{"id":"b7118a6b-743f-4e30-ab6a-110980ee14ed","arxiv_id":"2504.17539","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"The paper presents PoUI, a hybrid proof-of-stake and useful-AI-work consensus design with a job marketplace and dynamic reward adjustment, backed by simple energy arithmetic and a stylized simulation.","lead":"This paper proposes Proof of Useful Intelligence (PoUI), a blockchain consensus mechanism in which workers earn cryptocurrency by completing AI tasks and then stake those coins to help secure the network. The authors estimate PoUI uses about 0.6 kWh per worker-hour, an 83% cut from a PoW miner's 3.51 kWh, while generating useful AI output.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"PoUI's ledger security rests on an unverified assumption: that a stake-weighted validator majority will honestly verify subjective AI outputs, but no slashing condition or objective success criterion is specified, leaving collusion undetectable.","rationale":"The reader's weakest_assumption identifies exactly the same load-bearing concern: validators' ability to reliably and cheaply distinguish acceptable from unacceptable AI outputs in an adversarial setting. My stress-test confirms this is the point where the central claim of 'dual-purpose design' is least secure. The paper's only security measures are (i) task screening, (ii) majority-vote output verification, (iii) random validator assignment and transparency logs, and (iv) a reputation system. Each of these presumes that a validator's 'approve' decision is evidence of output quality, but for open-ended AI tasks no ground truth exists that a smart contract can consult. The transparency log does not help because the contract cannot tell whether a validator's vote was honest or collusive; the reputation system penalizes 'invalid outputs' but cannot identify them without the very verification the protocol is supposed to provide. The paper's own acknowledgment—'Future work will explore advanced attack detection'—is an explicit admission that the current design lacks a detection mechanism. Independent support is absent: there is no formal proof, no simulation of adversarial behavior, and no reproducible implementation of the verification layer. The simulation in Section V validates a worker-count regulation model that follows directly from the equations, so it does not bear on security. I do not see a way to patch this with parameter tuning; it requires a fundamentally different verification mechanism (e.g., objective tasks with deterministic checks, or an oracle-based dispute resolution protocol). I therefore recommend no change to the reader's CONDITIONAL verdict: the paper is a plausible design sketch, but it must scope the security and verification claims as open problems, correct the 97%/83% energy-reduction inconsistency, and specify a concrete verification scheme before the central claim can be accepted.","tokens_in":10109,"tokens_out":3376,"duration_ms":38670,"concrete_test":"Attempt to define an on-chain verifiable acceptance predicate for a representative PoUI task, e.g., 'generate a 100-word summary of a given news article.' Formalize this as a predicate P(output, task, validator_votes) that is (1) computable by a smart contract from data available on-chain, (2) returns true for all outputs produced by an honest worker following the task specification, and (3) returns false for a deliberately nonsensical output (e.g., random words). If no such predicate can be specified without an external oracle or without relying on validator votes themselves, then the protocol's approval mechanism is unfalsifiable.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim of PoUI is that the network is secure while producing useful AI output. Security here has two components: the PoS layer (validators selected by stake) and the output-verification layer (validators check workers' AI results). The load-bearing weakness is in the output-verification layer, as articulated in Section III-C.2 and III-C.3. The paper asserts that 'multiple validators review subjective tasks, requiring majority agreement for approval' and that 'smart contracts log all decisions transparently to detect collusion.' These assertions do not establish security because for subjective tasks—'text or image generation' (Section III-C, first sentence)—there is no objective, machine-checkable predicate that defines a correct or acceptable output. Without such a predicate, a smart contract cannot distinguish an honest approval from a collusive one; the log is just a record of votes, not evidence of correctness. The reputation system (Section III-C.4) is circular: it penalizes 'malicious actions, like submitting fraudulent tasks or invalid outputs,' but determining whether an output is invalid is precisely the verification problem the protocol fails to solve. The paper itself concedes the gap in Section III-C: 'Future work will explore advanced attack detection to further enhance security.' This is not a mere refinement; it is the absence of the mechanism that makes the ledger sound. A colluding set of validators controlling a majority of the review quorum—or, at the consensus level, a