{"id":"aa8308c0-b322-4ba9-9c27-a40b04cba683","arxiv_id":"2505.07828","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Most AI crypto tokens are not genuinely decentralized: they perform AI off-chain, add a token layer, and are driven heavily by speculation, according to this review.","lead":"This paper examines top AI-themed crypto tokens and finds that they mostly perform the actual artificial intelligence work on ordinary servers, with the blockchain handling payments and voting. It is a critical map for anyone deciding whether such tokens offer real decentralization or just the appearance of it.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The population-level conclusion extrapolates from a convenience sample of top-market-cap, self-identified AI tokens without on-chain usage data; the 'most/many' generalization is not empirically supported.","rationale":"Read in good faith, the paper is a competent critical review with a clear argument: since the reviewed leading projects execute AI off-chain and lack on-chain verification, they do not deliver fully trustless decentralized AI, and token value may decouple from actual use. The strongest claim inherits these observations, but the paper moves from a handful of case studies to a population statement ('most existing platforms,' 'many AI-token ecosystems'). The single most load-bearing assumption is therefore the representativeness of the market-cap-based, self-identified sample and the accuracy of the low-usage assertions. The reader identified this same issue, and I partially agree: I do not elevate the Section 4.1 decentralization standard (off-chain computation disqualifies genuine decentralization) to the primary concern because the paper argues for that standard and it is a definitional debate rather than an empirical gap. The more pressing weakness is evidential: Table 4's 'Notable Shortcomings' contain unsupported usage claims, and Section 4.4's assertion that usage metrics 'remain modest' is not backed by data. A concrete on-chain usage audit and a randomized sample check would settle whether the generalization is sound. This does not overturn the review's substance; it refines the epistemic status of the conclusion, so the conditional verdict remains appropriate and unchanged.","tokens_in":23645,"tokens_out":4254,"duration_ms":47149,"concrete_test":"Collect on-chain usage data (daily active addresses, transaction counts, compute-job volume, marketplace listing activity) for all projects in Table 4 from their respective chains or public explorers (e.g., Dune Analytics) over a defined window; then repeat the quality-of-decentralization assessment on a random sample of 50 lower-market-cap AI tokens from CoinGecko's AI category. If a substantial share of the random sample shows meaningful on-chain usage or verifiable on-chain inference, or if the Table 4 'no traffic' claims are contradicted by the data, the population-level conclusion requires substantial qualification.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The conclusion ('most existing platforms... many AI-token ecosystems... risk functioning primarily as speculative financial instruments') is a population-level claim, but the evidence is a non-random sample. Section 3 states the selected projects are 'drawn from the top-ranked AI tokens by market capitalization... limited to those that explicitly identify themselves as AI tokens.' No sampling frame, population size, or response rate is given, and the selection criteria plausibly over-sample hyped projects with strong token narratives rather than representative AI-token ecosystems. The paper's supporting usage claims are also unsourced: Table 4 asserts AGIX 'platform usage is low, with many listed services seeing no traffic' and OCEAN 'The marketplace sees limited real demand,' and Section 4.4 states 'Transaction volumes and usage metrics remain modest' without reporting any on-chain measurements or external usage data. Section 4.7 additionally extrapolates from non-AI tokens (Filecoin, Helium) to AI tokens. If the sample is biased toward speculative projects, or if the qualitative usage claims are inaccurate, the central generalization is overstated. The reader flagged this as the weakest assumption; I agree it is load-bearing.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper is a critical survey of AI-based crypto tokens. It reviews leading projects (Render, Bittensor, Fetch.ai, SingularityNET, Ocean Protocol, and several smaller ones), classifying them by blockchain base, AI role, token utility, consensus mechanism, and business model. The paper argues that despite technical ambition, most current AI-token platforms fall short of genuinely decentralized, trustless AI: they rely heavily on off-chain computation, lack robust on-chain verification, face scalability and network-effect problems, and in some cases function primarily as speculative instruments. It then discusses future directions such as zkML, verifiable off-chain computation, specialized blockchains, federated learning, improved tokenomics, and hybrid governance, and closes with practical guidelines for launching decentralized AI tokens.","tokens_in":23800,"tokens_out":4321,"duration_ms":43001,"significance":"The paper addresses a timely and important question: whether AI-token projects deliver real decentralization or primarily a narrative. Its strengths are a clear research-question structure, a useful comparative framework (Tables 1–4), integration of the empirical literature on token