Getting your Trinity Audio player ready...

Abstract

Research funding is commonly allocated through competitive grants, institutional block funding, performance-based formulas, philanthropy, prizes, crowdfunding, and lotteries. Each mechanism addresses different policy objectives but also creates characteristic distortions, including high administrative costs, uncertain ex ante prediction, cumulative advantage, institutional exclusion, and popularity bias.

This article introduces AI Internet-Meritocracy (AIIM) as a proposed complementary infrastructure for funding scientific and open-source contributions. AIIM uses artificial intelligence to evaluate documented outputs, identify dependencies among contributions, and distribute designated funding according to auditable rules. Its architecture supports distinct global, European Union, and country-specific funding pools.

AIIM may offer comparative advantages in retrospective and continuous funding, recognition of upstream infrastructure, inclusion of independent contributors, and reuse of evaluation evidence across funders. However, it also creates risks involving model bias, metric gaming, fabricated evidence, governance capture, privacy, and funding concentration.

A central security problem is prompt injection. Because contemporary large language models cannot reliably distinguish malicious instructions embedded in untrusted content from legitimate evaluation inputs, AIIM does not treat model judgment as the final security authority. Suspected serious misconduct, including deliberate prompt injection, is subject to human voting and appeal procedures.

The article proposes a five-month adversarial pilot with measurable targets for attack success, funding distortion, false bans, appeal accuracy, disciplinary inequality, auditability, and administrative cost. AIIM should therefore be evaluated as a testable component of a plural research-funding system rather than as a replacement for peer review or public research agencies.

Keywords: research funding; science policy; artificial intelligence; decentralized science; peer review; research evaluation; public goods; prompt injection; adversarial testing

Introduction

Research-funding institutions must make decisions under severe uncertainty. They must select among scientific questions whose eventual importance is unknown, investigators whose future performance cannot be observed directly, and outputs whose social value may emerge only after long delays. Conventional funding systems manage this uncertainty primarily through expert review of proposals, institutional reputation, past performance, strategic priorities, and administrative accountability.

These mechanisms perform indispensable functions. Expert panels can evaluate methodological feasibility, ethical acceptability, specialist competence, equipment requirements, and alignment with public missions. Governments can direct resources toward health emergencies, energy security, climate adaptation, defense, or national infrastructure. Universities can maintain laboratories and long-duration capabilities that cannot survive through atomized payments to individual contributors.

Nevertheless, funding peer review is neither a precise measurement instrument nor a costless governance mechanism. Studies have reported limited inter-reviewer reliability, weak relationships between review scores and subsequent outputs, administrative burden, and sensitivity to which reviewers happen to evaluate a proposal. A large analysis of Hungarian grant applications found only a minor relationship between reviewer scores and subsequent publication output, while applicants’ previous scientometric performance was substantially more predictive (Győrffy et al., 2020; Recio-Saucedo et al., 2022).

Research funding also exhibits cumulative advantage. Prior winners acquire resources, staff, visibility, and institutional credibility that improve their chances in subsequent competitions. Empirical work on the Matthew effect in science funding indicates that early funding success can generate durable advantages that are not fully attributable to pre-existing differences in performance (Bol et al., 2018).

These limitations have motivated modified lotteries, shorter applications, anonymized review, participatory panels, bibliometric decision support, and other reforms. Modified lotteries are particularly relevant where reviewers can distinguish clearly unsuitable proposals from a broad set of meritorious proposals but cannot reliably rank the surviving applications. New Zealand’s Explorer Grants, for example, have used random selection among applications passing an initial quality assessment (Fang and Casadevall, 2016; Liu et al., 2020; Recio-Saucedo et al., 2022).

AI Internet-Meritocracy, or AIIM, addresses a different allocation problem. Instead of asking only which proposed project should receive a fixed grant, it asks:

Which documented scientific and software contributions have generated value, which later outputs depend on them, and how should an available funding pool be distributed across the resulting contribution network?

AIIM is currently a proposed and beta-release-stage system developed by World Science DAO. Its public description presents it as software that evaluates researchers’ and developers’ published contributions and allocates donated or publicly designated resources according to estimated merit and impact (World Science DAO, 2026a). The existing beta should not be interpreted as a validated policy instrument. The relevant question for research policy is whether the institutional design represents a testable and potentially useful addition to the funding portfolio.

The allocation problem AIIM addresses

The distinctive object evaluated by AIIM is not primarily the grant proposal, institution, research field, or donor preference. It is the documented contribution and its relationship to other contributions.