majority of staked coins—can approve arbitrary worker outputs, collect fees, and have those outputs permanently recorded in the blockchain, because no on-chain evidence can prove their vote was dishonest. Consequently, the paper does not substantiate the claim that PoUI 'ensures resilience against malicious attacks' (Section III-C).","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Proof of Useful Intelligence (PoUI), a hybrid blockchain consensus mechanism in which workers execute AI tasks, such as text or image generation, to earn coins that can be staked, while validators selected by stake verify outputs and append blocks. The architecture introduces four functional node roles (job posters, market coordinators, workers, validators), a smart-contract-based workflow, an energy-consumption model distinguishing security energy from useful-work energy, and a dynamic reward adjustment rule designed to keep the worker population near a target. The central claims are that PoUI reduces energy relative to PoW while producing useful AI output, and that it resists malicious attacks through task screening, majority-based output verification, collusion prevention, and reputation scoring.","tokens_in":10392,"tokens_out":8728,"duration_ms":84820,"significance":"If substantiated, PoUI would be a timely contribution to the useful-proof-of-work and blockchain-sustainability literature, combining the energy advantages of stake-based validation with a market for AI tasks. The paper has clear virtues: the workflow is explicitly specified, the per-role energy arithmetic is transparent and easy to reproduce, and the simulation setup in Section V-B is stated with concrete parameters. However, the current contribution is closer to a design sketch than a validated system. The security claims rely on an unverified verification mechanism for subjective AI outputs, the energy comparison mixes per-miner, per-validator, and per-worker quantities without network-level scaling, and the simulation largely validates the assumptions built into the worker-response model. These gaps directly affect the paper's headline claims of consensus security and system-level energy savings.","major_comments":[{"comment":"The security claim is not established. For tasks such as text or image generation, the paper itself notes that no single correct answer exists, yet it asserts that multiple validators can review subjective tasks with majority agreement and that logging decisions detects collusion. Without an objective, machine-checkable acceptance predicate, a smart contract cannot distinguish an honest validator vote from a collusive one, and the reputation system in Section III-C.4 is circular because it presupposes that invalid outputs can be identified. The cited frameworks MCP [18] and CodeAct [19] are not output-verification oracles with defined correctness criteria, and no slashing condition is specified for validators who approve invalid work. The paper's own admission in Section III-C that 'Future work will explore advanced attack detection' acknowledges this missing core, which is load-bearing for the claim that PoUI maintains network trust and ledger integrity.","section":"Section III-C.2/C.3/C.4"},{"comment":"The energy-reduction claim is internally inconsistent and mixes incomparable units. The abstract states a 97% energy reduction from PoW while adding value, but Sections IV-D.3 and VI report an 83% reduction based on PoUI's total 0.6 kWh/node versus PoW's 3.51 kWh/miner; 97% is only the security-energy comparison (0.1 versus 3.51). In addition, Table II's column heading 'kWh/node' actually lists per-miner, per-validator, and per-worker numbers. A system-level claim requires multiplying by the number of nodes of each type in the network, which the paper does not do. The claims should be re-scoped to per-role hardware comparisons or extended to a network-level energy model.","section":"Section IV and Table II; Abstract"},{"comment":"The simulation's apparent convergence is built into the worker-response model. Equation (11) sets w_{i+1} = w_i * beta * (1 + (r_{i+1} - r_i)/r_i) + N, with beta = 1, so the worker count responds mechanically and proportionally to reward changes, and equation (10) closes the loop by adjusting rewards to the gap. The simulation therefore demonstrates the internal consistency of the control rule, not an empirically validated property of worker participation. The claim that the mechanism 'successfully maintains the worker count near the target' should be presented as a consistency check, with sensitivity analysis and, ideally, independent behavioral evidence or a more defensive model of worker response.","section":"Section V-B, Eq. 11; Section V-C"},{"comment":"The protocol does not make useful work necessary for consensus participation. The abstract says workers earn coins 'which are then staked to secure the network,' but Section III-D only says workers 'can then be staked' and validators are selected by stake. Any coin holder can become a validator without performing AI tasks, so the security layer is, as specified, standard proof of stake with an optional task market attached. If the authors intend