speculation, and a constructive set of future directions and launch guidelines. The qualitative central claim is consistent with much of the cited external evidence and with the public behavior of the projects described. However, the paper is an interpretive review rather than an empirical study: it introduces no original dataset or measurement, and its strongest population-level conclusions rest on a convenience sample and on unsourced claims about platform usage. If the evidentiary gaps are addressed, the paper could be a valuable reference for investors, regulators, and builders evaluating decentralized AI claims.","major_comments":[{"comment":"The sample is described as 'drawn from the top-ranked AI tokens by market capitalization ... limited to those that explicitly identify themselves as AI tokens,' but no sampling frame, population size, selection date, or inclusion/exclusion criteria beyond market-cap ranking are reported. Because Section 6 generalizes to 'most existing platforms' and 'many AI-token ecosystems,' the central conclusion is not supported by a non-representative convenience sample; either narrow the claims to the reviewed projects or add a principled sampling methodology and a defined population.","section":"Section 3, opening paragraph; Section 6"},{"comment":"Several load-bearing usage claims are made without evidence. Table 4 states that OCEAN's 'marketplace sees limited real demand' and that AGIX 'platform usage is low, with many listed services seeing no traffic,' and Section 4.4 asserts that 'transaction volumes and usage metrics remain modest,' yet no on-chain measurements, traffic data, or external citations are provided for these specific assertions. Since these claims underpin the speculative-instrument conclusion, they should either be supported with data/sources or explicitly framed as qualitative author assessments rather than empirical findings.","section":"Table 4; Section 4.4"},{"comment":"The paper treats heavy reliance on off-chain computation as evidence of a 'decentralization illusion' without defining the threshold at which token-coordinated off-chain computation ceases to count as decentralized. This evaluative standard is load-bearing because it drives the central critique of most reviewed projects. The paper should define the evaluative dimensions of decentralization (e.g., governance, infrastructure provision, verifiability, exit rights) and state concretely how each reviewed project fails them.","section":"Section 4.1; Section 2.2"},{"comment":"The argument that AI-token limitations are systemic draws on non-AI examples (Helium, Filecoin) and on ICO-era utility tokens. These examples are labeled as outside the AI domain, but they are nonetheless used to support the population-level conclusion about AI tokens. They should be presented strictly as analogies, with an explicit statement that they do not provide direct evidence about the representativeness of the AI-token sample itself.","section":"Section 4.7"}],"minor_comments":[{"comment":"Figure 1 reports market capitalizations on a log scale but provides no exact values or source-level data beyond 'April 2025'; adding a small data table or source URLs would make the figure reproducible.","section":"Figure 1"},{"comment":"References [34] and [72] are the same paper (Wang et al., on sandwich attacks in DeFi); one should be removed or cross-referenced.","section":"References [34] and [72]"},{"comment":"The sentence 'DeFi's built on blockchain infrastructures' contains a typographical error; it should read 'DeFi is built on blockchain infrastructures.'","section":"Section 2.1"},{"comment":"The sentence beginning 'Golem (GLM) iExec (RLC), which focuses on decentralized GPU computing, has migrated from Ethereum Layer 1 to Ethereum's Layer 2' is grammatically incomplete and conflates two distinct projects; it should be rewritten to describe Golem and iExec separately.","section":"Section 5.2"},{"comment":"The URL footnotes in Section 4.6 (the blockchainandthelaw.com post and the CFTC press release) should be converted into formal numbered references with access dates, consistent with the rest of the bibliography.","section":"Section 4.6"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a survey-style critical review rather than an original empirical or technical contribution, so the editor should confirm that this fits the journal's scope. I see no circularity concern: the author's self-citation [38] is used only as background and is not load-bearing for the central claim. The main risk is overgeneralization from a convenience sample and unsourced usage assertions, both of which can be addressed in revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's the short version: this is a solid, readable review of AI crypto tokens that makes a plausible critical argument, but it overreaches when it generalizes from its sample. The paper's real contribution is the structured comparison of eleven projects and a sensible design checklist, not new evidence.\n\nWhat it does well: the reading of the architectures is accurate. The central distinction between on-chain coordination and off-chain compute is correct and drives the critique home. The future-directions section is current (zkML, TEEs, federated learning), and the guidelines in 5.9 are practical and grounded in the case studies. The author is honest about the limitations and does not pretend the field is mature.