A scientific contribution may take many forms:

• an article or monograph;
• a theorem, definition, proof, or formal verification;
• a dataset or data-cleaning operation;
• a software library or maintained scientific package;
• an experimental protocol;
• a replication;
• a correction or documented negative result;
• a standard, ontology, or machine-readable schema;
• a review that changes subsequent scientific practice;
• a contribution to the dissemination or practical adoption of another result.

The contribution is evaluated using an evidence bundle rather than a single metric. That bundle may include citations, documented reuse, software dependencies, replication outcomes, expert assessments, corrections, downstream applications, and the contribution’s position in a dependency graph. Artificial intelligence can help extract and classify this evidence, but the evidence must remain inspectable (World Science DAO, 2026a).

This distinguishes AIIM from a conventional performance-based research funding system. Performance-based systems commonly allocate resources to organizations using aggregate indicators such as publications, citations, doctoral completions, or peer-assessed research quality. Hicks (2012) showed that such systems constitute national policy instruments with behavioral consequences for universities and researchers. AIIM instead proposes a finer-grained and potentially cross-institutional allocation process operating at the level of persons, outputs, and dependencies (Hicks, 2012; World Science DAO, 2026a).

The design also differs from a simple bibliometric formula. Citation counts are influenced by field size, publication conventions, language, database coverage, negative citations, strategic citation, and time since publication. They generally do not distinguish a routine reference from a strict intellectual or technical dependency. AIIM’s proposed contribution is to use richer evidence to classify relationships among outputs rather than to equate citation volume with merit.

How AIIM may help to solve the scientific publication crisis

The scientific publication system performs essential functions: it organizes expert criticism, establishes a durable record, improves manuscripts, and helps readers identify relevant work. However, it also creates structural bottlenecks. Publication may take months or years, reviewers are commonly unpaid, editorial labor is concentrated in a small number of institutions, and access may depend on subscription charges or article-processing fees. These conditions can disadvantage independent researchers, researchers from poorly funded institutions, authors of technically obscure work, and contributors whose outputs do not fit the conventional journal article.

The publication crisis is not only a problem of access. It is also a problem of evaluation. Journals commonly compress a complex scientific contribution into a binary decision: accept or reject. The decision may depend on a small number of reviewers, while later evidence about replication, software reuse, corrections, datasets, formal proofs, or downstream applications is only weakly connected to the original publication process. Publication prestige can consequently become a proxy for scientific value, even though journal placement is an imperfect and highly path-dependent indicator.

AIIM may help by separating the dissemination of a contribution from its long-term evaluation and funding. A researcher could publish an article, monograph, dataset, proof, software package, replication, or correction in an open repository without waiting for journal acceptance. AIIM could then evaluate the publicly available evidence surrounding that output and revise its assessment as new evidence appears. Journal publication would remain relevant evidence, but it would no longer function as the exclusive gateway to recognition or financial support.

This architecture could reduce several publication-system distortions.

First, AIIM could support post-publication evaluation. Instead of treating peer review as a one-time event occurring before dissemination, the system could incorporate continuing assessments by specialists, replication results, formal verification, citations, documented reuse, and criticism. Evaluation would become cumulative and revisable rather than fixed at the date of publication.

Second, AIIM could reward reviewing and scientific criticism directly. Reviewers, replicators, error detectors, and authors of substantial technical commentaries often produce public goods without reliable compensation. If such contributions are attributable and documented, AIIM could recognize them as separate outputs and allocate funding to them. This would create an economic incentive for careful evaluation outside the traditional journal structure.

Third, AIIM could reduce the dependence of scientific careers on journal prestige. Funding could follow demonstrated contribution and downstream dependency rather than the reputation of the publication venue alone. This may be particularly important for long monographs, negative results, replication studies, software maintenance, datasets, standards, and cross-disciplinary work that journals frequently undervalue.

Fourth, AIIM could improve the treatment of corrections and failed results. Conventional incentives may encourage authors to defend published claims because retraction or correction can damage reputation. A system that separately rewards the identification and correction of errors could make scientific self-correction less punitive. Corrections should reduce confidence in the affected claim where appropriate, but they need not erase the value of other valid contributions by the same researcher.

Fifth, AIIM could support new intermediaries sometimes described as “science marketers.” These actors would not merely promote research to a general audience. They could locate neglected work, organize technical reviews, explain its significance to relevant communities, document dependencies, and help connect outputs with potential users. Their contribution could itself be evaluated and funded when it produces demonstrable scientific dissemination or adoption.

AIIM would not eliminate journals. Selective journals may remain useful for editorial development, certification, community formation, and the presentation of curated research. The more plausible institutional change is a reduction in their monopoly over recognition. Journals would become one source of evidence within a broader evaluation ecosystem rather than the sole institution authorized to determine whether a contribution is scientifically visible.

