Commons Collective GitHub

Commons Collective / open institutional research

Commons Collective: Institutions for Shared Agency in the AI Data Economy

Date: 2026-09-19 · Version: 0.1
Status: AI-assisted discussion draft; not peer-reviewed. Human authorship, review, and approval pending. No participant endorsement is claimed.
Project: Commons Collective / proposed Commons Collective Federation (CCF)
Project URL: https://ethical-tech-colab.github.io/commons-collective/

Abstract

AI systems depend on shared knowledge whose reproduction is inexpensive but whose maintenance, legitimate governance, and renewal are not. This paper proposes Commons Collective as an institutional response that aggregates bounded negotiating mandates rather than indiscriminately pooling data. Five claims organize the argument: nonrival knowledge requires support without enclosure; purchasing, contributor bargaining, and public-good representation exercise different powers; federation should preserve accountable local authority and exit; maintained services can finance stewardship without creating royalties over existing open permissions; and provenance, contextual evaluation, and public participation are contestable mechanisms rather than certificates of truth or legitimacy. Drawing on an existing scoping review, legal and institutional sources, and supplied workshop materials, the paper develops a separated-function architecture and a narrow paid pilot. The proposal is unvalidated. Its success would require measurable benefits after governance costs, meaningful refusal, preserved public access, and independent rights-based stopping rules, not simply more members, collected data, or payments.

1. Research question and evidence boundary

The question is whether collective institutions can improve people's agency over AI services and contributions while sustaining knowledge available to everyone. Individual consent and small payments are incomplete answers: an individual transaction can affect nonparticipants, and neither payment nor permission maintains a shared resource by itself. Relational data-governance scholarship supplies the central objection to contributor-only representation: population-level effects extend beyond those who supplied the records. [S25]

This paper is a selective institutional-design synthesis of the existing Commons Collective report, source registers, replication blueprint, and supplied workshop materials. It is not a systematic review, legal opinion, or report of completed fieldwork. No new broad literature search was undertaken for this draft. Economic theory, normative scholarship, legal instruments, institutional self-descriptions, and empirical studies are distinguished below; none demonstrates that CCF already works.

The contribution is five connected propositions, developed in Sections 2–6, and a design that could disconfirm them in practice. CCF remains a proposed coordination framework, not an incorporated or authorized representative. The intended advance is not a universal organizational form for AI, but a way to determine which collective functions justify their costs in a particular setting.

2. Claim one: nonrival knowledge still requires costly stewardship

Copying information normally leaves the original available. Jones and Tonetti model gains from such nonrival use and inefficiencies associated with hoarding; their model does not establish that unrestricted sharing is always beneficial. [S31] The institutional problem is therefore not identical to grazing on subtractable pasture. What becomes scarce includes maintenance labor, attention, validation capacity, privacy, and legitimate authority to make consequential decisions.

This distinction changes the financing argument. AI developers may benefit from maintained archives, documentation, and reference resources without adequately supporting their renewal. But that possibility does not establish a universal royalty right over everything used by AI. Public facts have no single collective owner; existing licenses, exceptions, and other rights differ.

Ostrom's polycentric approach motivates participation, monitoring, conflict resolution, and nested institutions suited to context, rather than a choice between enclosure and unmanaged access. [S03] Applied here, stewardship should maintain the resource and its governing capacity while preserving lawful reuse. A voluntary commons-benefit covenant could purchase maintenance or commit a buyer to public-benefit funding. It would bind its contracting parties, not invent obligations for all downstream users. Any mandatory levy would be a separate policy proposal requiring its own legal and distributive justification.

3. Claim two: aggregate mandates, not indiscriminate data

Collective scale has at least three distinct meanings. Demand-side procurement aggregates purchases of AI assistance and related services. Members can negotiate retention limits, accessibility, portability, prices, and remedies without supplying training data. Its leverage rests partly on credible alternatives and switching capacity.

Contributor bargaining concerns specific rights or services that contributors can actually authorize. The institution needs evidence of authority and a mandate identifying material, purpose, recipient, duration, payment, and remedies. Possessing an agent log, for example, does not authorize licensing every message or professional work within it. Data-leverage scholarship identifies withholding and contributions to alternatives as possible strategies, not demonstrated bargaining gains for every dataset. [S06]

Public-good representation concerns maintenance and affected people who may neither purchase the service nor own a valuable input. It cannot be reduced to a dividend allocated among sellers. This third function follows from the relational interests that individual transactions leave unrepresented. [S25]