useful work to be a prerequisite for consensus influence, the rule and its game-theoretic consequences should be specified; otherwise the claim that PoUI 'integrates security with real-world utility' should be weakened to a description of parallel mechanisms.","section":"Sections I and III-D"}],"minor_comments":[{"comment":"There is a duplicated word in the sentence 'Assuming Pval = 100 W = 0.1 kW for an Ethereum-like validator node (e.g., a standard PC) and and t = 1 hour'; 'and and' should be a single 'and'.","section":"Section IV-B"},{"comment":"The noise term uses w_{i+1}^{(c)} without defining the superscript, and beta = 1 makes the beta factor redundant; please clarify the intended parameterization and define all symbols.","section":"Eq. (11)"},{"comment":"The 'informed analysis' that determines the target worker count w_tilde is never described; the simulation conclusions depend on this exogenous value, so the paper should at least discuss how it would be obtained in practice.","section":"Section V-A"},{"comment":"The phrase 'hardware overlap resulting in' is ambiguous: if a node both validates and performs useful work, the paper should justify why the energies add (0.1 + 0.5) rather than overlap or share components.","section":"Eq. (7) and Section IV-C"},{"comment":"The references [18] and [19] are named as frameworks for assessing output quality, but the paper does not explain how either framework provides an objective correctness criterion for arbitrary LLM outputs in an adversarial setting; please either specify the concrete verification protocol or remove the implication that these references close the security gap.","section":"Section III-C.2"}],"recommendation":"major_revision","confidential_remarks":"This paper is best assessed as a design/position proposal rather than a completed security protocol. The missing adversary model for output verification and the absence of a slashing condition are the critical gaps; if the journal accepts position papers, this could be considered after substantial revision, but for a regular research paper the authors should either provide a concrete verification mechanism with a threat model or explicitly scope security analysis out. The 97% versus 83% inconsistency and the unit mixing in the energy comparison should be fixed in any revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my read on PoUI. It's a short design proposal, not a security paper, and it should be read that way. The novel part is the combination: a decentralized AI job market, PoS-style validator selection, and dynamic reward balancing. That combination doesn't appear verbatim in the cited PoUW/PoI work, and the architecture with four functional roles is clearly described. What it does well: the energy arithmetic is transparent and correct for the per-node numbers, and the paper is honest about its own limits (\"Due to page limitations...\"). The simulation is simple but at least demonstrates the feedback loop behaves as intended.\n\nThe soft spots are real. First, the headline is wrong: the abstract says 97% energy reduction, but the body's own comparison for PoUI is 83% (0.6 vs 3.51 kWh/node). The 97% figure is the PoS-only reduction and should not be applied to PoUI. Second, the simulation in Section V-B is circular: Eq. 11 assumes workers respond to reward changes in exactly the way Eq. 10 anticipates, so convergence to the target worker count is built into the model, not an empirical finding. That's fine if presented as a consistency check, but the text says it \"validates\" the mechanism. Third, and most importantly, the security claims in Section III-C don't hold. For subjective AI tasks, \"multiple validators review... and require majority agreement\" is not a verification protocol. There is no objective, machine-checkable success criterion, so a colluding set of validators can approve arbitrary outputs, and the smart-contract log only records votes, not correctness. The reputation system is circular because it penalizes \"invalid outputs\"—but determining invalidity is precisely the unsolved problem. The paper's own \"future work will explore advanced attack detection\" is an admission that the load-bearing mechanism is missing.\n\nFor a design sketch, this is a reasonable contribution to the PoUW/PoI conversation. I'd send it to peer review, but with the clear expectation that the abstract be corrected and the security analysis either supplied or explicitly scoped as an assumption of honest validators. Without that, the main claim of \"resilience against malicious attacks\" is overstated. If it's a workshop, accept conditionally; if a journal, it needs major revision. A reader tracking useful-proof-of-work proposals or AI decentralized marketplaces will find value here; the core ledger security question remains open.","headline":"A plausible design sketch for AI-backed consensus that overstates both its energy savings and its security, but is worth engaging as a vision piece.","tokens_in":11036,"tokens_out":3044,"would_cite":false,"duration_ms":29281,"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":"The paper proposes Proof of Useful Intelligence (PoUI), a hybrid