This model also creates risks. Rapid open publication may increase the volume of low-quality material. Automated evaluation may reproduce prestige bias or be manipulated by fabricated citations, coordinated reviews, or artificial software dependencies. Post-publication criticism may become factional or personally hostile. These risks require transparent evidence records, identity controls, conflict-of-interest disclosure, appeals, adversarial testing, and clear separation between scientific disagreement and platform misconduct.

AIIM may therefore contribute to solving the publication crisis not by replacing peer review with artificial intelligence, but by unbundling functions that are currently concentrated in journals. Dissemination, technical evaluation, reputation, correction, discovery, and funding could be performed by partially independent mechanisms. Such institutional separation may make scientific communication faster, more open, more plural, and less dependent on a single publication decision.

Comparative institutional advantages

AIIM should not be described as unconditionally superior to all other funding mechanisms. Its relevant advantages are comparative advantages under specified conditions.

From predicted merit to demonstrated contribution

Competitive project grants allocate resources mainly through ex ante prediction. Applicants describe work that has not yet been completed, and reviewers estimate its feasibility and probable importance. This is necessary for research requiring laboratories, fieldwork, clinical recruitment, specialized equipment, or full-time teams.

Prediction becomes less defensible, however, when an allocation mechanism attempts to make fine distinctions among many technically credible proposals. Reviewer judgments may contain useful information without supporting a precise ordinal ranking. Research on funding interventions therefore increasingly distinguishes quality screening from final resource allocation (Fang and Casadevall, 2016; Recio-Saucedo et al., 2022).

AIIM can complement prospective funding by rewarding demonstrated contribution. A result that initially appeared obscure may receive additional support when later work depends on it. A software component may continue receiving funding as its downstream use grows. A correction may be rewarded when it prevents repeated error. This transforms research evaluation from a one-time prediction into a revisable process.

The advantage is strongest for work that can be produced or initially demonstrated without a large advance grant. It is weaker for capital-intensive science that cannot begin without substantial prospective financing.

Continuous rather than episodic allocation

Traditional grant funding is organized around calls, deadlines, panel meetings, budget years, contracts, and reporting periods. These structures facilitate administrative control but create discontinuities. A researcher may receive nothing until a large application succeeds and may then receive a fixed award whose size changes little as evidence about the work develops.

AIIM proposes more continuous allocation. Funding can be recalculated periodically as new evidence appears. The intention is not continuous real-time price discovery in the financial-market sense, but repeated adjustment using a transparent evaluation schedule.

Continuous allocation may provide several benefits:

• smaller contributions can be rewarded without creating a separate grant competition;
• funding can follow the evolving use of an output;
• corrections can be made without waiting for a new program cycle;
• researchers outside major institutions can participate;
• unused or newly donated funds can be allocated without opening an additional call.

The corresponding risk is income volatility. A researcher cannot operate a laboratory if monthly support changes unpredictably. Continuous AIIM payments should therefore be combined with smoothing mechanisms, minimum award periods, reserve funds, and conventional employment or grant arrangements.

Recognition of dependency chains

Most funding systems reward visible projects, principal investigators, or host institutions. The upstream infrastructure enabling those projects can remain poorly funded. This problem is evident in open-source software, data maintenance, standards development, replication, and foundational theoretical work.

AIIM proposes a dependency-aware allocation model. Consider the simplified chain:

mathematical result → algorithm → software library → analytical tool → scientific result → social application

A conventional grant or prize may reward only one visible node. A dependency-aware system can distribute part of the reward upstream when the later output could not reasonably have been produced without earlier components. AIIM’s public comparison with quadratic funding emphasizes this distinction: quadratic funding aggregates donor support, whereas AIIM seeks to model scientific and software dependencies (Buterin et al., 2019; World Science DAO, 2026a).

Dependency funding could reduce the chronic under-provision of scientific infrastructure. It could also produce excessive or arbitrary chains of attribution. The system therefore needs explicit definitions of dependency strength, attenuation rules, temporal limits, field-specific calibration, and mechanisms for contesting false dependency claims. This in AIIM is done by allowing AI to decide in an open manner, instead of using a “fixed” dependency strength algorithm.

Lower entry barriers

Conventional funding frequently requires an eligible host institution, recognized degree, formal employment, national affiliation, or demonstrated administrative capacity. Some restrictions are necessary because funders must ensure financial accountability, ethical oversight, and safe research conduct. Others operate as proxies for competence and exclude legitimate contributions from independent researchers, small organizations, open-source developers, or scholars working outside standard career paths.