Delegation should therefore be aggregated before custody. A register of bounded mandates can support negotiation while source material remains with an authorized custodian. Data-trust scholarship offers a model of purpose-sensitive delegation, not a universal property right or automatically effective fiduciary regime. [S07] Where applicable, GDPR duties remain distinct from copyright permissions; withdrawing consent does not retroactively invalidate prior lawful processing. [S10]

Contribution and sustained withholding also differ. Multihoming, or joining several collectives, can protect choice, but overlapping mandates may weaken exclusive commitments. Coordinated refusal requires organizing capacity, participant support, a lawful target, duration, and protected dissent. It cannot retract compliant CC BY permissions or promise removal from already trained models. [S26] European competition guidance for specified solo-self-employed working conditions is not blanket immunity for data licensing. [S13] Thus a procurement pilot need not depend on a speculative economy-wide data strike.

4. Claim three: federation must distribute authority, not merely infrastructure

The proposed CCF should coordinate interoperable records, shared procurement support, and assurance methods across autonomous collectives. It should not become the compulsory route to public knowledge. Existing libraries, associations, or community institutions should host the functions where legitimate and capable; new incorporation is justified only by a documented gap.

The architecture separates responsibilities, not necessarily corporations:

Function Delegated responsibility Accountability boundary
Member purchasing body Select services and negotiate bounded procurement contracts No automatic contributor-licensing authority
Contributor representation Verify and negotiate specified mandates Separate accounts, refusal paths, and conflict rules
Commons and affected-person council Set stewardship priorities and protected public-access rules Includes noncontributors; cannot waive third-party rights
Replaceable service operator Deliver contracts, support, records, and corrections Published charges; no unilateral mandate changes
Independent assurance and redress Inspect authority, finances, outcomes, and complaints Funded separately from favorable findings; power to pause affected activity

Local bodies retain membership rules, treasury, and lawful duties. Shared mandates are transaction-specific; federation membership is nonexclusive. Portable records, open interfaces, competing operators, and funded exit make replacement practical rather than merely nominal. Payment handling requires appropriate legal safeguards and regulated providers where applicable.

Governance should combine accountable member representatives, supplier voice, and compensated affected-person participation. Purchasing volume and contributed data should not buy extra constitutional votes. Conflict disclosures, recall, written decisions, independent appeal, and protection for minority rights are essential. Sponsors and operators cannot veto adverse assurance findings.

This is a demanding arrangement. Its justification is functional: the buyer, seller, operator, and public face different incentives. Separation helps expose conflicts but does not remove them. A federation is preferable only if its benefits exceed coordination costs and its members can contest or leave it without losing existing public access.

5. Claim four: sell maintained services without enclosing the commons

An open resource and a paid service are not mutually exclusive. Wikimedia Enterprise's published principles distinguish paid delivery services from restrictions on free use and customer editorial control. This is an institutional example, not independent evidence that the proposed collective will be profitable. [S83]

CCF participants could sell reliable updates, accessibility work, integration, provenance diligence, evaluation, support, or commissioned curation. Buyers would pay for accountable labor and delivery, not permission retrospectively imposed on public-domain works. CC BY 4.0 grants a royalty-free, nonexclusive, irrevocable license under its terms; collective membership cannot rewrite those existing permissions. [S26]

The financial architecture should distinguish operator earnings, contributor payables, and restricted commons funds. Money collected for contributors is not unrestricted operating revenue. Grants can support public interfaces, participation, and underserved resources, but must not be reported as proof of recurring customer demand. Service fees should cover delivery, administration, assurance, reserves, and exit rather than relying on hidden volunteer labor.

A public core of quality observations and correction records could coexist with paid freshness guarantees or integration. Paying must never purchase a favorable rank. The relevant viability test is whether customers renew useful services at a cost that sustains workers and safeguards public access. This draft offers no revenue forecast or universal self-financing claim; some socially valuable resources will require continuing public or philanthropic support.

6. Claim five: evidence and participation are mechanisms, not certificates

Shared records should distinguish provenance, authority claims, task validation, and reputation. Provenance describes a recorded history; it does not establish the history's truth. A signature identifies the signer, not a rightful owner. Validation supports a specified use and version, not every future application.

The proposed mechanism is an inspectable, challengeable evidence record: object version, asserted authority, evaluator, method, task, date, conflicts, limitations, and dispute status. Sensitive evidence remains controlled. Ranking first applies mandatory permission and safety eligibility checks; only eligible offerings are compared for task-specific utility, reliability, freshness, and cost. Missing evidence is not equivalent to poor quality, and popularity cannot cure invalid authority. Competing evaluators and appeals prevent an operator's commercial incentives from becoming an unquestionable quality judgment.