\n\nThe main soft spot is the population-level claim. The sample is top-market-cap, self-identified AI tokens, which over-represents hyped projects. That is fine if the conclusion is 'these projects show X,' but the paper says 'most existing platforms' and 'many AI-token ecosystems.' The usage claims in Table 4 (AGIX services with no traffic, OCEAN marketplace with limited real demand) have no sources; they may be right, but the paper does not show the data. The Filecoin/Helium analogy is suggestive, not evidence.\n\nThat said, the core architectural critique does not depend on those usage claims. The fact that the flagship projects rely on off-chain computation and lack verifiability is enough to support a cautious version of the conclusion. I do not think the sample issue sinks the paper; it just means the strongest claims need to be dialed back.\n\nBottom line: this is a review paper, not a new result, and I would not cite it as primary evidence. But it is a fair, competent synthesis that deserves referee time. I would send it out with a request for a methods paragraph on sample selection and sourcing for the empirical usage claims. A revision that softens the generalizations would make it a useful reference for the intersection of AI and blockchain.","headline":"A credible review of AI tokens whose central critique survives its weak sample, but whose 'most/many' generalizations overshoot the evidence.","tokens_in":24335,"tokens_out":2226,"would_cite":false,"duration_ms":25000,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A review of leading AI-token projects concludes that most run their AI off-chain, so 'decentralized AI' is currently more aspiration than reality.","keywords":["AI tokens","decentralized AI","blockchain","off-chain computation","verification dilemma","tokenomics","speculative tokens","zkML"],"falsifier":"A direct test would be an on-chain audit of the leading AI-token networks counting how many AI service calls are actually executed and verified on-chain rather than merely paid for on-chain. If, for example, Bittensor's subnets were shown to verify model outputs on-chain at meaningful scale, and if usage metrics for SingularityNET or Ocean Protocol approached centralized alternatives like Hugging Face or Kaggle, the 'illusion of decentralization' claim would lose its empirical footing.","tokens_in":23411,"feed_emoji":"🤖","tokens_out":6398,"duration_ms":61797,"temperature":0.7,"pith_summary":"This review asks whether AI-based crypto tokens deliver genuine decentralized AI or merely the appearance of it. It studies leading AI-token projects, including Render, Bittensor, Fetch.ai, SingularityNET, Ocean Protocol, and Numerai, and concludes that most rely on off-chain computation, so blockchains act mostly as payment and governance layers rather than as trustless AI execution environments. The paper argues that tokenized platforms currently struggle with scalability, verification, network effects, and accountability, and that many function closer to speculative instruments than to engines of innovation. It does not dismiss the field: it identifies on-chain verification, federated learning, modular chains, and better incentive design as the paths that could close the gap.","feed_headline":"Most AI tokens run AI off-chain, review concludes","feed_subtitle":"Top AI-token projects use blockchains for payment and governance while AI computation stays off-chain, leaving 'decentralized AI' mostly…","key_machinery":"The load-bearing analytical device is the distinction between on-chain coordination and off-chain computation, combined with the 'verification dilemma': the correctness of an AI output is hard to confirm without re-executing it, so delegated computation cannot be trusted at face value. The review applies this lens across a comparison table of token projects, classifying each by blockchain base, AI role, computation location, token utility, consensus, and business model, and then evaluates the same projects against centralized AI services across compute execution, hosting, access, monetization, governance, transparency, and verification. This framework carries the argument because every central limitation, including off-chain reliance, scalability constraints, quality-control gaps, and weak network effects, follows from where computation lives and whether its result can be checked.","core_discovery":"The central claim is that current AI-token ecosystems have not yet delivered decentralized, trustless AI and in some cases risk functioning primarily as speculative financial instruments. The review substantiates this by tracing each major project's architecture: Render, Bittensor, Fetch.ai, SingularityNET, Ocean Protocol, Numerai, Cortex, and others place training and inference off-chain, reducing the blockchain to a payment, staking, and governance layer. Because on-chain verification of model outputs is still experimental, users must trust node operators, and the paper argues the resulting systems inherit many trust problems of centralized AI while adding crypto-economic overhead. The paper's constructive claim is that the gap is closable through zkML, trusted execution environments, proof-of-useful-work, blockchain-coordinated federated learning, and modular execution layers.","pith_inferences":["An unstated corollary of the paper's own evidence: the speculative component of AI-token valuations could be quantified directly from on-chain data by comparing token transfer volume with the frequency of verifiable service calls on the same network; a large divergence would convert the review's qualitative critique