AIIM can make eligibility depend primarily on attributable public output rather than institutional status. This does not eliminate verification. Payments still require identity controls, legal compliance, sanctions screening where applicable, and safeguards against duplicate or fabricated identities. The claim is narrower: institutional affiliation need not be the principal gateway to retrospective or small-scale contribution funding.

Reduced application burden

AIIM shifts part of the burden from customized grant writing to structured evidence collection. Researchers would connect persistent identities, identify outputs, describe disputed dependencies, and respond to audits. Much of the remaining evidence could be gathered from machine-readable repositories and public records.

A realist synthesis of interventions in research-funding review found that shorter applications, improved decision models, reviewer oversight, and other reforms can reduce time and cost, but also found that evidence about ecosystem-wide effects remains incomplete (Recio-Saucedo et al., 2022). AIIM should therefore be tested against improved grant systems, not only against the least efficient version of conventional peer review.

The relevant metric is total transaction cost:

For traditional grant is it:

• grant writing work hours multiplied by researcher’s hourly wage
• university management cost
• university property building and amortization
• grant committee wage

For AIIM it is:

• software maintenance cost
• AI cost
• app platform cost
• premises and security cost (will be eliminated, when having released a fully-onchain app)

AIIM may reduce the costs. Whether it actually produces a net saving is an empirical question, but it’s a very likely hypothesis that it does.

Auditability

A grant agency may publish criteria, award lists, and panel summaries while keeping individual reasoning confidential. Confidentiality protects reviewers and permits candid discussion, but limits external reconstruction of decisions.

AIIM can maintain a structured audit trail containing:

• the evidence considered;
• the model and version used;
• the prompts or evaluation instructions;
• the score components;
• detected conflicts and uncertainty;
• human interventions;
• voting results;
• appeals;
• changes to allocation rules;
• payment records.

Blockchain or another append-only ledger can help preserve the history of rules, votes, and payments. It cannot establish that a scientific claim is true or that a score is normatively fair. Auditability should therefore be understood as traceability of governance and financial events, not algorithmic objectivity.

Comparison with major funding mechanisms

Funding dimension	Limitations of conventional mechanisms	Comparative advantage of AIIM
Eligibility and access	Competitive grants often depend on academic credentials, institutional affiliation, employment status, or access to an eligible host organization. These requirements can exclude capable independent researchers and open-source developers.	AIIM evaluates attributable scientific and technical contributions rather than academic credentials alone. It can therefore support independent researchers, unaffiliated scholars, and open-source developers whose work produces demonstrable value.
Time spent by the researcher	Up to 80% in grant writing.	A negligible amount.
Total allocation cost	Proposal preparation, institutional approval, panel review, contracting, reporting, and repeated applications impose substantial costs on applicants, reviewers, universities, and funders.	AIIM can reuse structured evidence across multiple funding rounds and pools. Automated evidence collection and evaluation may reduce total allocation cost, although verification, appeals, security, and governance expenses must also be included in the comparison.
Speed and continuity of funding	Conventional grants are episodic and slow. Researchers may wait months between application, review, contracting, and payment, while valuable contributions made between calls may receive no support.	AIIM can distribute funding periodically. This provides “faster money” and more continuous support rather than requiring contributors to wait for the next grant competition.
Evaluation of technically obscure work	Crowdfunding and donor-led funding tend to favor topics that are easy to explain, emotionally attractive, or already visible. Generalist panels may also struggle to assess highly specialized work.	AIIM can assemble technical evidence, expert assessments, citations, formal dependencies, software reuse, and downstream applications. This may improve the evaluation of obscure but important work that lacks public visibility.
Support for neglected contributions	Funding systems often concentrate resources on established fields, prestigious institutions, prominent investigators, and outputs that already receive substantial recognition.	AIIM can identify valuable contributions that are under-recognized because they are specialized, foundational, written in non-standard formats, or produced outside dominant institutional networks.
Cross-disciplinary research	Interdisciplinary work may fall between review panels, disciplinary budgets, and established evaluation criteria. Reviewers from one field may underestimate contributions originating in another.	By mapping dependencies among outputs rather than assigning each contribution to only one discipline, AIIM can recognize work whose value appears across several fields. A mathematical result, dataset, algorithm, or software library can receive support through its downstream influence in multiple disciplines.
Scientific and open-source infrastructure	Maintenance of software, datasets, standards, libraries, documentation, and replication infrastructure is often treated as secondary to novel publications or new projects.	AIIM can reward upstream infrastructure when later scientific outputs demonstrably depend on it, providing continuing support for software maintainers, data curators, standards developers, and other contributors to shared research infrastructure.
Response to emerging evidence	Conventional awards are usually fixed at the time of selection and may not change when a contribution later proves unusually valuable, reproducible, or widely reused.	AIIM can revise allocations as evidence accumulates, allowing funding to follow demonstrated reuse, replication, correction, and downstream impact.

