92]. However, sharing model information introduces the possibility of model inversion attacks, necessitating the use of encryption or other techniques, such as ZKPs, to secure the model data [11,45,92]. Additionally, Byzantine and poisoning attacks are addressed through a combination of auditing, reputation mechanisms, and access control strategies [1,39,40,93]. Despite these mitigations, key vulnerabilities persist, including leakage via gradient vectors, Sybil attacks, and the complexity of implementing secure aggregation using homomorphic encryption or MPC. A pressing open problem is whether ZKPs or verifiable computation frameworks can scale sufficiently to enable real-time model auditing without compromising system performance or increasing architectural complexity [1,2,8].

7.3. Incentive Misalignment

A major challenge associated with incentive mechanisms in decentralized systems is the difficulty of designing mechanisms that simultaneously encourage participants to contribute valuable data and computational resources while discouraging the submission of poor-quality data [45,62]. For instance, some contributors may intentionally train the model using only a small subset of data to reduce their computing costs, which can negatively impact the performance of the global model [42]. One proposed solution to address this issue is to filter out bad actors through the use of smart contracts. However, this approach introduces the downside of increased complexity within the blockchain, potentially leading to delays as the system scales [35,61]. Another possible solution involves requiring participants to submit a deposit before contributing to the model. While this method may help to mitigate the issue of poor data submission, it has the unintended consequence of discouraging legitimate participants from contributing to the model [62]. Incentive misalignment may also manifest through free-riding behavior, short-term reward gaming, and overreliance on centralized oracles for verification. A critical open challenge is identifying the optimal design space that balances staking, slashing, reputation, and DAO governance to sustainably align individual incentives with collaborative, long-term model performance [10,22,62].

7.4. Governance and Trust

Due to the absence of a universally accepted framework for blockchain governance, there exists considerable variability in its different implementations [46]. Many federated learning systems continue to rely on a central server for the distribution of aggregated models, while others have fully embraced decentralization through the use of DAOs [1,25,35]. However, existing governance structures within DAOs have been shown to suffer from issues such as “whale dominance,” where a small number of wealthy participants disproportionately influence decision-making, thereby undermining the effectiveness of voting mechanisms [2]. It is important to note that DAOs are still in the process of development, with their definitions and structural possibilities remaining ambiguous, which complicates their widespread adoption [20,25]. A related challenge is the lack of transparent, auditable standards to ensure model integrity and compliance in decentralized settings. In practice, human intervention remains necessary for dispute resolution and quality assurance. An emerging open problem is whether decentralized model certification frameworks, such as audit tokens or verifiable behavior logs, can be formalized and widely adopted to establish trust and accountability in autonomous AI governance [2,9,20,45].

7.5. Interoperability and Standardization

The lack of standardization in both decentralized systems and AI models supported by blockchain technology impedes their adoption and integration [5,73]. Consequently, decentralized systems often comprise heterogeneous components that necessitate the use of AI to bridge gaps, thereby enabling effective communication and data exchange [5]. Furthermore, integrating pre-existing systems into decentralized AI frameworks presents significant challenges [5]. To facilitate this integration, middleware development is often required to enable communication between different platforms. However, due to the lack of standardization, this integration process must be tailored individually for nearly every AI model [5,73]. Moreover, the absence of standardized frameworks complicates the accurate comparison of AI models and the assurance of their compatibility across various organizations [73]. These issues are exacerbated in multi-chain environments, where cross-chain coordination of training, aggregation, and incentives remains complex. A key open problem is how to develop standardized benchmarks and cross-chain protocols that preserve privacy, integrity, and fairness while enabling interoperability across diverse blockchain platforms [22,73].

7.6. Ethical, Legal and Social Implications

Ethical concerns surrounding DAI primarily focus on AI model bias, fairness, and trust. AI models trained on biased data inherit these biases, which can perpetuate discriminatory outcomes in model predictions [22]. One approach to counteracting bias is through model aggregation, which helps to mitigate the effects of biased data by combining models trained on diverse datasets [23]. Another significant ethical concern is related to incentive fairness, where participants are rewarded based on their contributions to the model. Some nodes may exploit the incentive structure by training the model with minimal data, thereby reducing computational costs while still receiving rewards [10,42]. Lastly, trust in AI models is a critical ethical consideration. The integration of blockchain technology enhances trust by providing features such as transparency and immutability, which ensure auditable records and verifiable data, thereby promoting data integrity [6,9].