blockchain consensus in which workers earn coins by completing AI tasks and then stake those coins to secure the ledger, merging security with useful computation.","keywords":["proof of useful intelligence","blockchain consensus","proof of work","proof of stake","AI task marketplace","decentralized job market","dynamic reward adjustment","energy efficiency"],"falsifier":"A concrete test would be a small adversarial trial in which several validators independently review the same open-ended AI outputs, some genuine and some deliberately fabricated, measuring the agreement rate and whether majority rule rejects the fabricated ones; the central claim collapses if honest validators cannot agree on acceptable output often enough to keep fabricated work out of the chain.","tokens_in":9867,"feed_emoji":"🤖","tokens_out":7450,"duration_ms":68732,"temperature":0.7,"pith_summary":"Proof of Useful Intelligence (PoUI) is a consensus mechanism that pays workers in cryptocurrency for completing AI tasks such as text generation, image analysis, and public-good verification, and then uses the same workers' staked coins to secure the blockchain through a proof-of-stake validator layer. The paper's central claim is that this arrangement keeps the energy efficiency of proof of stake while adding real-world utility, replacing proof of work's puzzle-solving waste. Its accounting puts PoUI at 0.6 kWh per worker-node per hour, compared with 0.1 kWh for a proof-of-stake validator and 3.51 kWh for a proof-of-work miner. If correct, PoUI would be a consensus design that produces economically valuable output instead of only consuming energy.","feed_headline":"PoUI consensus runs AI tasks while securing the chain at 0.6 kWh","feed_subtitle":"PoS-level security plus paid AI work replaces wasteful mining hashes, cutting PoW energy by about 83 percent.","key_machinery":"The load-bearing objects are the hybrid role structure, the energy identity $E_{\\mathrm{tot}}=E_{\\mathrm{sec}}+E_{\\mathrm{use}}$, and the dynamic reward rule $r_{i+1}=r_i\\left(1+\\alpha \\frac{\\tilde{w}-w_i}{w_i}\\right)$ applied whenever $|\\tilde{w}-w_i|/w_i \\ge \\Delta$. The energy identity lets the paper count security energy separately from useful-work energy, so PoUI is presented as proof-of-stake security (0.1 kWh per validator) plus AI inference energy (0.5 kWh per worker). The reward rule keeps the job market stable: when workers fall below the target $\\tilde{w}=250$, rewards rise to attract participation; when they exceed it, rewards fall, with a deadband $\\Delta=0.05$ to suppress jitter. Validator verification, based on stake-based selection, a ban on self-validation, and majority approval of subjective outputs, is the security mechanism that connects useful tasks to ledger integrity.","core_discovery":"PoUI's central claim is that blockchain security and useful AI computation can be merged without sacrificing proof-of-stake-level energy efficiency. The network separates four functional roles: job posters submit divisible tasks, market coordinators screen and queue jobs, workers run AI models, and validators, selected by stake, check outputs before adding blocks. Workers earn coins that can be staked, creating a path from contributor to validator; validators cannot verify their own work, and stake caps plus random selection adjustments counter concentration. The energy identity $E_{\\mathrm{tot}}=E_{\\mathrm{sec}}+E_{\\mathrm{use}}$ separates security energy from useful-work energy, placing PoUI at 0.1 kWh per validator for security and 0.5 kWh per worker for useful AI work, for a total of 0.6 kWh per node, an 83% reduction from PoW while producing value that PoS lacks. A simulation with a target of 250 workers shows the dynamic reward rule bringing participation near target within about 25 steps and holding it there, with deviations typically within $\\pm 20$ workers.","pith_inferences":["A natural next experiment is measuring inter-validator agreement on real open-ended AI tasks; if honest validators cannot agree on what counts as acceptable output at a high rate, PoUI would need a different verification layer, such as objective benchmarks or reputation-weighted review, before it can secure a ledger.","The energy comparison is order-of-magnitude: the 0.5 kWh per worker assumes full A100 utilization and ignores idle time, queuing, and network overhead, so the real-world gap between PoUI and proof of stake depends on utilization rates.","If PoUI worked, the same mechanism could subsidize public-good AI tasks, such as dataset cleaning or Wikipedia verification, by treating block rewards as a community fund, effectively turning consensus into a public-works program.","The simulation's convergence assumes workers respond linearly to reward changes with bounded noise; real human or institutional workers may respond more slowly or strategically, so the deadband and sensitivity parameters would need empirical calibration."],"forward_implications":["PoUI would let a blockchain pay for useful AI work out of its normal block and task rewards, so network security and AI inference share the same hardware and energy budget.","A node that only validates spends 0.1 