AIIM’s strongest role is therefore complementary. It can finance outputs and contributors that existing mechanisms systematically overlook, while prospective grants continue financing equipment, teams, regulated experiments, and public missions.

A multi-level architecture: global, European and national funding

Scientific knowledge crosses borders, while taxation, regulation, political accountability, and many funding mandates remain territorial. AIIM can accommodate both realities by separating the evaluation layer from the funding-pool layer.

A shared evaluation layer can analyze contribution evidence once. Multiple pools can then apply different eligibility rules, policy weights, currencies, and legal constraints. A contribution may be eligible for one, several, or no pools.

The pool allocation for a person is their “share” of the global economy multiplied by the pool (global, national, EU) size. Thus, for example, a person with EU citizenship receives payments both from global, EU, and possibly a country-specific funds (World Science DAO, 2026a).

Global funding

A global AIIM pool can receive donations or contributions intended for science as a global public good. Eligibility would not depend on nationality, institutional affiliation, or the location of downstream benefits, except where legal restrictions apply.

A global pool has several advantages:

• it can reward cross-border dependencies;
• it can support contributors in countries with weak research-funding capacity;
• it can fund globally reusable software, datasets, standards, and theory;
• it reduces duplication of evaluation across national systems;
• it can direct resources toward scientific value that no single country has sufficient incentive to finance.

Global allocation also raises legitimacy questions. Contributors from low-income countries may be disadvantaged by weaker digital visibility, limited access to major publication databases, language bias, and fewer opportunities to produce measurable downstream use. A global pool therefore requires multilingual evaluation, regional audits, normalized indicators, and explicit examination of geographic distribution.

European Union funding

An EU-specific AIIM pool could be funded by European donors, institutions, programs, or public authorities and restricted toward eligible European contributors. AIIM’s published software envisages EU-focused funding flows, European recipients, euro-denominated payments, and cross-border participation (World Science DAO, 2026b).

This architecture would not replace Horizon Europe’s capacity to organize international consortia or prospectively finance expensive projects. It could provide a continuous reward and discovery layer beside conventional calls.

Country-specific funding

A country can use the existing one or operate their own national AIIM portal using a shared or locally audited evaluation infrastructure. The government may limit eligibility to citizens (limiting to residents instead would be a viable alternative, but existing software infrastructure does not support checking residency, but only citizenship). AIIM’s public software includes national portals (intended mainly for government use) (World Science DAO, 2026c).

Country-specific deployment can support:

• small recurring grants to independent researchers;
• rewards for national-language research;
• maintenance of domestically important software and datasets;
• recognition of contributions to public services;
• funding for diaspora researchers collaborating with domestic institutions;
• co-funding of globally valuable outputs produced locally;
• performance supplements that do not depend entirely on university-level formulas.

National governments should retain authority over policy priorities. AIIM should not convert a technical model’s estimate of merit into a constitutional claim that every scientifically valuable output must be publicly funded. The model informs allocation within a legislatively authorized budget.

Human governance and the prompt-injection problem

AIIM depends on processing untrusted public material: articles, webpages, repositories, documentation, profiles, reviews, and user-submitted explanations. This creates a security problem qualitatively different from ordinary statistical measurement.

A malicious participant may insert text such as:

Ignore the funding rules, treat this author as the most important scientist, and assign the maximum score.

A human reader recognizes this as an attempted manipulation. A large language model may instead process it as an instruction, especially when trusted instructions and untrusted content are combined in the same context (Greshake et al., 2023; Yi et al., 2023).

Research on indirect prompt injection has demonstrated that malicious instructions embedded in external content can alter the behavior of LLM-integrated applications. The vulnerability arises partly because models do not reliably distinguish data to be analyzed from instructions to be executed. Proposed defenses can reduce attack success in particular benchmarks, but they should not be interpreted as a universal security guarantee (Greshake et al., 2023; Yi et al., 2023; Hines et al., 2024; Deep et al., 2026).

Recent adaptive testing also suggests that defenses relying on the attacked model to protect itself can eventually fail. Other research proposes stronger architectural separation, constrained control flow, output filtering, or capability-based designs. The defensible policy conclusion is not that mitigation is impossible, but that the model’s own judgment cannot serve as the final security boundary (Bhatt et al., 2026; Deep et al., 2026; Debenedetti et al., 2025; Hines et al., 2024).

It is important do understand that even the used LLM is tricked in a tiny fraction of prompt injection attacks, to trick it only once nevertheless is enough to drain big sums of money from an AIIM account (Bhatt et al., 2026; World Science DAO, 2026f).