In addition to ethical issues, regulatory compliance presents a substantial challenge for DAI, particularly in healthcare, where patient data is protected by stringent regulations such as HIPAA [19,94]. Furthermore, there exists legal ambiguity concerning DAOs, as they do not fall clearly under existing regulatory categories, making it difficult to determine the applicable legal framework [46]. The lack of recognition of AI systems, models, and agents as legal entities compounds this issue, as it remains unclear who is accountable for their failures [2]. DAI systems also face jurisdictional challenges, particularly when nodes are dispersed across multiple countries, complicating the determination of the appropriate legal frameworks [2,25]. Additionally, issues related to intellectual property and copyright arise in the context of data stored on-chain. The resistance of public blockchains to censorship further complicates enforcement, as legal authorities may struggle to impose bans effectively [2,19,22]. Collectively, these concerns underscore an open problem in embedding ethical principles, fairness guarantees, and legal accountability into DAI protocols in a way that is enforceable across decentralized, jurisdictionally fragmented, and technically diverse infrastructures.

Table 1 summarizes the key challenges faced by decentralized AI systems, outlining critical issues and highlighting open problems that require focused research to advance the field.

Table 1. Key Challenges and Open Problems in Decentralized AI.

Challenge Area	Key Issues	Open Problems	Sources
Scalability and Performance	Low throughput, high gas fees, coordination latency	Layer-2 and modular blockchains for faster, cheaper ops	[4,6,8,10,13,19,35,39,40,59,73,87,88]
Privacy and Security	Data leakage, poisoning attacks, secure aggregation difficulty	Scalable ZKPs and privacy-preserving verifiable computation	[1,2,8,11,14,35,39,40,45,89,90,91,92,93]
Incentive Misalignment	Free-riding, manipulation of rewards, oracle reliance	Optimal mix of staking, slashing, and reputation mechanisms	[10,22,35,42,45,61,62]
Governance and Trust	Opaque or slow DAO decisions, human intervention needed	Model certification and decentralized quality assurance	[1,2,9,20,25,35,45,46]
Interoperability and Standardization	Incompatible protocols, lack of cross-chain standards	Cross-chain model training and benchmark creation	[5,22,73]
Ethical, Legal, and Social Implications (ELSI)	Undefined liability, ownership ambiguity, fairness gaps	Embedding ethics and bias detection into protocols	[2,6,9,10,19,22,23,42,46,94]

Table 1. Key Challenges and Open Problems in Decentralized AI.

Challenge Area	Key Issues	Open Problems	Sources
Scalability and Performance	Low throughput, high gas fees, coordination latency	Layer-2 and modular blockchains for faster, cheaper ops	[4,6,8,10,13,19,35,39,40,59,73,87,88]
Privacy and Security	Data leakage, poisoning attacks, secure aggregation difficulty	Scalable ZKPs and privacy-preserving verifiable computation	[1,2,8,11,14,35,39,40,45,89,90,91,92,93]
Incentive Misalignment	Free-riding, manipulation of rewards, oracle reliance	Optimal mix of staking, slashing, and reputation mechanisms	[10,22,35,42,45,61,62]
Governance and Trust	Opaque or slow DAO decisions, human intervention needed	Model certification and decentralized quality assurance	[1,2,9,20,25,35,45,46]
Interoperability and Standardization	Incompatible protocols, lack of cross-chain standards	Cross-chain model training and benchmark creation	[5,22,73]
Ethical, Legal, and Social Implications (ELSI)	Undefined liability, ownership ambiguity, fairness gaps	Embedding ethics and bias detection into protocols	[2,6,9,10,19,22,23,42,46,94]

8. Limitations

While this review employed a rigorous and transparent methodology, several limitations should be acknowledged. First, the exclusion of non–peer-reviewed sources and gray literature may have restricted the scope of emerging or practice-based insights that often appear outside formal publication channels. Second, although the screening and data extraction procedures incorporated reliability checks, the absence of a formal critical appraisal limits the interpretability of findings with respect to the methodological quality of individual studies. Third, the review focused on publications in English between 2016 and 2025; therefore, relevant contributions published in other languages or outside this timeframe may have been overlooked. Finally, the rapidly evolving nature of decentralized AI and blockchain research means that the findings provide a snapshot of the field rather than a definitive or exhaustive account. These limitations underscore the importance of viewing the synthesis as a mapping of current knowledge and open challenges rather than an empirical validation of outcomes.