kWh per hour, the same as a proof-of-stake validator, while a node that both works and validates spends 0.6 kWh, still about 83% below the 3.51 kWh of a proof-of-work miner.","Workers who earn coins can stake them into validator selection, giving active contributors a path to security and governance influence, while stake caps and random selection adjustments limit dominance by large stakers.","Dynamic reward adjustment should keep the worker pool near the target set by job demand, clearing backlogs when rewards rise and avoiding over-participation when rewards fall.","PoUI broadens earlier proof-of-useful-work ideas from specialized ML training to diverse AI tasks such as text and image generation, lowering hardware and expertise barriers to participation."],"supporting_citations":[{"why":"Supplies the Antminer S21 Pro hash rate and power draw used to compute the 3.51 kWh/miner proof-of-work baseline.","marker":"[20]"},{"why":"Supplies the NVIDIA A100 active power draw used to set PoUI useful-work energy at 0.5 kWh/worker.","marker":"[21]"},{"why":"Named as the Model Context Protocol framework that validators use to assess AI output quality.","marker":"[18]"},{"why":"Named as the CodeAct framework for executable code actions that validators use to verify subjective AI tasks.","marker":"[19]"},{"why":"Introduces proof of useful work for AI, the baseline idea that PoUI extends with a decentralized job market.","marker":"[10]"},{"why":"Describes proof of intelligence, the closest prior consensus that PoUI contrasts with on task diversity and decentralization.","marker":"[14]"},{"why":"Provides the provably secure proof-of-stake protocol from which PoUI takes stake caps to limit dominance.","marker":"[17]"},{"why":"Supplies the random selection adjustments that PoUI uses to give less-staked nodes better odds.","marker":"[8]"}],"fun_headline_variants":["PoUI: blockchain security with useful AI work at 0.6 kWh","PoUI: mining replaced by AI tasks, slashing energy 83%","PoUI consensus runs useful AI work, not wasteful hashes","Proof of Useful Intelligence: AI work secures chain at 0.6 kWh"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole design assumes validators can cheaply and reliably decide, by running standard output-verification frameworks and taking a majority vote, whether an AI-produced answer is good enough to pay for; if subjective outputs cannot be checked in an adversarial setting, workers could submit garbage and colluding validators could approve it, breaking the ledger.","fun_headline_variants_meta":{"raw":{"variants":["PoUI: blockchain security with useful AI work at 0.6 kWh","PoUI: mining replaced by AI tasks, slashing energy 83%","PoUI consensus runs useful AI work, not wasteful hashes","Proof of Useful Intelligence: AI work secures chain at 0.6 kWh"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000838,"raw_usage":{"total_tokens":3641,"prompt_tokens":923,"completion_tokens":2718,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":539,"completion_tokens_details":{"reasoning_tokens":2637}},"tokens_in":539,"tokens_out":2718,"duration_ms":17214,"temperature":1.0,"reasoning_tokens":2637,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T10:37:08.920759+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A concrete test would be a small adversarial trial in which several validators independently review the same open-ended AI outputs, some genuine and some deliberately fabricated, measuring the agreement rate and whether majority rule rejects the fabricated ones; the central claim collapses if honest validators cannot agree on acceptable output often enough to keep fabricated work out of the chain.","supporting_citations":[{"cited_title":"Bitmain shop,","cited_arxiv_id":null,"evidence_quote":"Supplies the Antminer S21 Pro hash rate and power draw used to compute the 3.51 kWh/miner proof-of-work baseline."},{"cited_title":"NVIDIA A100 Tensor Core GPU Datasheet,","cited_arxiv_id":null,"evidence_quote":"Supplies the NVIDIA A100 active power draw used to set PoUI useful-work energy at 0.5 kWh/worker."},{"cited_title":"Introducing the Model Context Protocol,","cited_arxiv_id":null,"evidence_quote":"Named as the Model Context Protocol framework that validators use to assess AI output quality."},{"cited_title":"Executable code actions elicit better LLM agents,","cited_arxiv_id":null,"evidence_quote":"Named as the CodeAct framework for executable code actions that validators use to verify subjective AI tasks."},{"cited_title":"Cortex: The decentralized ai autonomous sys- tem,","cited_arxiv_id":null,"evidence_quote":"Describes proof of intelligence, the closest prior consensus that PoUI contrasts with on task diversity and decentralization."},{"cited_title":"Ouroboros: A provably secure proof-of-stake blockchain protocol,","cited_arxiv_id":null,"evidence_quote":"Provides the provably secure proof-of-stake protocol from which PoUI takes stake caps to limit dominance."},{"cited_title":"Incentives in ethereum’s hybrid casper protocol,","cited_arxiv_id":null,"evidence_quote":"Supplies the random selection adjustments that PoUI uses to give less-staked nodes better odds."}],"review_version":1}