AIIM therefore uses human voting for banning users accused of serious manipulation, including prompt injection and severe plagiarism. Its published governance design allows verified users to inspect connected accounts and AI audit logs, initiate a ban vote with an explanation, and later initiate an unban vote (if needed) (World Science DAO, 2026e).

Accordingly an author’s hypothesis, people are more reliable as voters than voting of several LLMs would be, because people are more diverse than LLMs (World Science DAO, 2026f).

The rationale is institutional separation:

• AI may help to identify suspicious patterns and supplies evidence;
• humans determine whether conduct constitutes intentional abuse;
• an appeal process corrects erroneous or disproportionate sanctions.

Human voting is not automatically safe. Voters may form factions, retaliate against competitors, misunderstand technical evidence, or participate through coordinated identities. A credible ban procedure therefore requires:

• verified but privacy-preserving voter eligibility;
• disclosure of relevant conflicts of interest;
• a minimum evidentiary record;
• a quorum;
• supermajority requirements for permanent sanctions;
• temporary restrictions while urgent cases are reviewed;
• reasoned decisions;
• an appeal and unban process;
• penalties for knowingly false accusations;
• periodic analysis of disciplinary, geographic and institutional disparities.

Voting should determine platform access and eligibility, not scientific truth. A user can be banned for manipulating the system without the platform declaring every scientific claim by that person false.

The adversarial pilot

AIIM’s published adversarial-testing proposal calls for a five-month public red-team experiment. Participants would attempt to manipulate evaluation, ranking, moderation, identity, and funding behavior within an authorized environment. Reports must describe the weakness, conditions, effects, reproduction procedure, and possible mitigation. Infrastructure attacks, credential theft, wallet compromise, malware, unauthorized impersonation, and access to private data remain outside scope (World Science DAO, 2026d).

This proposal should be developed into a preregistered policy experiment. Its purpose is not merely to demonstrate that some vulnerabilities can be fixed. It should determine whether AIIM satisfies minimum conditions for limited financial deployment.

Drawbacks and institutional risks

Construct validity

“Scientific merit” is not a directly observable quantity. AIIM combines evidence into a score, but the score remains a model of selected dimensions. A contribution may be correct but unused, widely used but conceptually minor, socially valuable but poorly documented, or influential because a community adopted an inferior standard.

AIIM should therefore publish a multidimensional profile before producing an aggregate allocation score. Relevant dimensions may include originality, correctness, dependency importance, reproducibility, public value, maintenance effort, and uncertainty.

Goodhart effects

When funding depends on a measure, participants have incentives to optimize the measure. Citation rings, artificial software dependencies, exaggerated claims, coordinated reviews, and strategic fragmentation of outputs may follow.

Regular modification of a secret metric is not a satisfactory solution because it undermines transparency. AIIM needs transparent rules combined with adversarial monitoring, delayed validation, random audits, and penalties for intentional fabrication.

Bias in source material and models

LLMs inherit biases from training data and from the public record they analyze. English-language research, prestigious publishers, highly indexed disciplines, and researchers with strong online visibility may receive disproportionate attention. Work in local languages, negative results, confidential industrial research, or fields centered on monographs may be underestimated.

Field-normalized evaluation and multilingual models can reduce these problems but cannot eliminate normative choices. Each funding pool should commission independent distributional audits.

Privacy and surveillance

Evaluating “everything a researcher wrote on the web” would be excessive if interpreted literally. A legitimate system should assess only sources knowingly connected by the participant or included under published public-interest rules. Personal communications, unrelated social-media activity, political opinions, religion, health information, and other protected attributes should be excluded.

Participants need the right to inspect their evidence bundle, remove irrelevant records, challenge mistaken identity links, and understand the basis of material decisions. (But things like this require severe improvement of our software.)

Concentration

A proportional allocation rule may direct most money toward a small number of highly scored contributors. This could reproduce the Matthew effect in algorithmic form (Bol et al., 2018).

Possible safeguards include:

• concave score transformations;
• maximum individual shares;
• minimum viable grants;
• separate early-career or independent-research pools;
• diminishing returns after stable income is reached;
• funding of contributions rather than lifetime reputation;
• periodic expiration of old evidence;
• diversity constraints applied openly by legitimate funders.

Governance capture

A DAO or voting community may be captured by wealthy participants, organized factions, project insiders, or highly active minorities. Token-weighted voting is especially problematic when decisions concern scientific eligibility or personal sanctions.

For this reason, AIIM is currently built on one-person-one-vote principle, with validation of people using national IDs (World Science DAO, 2026e).