9. Future Research Directions

Despite the growing body of literature on DAI, several gaps in the research remain that warrant further exploration. One such gap pertains to the development of more efficient consensus algorithms tailored for blockchain-based AI models [9]. Furthermore, there are ongoing opportunities to optimize existing blockchain configurations for federated learning, including enhancing the block generation rate [73].

Another important area for future research involves privacy and security. Specifically, the integration of post-quantum cryptography into current DAI models is essential to safeguard against potential quantum attacks [3]. In the broader context of AI, it is critical to develop new encryption techniques, data anonymization methods, and data-sharing protocols to strengthen the data privacy of AI models [5]. This issue is particularly pressing for DAI systems operating within DAOs, which still face significant legal and regulatory challenges. At present, the legal categorization of DAOs remains ambiguous, which complicates their regulatory framework [46]. Additionally, there is a notable absence of real-world case studies on the implementation of decentralized AI operating within DAOs.

Further research is also needed in the areas of interoperability and ethics. For instance, adopting a hybrid approach that leverages multiple AI models to optimize performance by capitalizing on each model’s unique strengths presents an intriguing avenue for investigation [6]. Standardizing protocols, APIs, and model formats is vital to enable seamless integration across heterogeneous DAI ecosystems. From an ethical perspective, there is a need for the development of comprehensive frameworks and guidelines to address the ethical implications of automated decision-making, particularly concerning fairness and bias mitigation. Additionally, research into international cooperation on cybersecurity threats and data sovereignty in the context of DAI remains sparse and underexplored at the time of writing.

Looking ahead, scalable and resource-efficient decentralized learning architectures are essential for the advancement of DAI. This includes leveraging modular blockchain layers and Layer-2 solutions such as rollups to achieve fast, cost-effective coordination. Asynchronous learning approaches and compressed model updates can mitigate latency and bandwidth constraints, facilitating the decentralized training of large-scale AI models, including foundation models and large language models, across diverse infrastructures. The integration of verifiable off-chain computation and zero-knowledge machine learning frameworks offers promising paths to ensure integrity and auditability at scale.

In summary, addressing these research directions will be critical to unlocking the full potential of decentralized AI, ensuring its scalability, security, ethical alignment, and interoperability in increasingly complex and distributed environments.

10. Conclusions

This systematic review depicts the current state of the art of DAI systems and the transformative potential of DAI, which combines federated learning, blockchain technologies, smart contracts, and decentralized governance mechanisms to support privacy-preserving, secure, and transparent AI systems. Our analysis across various domains, including healthcare, mobility, and decentralized marketplaces, shows that DAI offers viable alternatives to centralized AI models, especially in contexts where data sensitivity, trust, and multi-stakeholder coordination are paramount. Key innovations such as Compute-to-Data frameworks, token-based marketplaces, Swarm Learning, and smart contract-based model validation demonstrate the breadth of architectural solutions currently explored in the literature. Despite this promise, significant barriers remain before DAI can reach widespread adoption. Challenges related to scalability, interoperability, and performance bottlenecks continue to hinder real-time deployments. Legal and regulatory uncertainties, especially regarding the classification and accountability of DAOs and AI agents, pose complex compliance hurdles. In addition, issues of incentive misalignment, data/model standardization, and governance inequities require further investigation to ensure that DAI systems are not only technically robust but also ethically sound and socially equitable. Future research should focus on improving consensus algorithms tailored to AI-specific needs, designing lightweight privacy-preserving computation techniques, and developing standardized frameworks to enable interoperability across decentralized platforms. Moreover, real-world implementation studies are urgently needed to validate the performance, scalability, and trustworthiness of DAI systems under practical conditions. Interdisciplinary efforts bridging AI, cryptography, policy, and ethics will be essential to guide the responsible evolution of DAI ecosystems. If these limitations are systematically addressed, DAI has the potential to redefine the foundations of trustworthy, democratized AI in an increasingly connected world.