into a measurable metric.","The paper's off-chain criterion implicitly defines decentralization as on-chain execution; a competing definition would treat token-coordinated off-chain compute as decentralization of coordination rather than of computation. Under that definition, some of the paper's sharpest criticisms would soften, and the testable question becomes whether governance and verification are distributed even when e","If the paper is right that verification is the bottleneck, then the first AI-token project to ship practical zero-knowledge proofs or TEE attestations for its models would create a categorical advantage over peers, making verifiability the likely competitive axis of the next cycle."],"forward_implications":["Investors and users should discount decentralization claims from AI-token projects that cannot show on-chain verification or distributed control of model execution.","Regulators examining utility-token classifications have grounds to ask whether tokens function as payment for services or as speculative assets, since the paper documents a systematic gap between token design and actual usage.","Builders should focus on services that centralized AI cannot provide, such as privacy-preserving computation, collective model ownership, and censorship-resistant access, because parity with centralized APIs is unlikely to win adoption.","The merging of Fetch.ai, SingularityNET, Ocean, and CUDOS into the Artificial Superintelligence Alliance is presented as a signal of maturing infrastructure, but the paper's analysis suggests consolidation alone does not solve off-chain computation and verification gaps.","On-chain verification technologies, including zkML, TEE attestations, and quorum-based AI oracles, are the concrete path the paper identifies for moving AI-token platforms from coordination layers to trustless service layers."],"supporting_citations":[{"why":"Defines the building blocks of decentralized AI that frame the review's evaluative standard.","marker":"[1]"},{"why":"Case study of Render's decentralized GPU network and its token utility.","marker":"[4]"},{"why":"Case study of SingularityNET's AI service marketplace with off-chain model hosting.","marker":"[5]"},{"why":"Case study of Ocean Protocol's data marketplace and compute-to-data mechanism.","marker":"[6]"},{"why":"Case study of Fetch.ai's agent economy and off-chain training with on-chain settlement.","marker":"[7]"},{"why":"Case study of Bittensor's Proof-of-Intelligence consensus and subnet architecture.","marker":"[10]"},{"why":"Empirical evidence that utility-token usage declines while speculation dominates, supporting the speculative-instrument critique.","marker":"[41]"},{"why":"Filecoin example of low utilization relative to centralized storage, supporting the weak-network-effects argument.","marker":"[44]"},{"why":"Supplies the verification dilemma that grounds the trust and quality-control critique.","marker":"[68]"},{"why":"Introduces zkML as the leading emerging path for on-chain verifiable AI inference.","marker":"[80]"}],"fun_headline_variants":["AI tokens: decentralized in name, centralized in compute","Tokens don't decentralize AI; they just charge for it","Off-chain AI: the dirty secret of 'decentralized' tokens","The illusion of decentralized AI: tokens just pay for off-chain compute"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The conclusion depends on treating the top market-cap tokens that self-identify as AI tokens as representative, and on the standard that heavy off-chain computation disqualifies a system from genuine decentralization; if token-coordinated off-chain compute counts as a legitimate form of decentralization, or the sample skews toward hyped projects, the critique is overstated.","fun_headline_variants_meta":{"raw":{"variants":["AI tokens: decentralized in name, centralized in compute","Tokens don't decentralize AI; they just charge for it","Off-chain AI: the dirty secret of 'decentralized' tokens","The illusion of decentralized AI: tokens just pay for off-chain compute"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000679,"raw_usage":{"total_tokens":3084,"prompt_tokens":945,"completion_tokens":2139,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":561,"completion_tokens_details":{"reasoning_tokens":2067}},"tokens_in":561,"tokens_out":2139,"duration_ms":14564,"temperature":1.0,"reasoning_tokens":2067,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T05:19:51.947578+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A direct test would be an on-chain audit of the leading AI-token networks counting how many AI service calls are actually executed and verified on-chain rather than merely paid for on-chain. If, for example, Bittensor's subnets were shown to verify model outputs on-chain at meaningful scale, and if usage metrics for SingularityNET or Ocean Protocol approached centralized alternatives like Hugging Face or Kaggle, the 'illusion of decentralization' claim would lose its empirical footing.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Filecoin example of low utilization relative to centralized storage, supporting the weak-network-effects argument."},{"cited_title":"Technol., 2021, 34, (4), pp","cited_arxiv_id":null,"evidence_quote":"Supplies the verification dilemma that grounds the trust and quality-control critique."},{"cited_title":"19th Eur","cited_arxiv_id":null,"evidence_quote":"Introduces zkML as the leading emerging path for on-chain verifiable AI inference."}],"review_version":1}