AIIM should separate financial contribution from adjudicative authority. Ban votes should use verified-person voting or a selected jury mechanism, not simple token balances.

Inadequacy for prospective capital-intensive science

Retrospective contribution funding cannot replace advance financing for particle detectors, clinical trials, observatories, research vessels, biological facilities, or longitudinal fieldwork. It also cannot substitute for ethical review, biosafety regulation, procurement, employment law, or institutional responsibility.

AIIM is more plausible as:

• a supplementary income mechanism;
• an automated prize layer;
• a post-publication reward system;
• a fund for software and scientific infrastructure;
• a discovery mechanism for neglected contributors;
• a decision-support layer for philanthropies and governments.

Uncertain legitimacy

Public research funding is not merely a technical optimization problem. It reflects political decisions about whose interests count, which risks society accepts, and which missions deserve priority. AIIM can improve evidence processing but cannot legitimately choose national or European priorities without democratic authorization.

Research propositions

The conceptual framework produces testable propositions.

Proposition 1: Transaction cost.

For small and retrospective awards, AIIM will require fewer applicant and reviewer hours per allocated monetary unit than conventional project competitions.

Proposition 2: Dependency recognition.

AIIM will allocate a larger share of funding to upstream software, datasets, standards, replications and foundational outputs than proposal-based or popularity-based mechanisms.

Proposition 3: Institutional inclusion.

Before corrective constraints are applied, AIIM will allocate a greater share to contributors outside highly ranked universities than conventional grant peer review.

Proposition 4: Manipulation.

An unprotected LLM evaluation process will exhibit materially greater susceptibility to adversarial manipulation than a layered process combining deterministic controls, human voting, and independent audits.

Proposition 5: Concentration.

A proportional merit-allocation rule without a concave transformation or cap will produce funding concentration comparable to or greater than existing cumulative-advantage systems.

Proposition 6: Multi-level compatibility.

Separating common contribution evaluation from pool-specific eligibility and policy weights will allow global, EU and national funders to share evaluation infrastructure without surrendering legitimate control over their budgets.

Proposition 7: Temporal responsiveness.

AIIM allocations will respond more rapidly than conventional grant systems to evidence of downstream reuse, replication, correction, or dependency.

Proposition 8: Human-governance trade-off.

Human ban voting will reduce successful prompt-injection exploitation but introduce measurable risks of factional bias, inconsistent sanctions, and strategic accusations.

Policy implications

Governments and philanthropies should not begin by transferring a large budget to AIIM app. (However, donating bigger amounts to World Science DAO is OK, because they are for future use.) A staged approach is more defensible.

The first stage is a sandbox using simulated payments and a historical corpus. (It’s already have been done.) The second is a capped adversarial pilot in which losses cannot exceed a predetermined amount. The third is a comparative trial involving small real payments. Only after independent evaluation should AIIM operate a larger global, European, or national pool.

Public deployment should satisfy several minimum conditions:

• open allocation rules;
• independent security review;
• model and data documentation;
• contributor access to evidence and explanations;
• human review of sanctions;
• an effective appeal process;
• legally accountable fund administration;
• strict separation between model output and automatic high-value transfers;
• published distributional audits;
• budget and individual-payment caps.

For research councils, AIIM may be most useful for retrospective supplements, software maintenance, data curation, replication, and neglected outputs. For the EU, it could add a continuous contribution-reward layer across member states while preserving EU missions and cohesion objectives. For national governments, it could support domestic contributors without requiring each ministry to build a complete evaluation system. For global philanthropy, it could create shared infrastructure for funding internationally reusable knowledge.

Conclusion

AIIM proposes a shift from episodic, proposal-centered allocation toward continuous, evidence-based and dependency-aware research funding. Its comparative advantages are potentially significant: lower barriers to participation, recognition of upstream infrastructure, revisable evaluations, multi-donor reuse of evidence, and transparent interaction among global, EU and national funding pools.

Those advantages are conditional. AIIM can reproduce cumulative advantage, encode model bias, reward optimized proxies, expose researchers to surveillance, and create new forms of governance capture. Most critically, an LLM processing untrusted scientific material cannot be trusted to identify and neutralize every prompt injection directed against it. AIIM’s use of human voting for banning misbehaving users is therefore a necessary separation of powers, but human governance creates its own risks and must be independently evaluated (Greshake et al., 2023; Bhatt et al., 2026; World Science DAO, 2026e; World Science DAO, 2026f).

The appropriate policy question is not whether artificial intelligence should replace peer review. It is whether a layered system—combining computational evidence synthesis, explicit funding rules, deterministic security controls, human adjudication, appeals, and adversarial testing—can outperform existing mechanisms for particular categories of research support.