Public participation can also shape evaluation rubrics and training objectives. Collective Constitutional AI connects a bounded public-input process involving 1,002 U.S. adults to model training and comparative evaluation. Its screened, U.S.-focused participation and interpretation choices do not establish global representation or democratic legitimacy. [S56] The result motivates experiments, not a claim that public preferences can simply be encoded once.

A collective could pay participants for deliberation, retain disagreement, publish how proposals changed a rubric, and commission independent held-out evaluation. Participation must have demonstrable decision consequences. An optimization reward is not contributor compensation; a numerical reward weight is not a measurable share of democratic control. Rights cannot be traded off against a higher aggregate performance score. Any future regulation of participation or outcomes would need separate justification; no universal commons-reward quota is asserted as enacted law here.

7. Research design: one paid pilot with disconfirming outcomes

The first test should be a six-month, low-risk service contract hosted by an existing library consortium in one jurisdiction. A small member group would purchase an AI-assisted reference service over an already shareable, nonpersonal collection. The paid bundle would include maintained metadata, attributable answers, correction handling, and accessibility support. Commissioned annotations would have explicit terms; private patron records and bulk agent traces would be excluded.

Before supplier selection, publish the baseline alternative, evaluation protocol, lawful authority review, cost categories, participant recruitment, and stopping rules. A procurement comparison must include legal review, assurance, member time, accessibility, migration, payment fees, and funded exit. Grants and volunteered labor should be disclosed separately. An executed paid contract tests demand more directly than expressions of interest, but one subsidized contract cannot validate a general business model.

Three proposed decision tests make the pilot falsifiable:

  1. Net service benefit: by day 180, quality-adjusted total cost is at least 10% below a documented comparable alternative, without worse access or service quality. Specify quality measures before bids; missing baseline evidence means the claim is untested.
  2. Effective agency: at least 90% of participating adults correctly explain their mandate and withdrawal rule. Exercise supplier-switching, mandate-withdrawal, and complaint drills; publish completion times and failures.
  3. Accountable stewardship: reconcile protected funds, deliver contracted maintenance, preserve existing open permissions, and report whether affected-person deliberation changed an actual decision.

These are proposed thresholds, not predicted effects. Comparable workloads and independently assessed answer samples should test service quality; subgroup results, uncertainty, attrition, and missing observations must be reported. A small pilot mainly establishes feasibility, not population-wide causality or contributor bargaining power.

At day 90, independent reviewers require complete authority records, reconciled funds, working redress, and a successful withdrawal drill. Immediately suspend affected activity for invalid authority, serious confidentiality failure, unremedied rights violations, unlawful coordination, or evaluator manipulation. Economic failure blocks expansion unless a separately justified public-benefit subsidy is approved. Rights failures cannot be offset by savings. Publish negative results and preserve lawful public access during wind-down.

8. Limitations and counterarguments

The strongest objection is that institutional overhead may consume the gains. Existing procurement consortia or public funding could perform better than a new federation. Large buyers may dominate despite formal voting equality; suppliers may multihome without offering scarce inputs; affected-person councils may become ceremonial. These are reasons to compare alternatives and observe exercised power, not merely adopt a charter.

Service revenue may favor wealthy customers and popular languages while neglecting valuable low-demand resources. Independent allocation and public subsidy can address that gap only if funded and accountable. Mandate review, privacy protection, and legal compliance may make some offerings impractical. Open publication of governance evidence can itself expose vulnerable people unless carefully minimized.

The evidence base is selective, largely Anglophone, and uneven across jurisdictions. No observed CCF bargaining result, validated willingness-to-pay estimate, or demonstrated representative legitimacy is available. Pilot success would justify another bounded test, not establish that all AI data deserve royalties or all future AI requires this organizational form.

9. Conclusion

Commons Collective is most defensible as a proposal for shared agency over services, authorized contributions, and maintenance priorities, not a claim to own the commons. Its central test is whether bounded delegation, functional separation, and contestable evidence improve real choices without enclosing open knowledge. A paid, stoppable pilot can reveal whether those institutional commitments survive their costs. Failure should narrow or reject the model rather than become an argument for compulsory participation.

Provenance and license

This draft draws on session-supplied workshop directions and the provided note excerpt recorded as [S97] and [S99]. It does not represent authenticated review of the complete Google Doc or its other tabs. Source attributions do not establish authorship, review, approval, or endorsement of this paper by named workshop participants.

Original draft text is licensed under Creative Commons Attribution 4.0 International (CC BY 4.0): https://creativecommons.org/licenses/by/4.0/. Cited and other third-party materials retain their own terms. Human authorship and publication approval remain pending.

Bibliography