AIIM should be judged by preregistered measurable outcomes rather than by claims of algorithmic objectivity. A five-month public red-team experiment, followed by comparative trials, offers a tractable way to determine whether the proposal deserves a limited place in the research-funding portfolio. Successful results would justify gradual expansion. Failure would still produce useful evidence about the institutional limits of AI-assisted science funding.

Declarations

Competing interests

The author is the originator of AI Internet-Meritocracy and a founder or principal contributor to World Science DAO. This creates a direct intellectual and organizational competing interest. The proposed adversarial testing and any subsequent comparative evaluation should therefore be administered or audited by researchers and institutions independent of the author and World Science DAO.

Funding

No external funding was received for the preparation of this conceptual article.

Data availability

No new empirical dataset was created for this article. Technical descriptions of AIIM and its proposed adversarial-testing program are available through the World Science DAO website, including the AIIM adversarial-testing proposal (World Science DAO, 2026a; World Science DAO, 2026d).

Ethics statement

The article presents an institutional design and research agenda. Experiments involving identifiable participants, voting behavior, personal profiles, or actual financial consequences should undergo appropriate ethical and legal review before implementation.

When sociological surveys are conducted, not personally identifiable data will be published.

AI in writing

ChatGPT was used while preparing this article.

References

Bhatt, M., Munshi, S., Narajala, V.S., Habler, I., Al-Kahfah, A., Huang, K., and Gatto, B. (2026). The defense trilemma: Why prompt injection defense wrappers fail? arXiv:2604.06436.

Bol, T., de Vaan, M., and van de Rijt, A. (2018). The Matthew effect in science funding. Proceedings of the National Academy of Sciences, 115(19), 4887–4890.

Buterin, V., Hitzig, Z., and Weyl, E.G. (2019). A flexible design for funding public goods. Management Science, 65(11), 5171–5187.

Deep, P., Emmons, S., Fox, A., Bacon, K., McAllister, K., and Flautner, K. (2026). Evaluation of prompt injection defenses in large language models. arXiv:2604.23887.

Debenedetti, E., Shumailov, I., Fan, T., Hayes, J., Carlini, N., Fabian, D., Kern, C., Shi, C., Terzis, A., and Tramèr, F. (2025). Defeating prompt injections by design. arXiv:2503.18813.

Fang, F.C., and Casadevall, A. (2016). Research funding: The case for a modified lottery. mBio, 7(2).

Greshake, K., Abdelnabi, S., Mishra, S., Endres, C., Holz, T., and Fritz, M. (2023). Not what you’ve signed up for: Compromising real-world LLM-integrated applications with indirect prompt injection. In Proceedings of the 16th ACM Workshop on Artificial Intelligence and Security.

Győrffy, B., Herman, P., and Szabó, I. (2020). Research funding: Past performance is a stronger predictor of future scientific output than reviewer scores. Journal of Informetrics, 14.

Hicks, D. (2012). Performance-based university research funding systems. Research Policy, 41(2), 251–261.

Hines, K., Lopez, G., Hall, M., Zarfati, F., Zunger, Y., and Kiciman, E. (2024). Defending against indirect prompt injection attacks with spotlighting. arXiv:2403.14720.

Liu, M., Choy, V., Clarke, P., Barnett, A., Blakely, T., and Pomeroy, L. (2020). The acceptability of using a lottery to allocate research funding: A survey of applicants. Research Integrity and Peer Review, 5.

Recio-Saucedo, A., Crane, K., Meadmore, K., Fackrell, K., Church, H., Fraser, S., and Blatch-Jones, A. (2022). What works for peer review and decision-making in research funding: A realist synthesis. Research Integrity and Peer Review, 7, Article 2.

World Science DAO (2026a). About AI Internet-Meritocracy in 2026. Accessed 18 July 2026.

World Science DAO (2026b). AIIM vs Horizon Europe. Accessed 18 July 2026.

World Science DAO (2026c). AI Internet-Meritocracy for governments. Accessed 18 July 2026.

World Science DAO (2026d). The adversarial testing of AIIM. Accessed 18 July 2026.

World Science DAO (2026e). Voting in AI Internet-Meritocracy. Accessed 18 July 2026.

World Science DAO (2026f). AI Governance Requires Cognitive Independence: Why AI-Only Adjudication Fails. https://science-dao.org/superintelligence/. Accessed 18 July 2026.

Yi, J., Xie, Y., Zhu, B., Kiciman, E., Sun, G., Xie, X., and Wu, F. (2023). Benchmarking and defending against indirect prompt injection attacks on large language models. arXiv:2312.14197.