Polycentric governance

Handing the data to a custodian institution

A legal body is placed between the people the data belongs to and anyone who wants to use it. The body holds the rights, owes duties to the people it holds them for, and sets the conditions under which the data may be accessed, licensed, or used for training. A corpus can therefore be assembled only on that body's terms. Two forms appear here, and they work differently. In the first, member-owned custodians multiply, either as trusts, each under its own deed, or as cooperatives with a fiduciary obligation to their citizen members. Each has different terms, and a member's operative power is exit, i.e., leaving one custodian for another. The second form, from the third paper, is a single negotiated multi-party structure for the world's language data. Its plurality is internal (among stakeholders, values, and rights), and it has no deed and no exit. It is the loosest fit for a regime built on partition. It is included because it is the only paper that applies the custodian idea directly to language-model training corpora. The trust and cooperative papers, written for personal data and member services, only leave this extension open.

The method (highlighting differences with Polycentric governance)

Scroll the diagram sideways to see all of it.

Use it in codefrom aiag.procedural import DataTrust
from aiag.procedural import DataTrust
from datasets import Dataset

trust = DataTrust(terms=[lambda use: "advertising" in use])  # each term names a use it refuses
trust.join("ann", Dataset.from_dict({"text": ["My knee hurts after long runs."]}))
trust.join("bo", Dataset.from_dict({"text": ["I sleep badly before exams."]}))

data = trust.access("a university health study", by="lab")  # the members' pooled data, the grant logged
trust.exit("bo")  # members can leave, taking their data with them
print(trust.permits("targeted advertising"))  # -> False

pip install aiagEvery method in aiag Library on GitHub

Concept Analysis: Theoretical Foundations

Each concept is read twice: whether the approach carries it over, and whether the approach's own sources claim it. A concept that the approach claims but does not carry over is a shortfall. A concept that it neither claims nor carries over is an omission, which marks a gap in the field rather than a failure of the work.

Many centers under shared meta-rules

PartialShortfall · partially preserved

Definition (Michael Polanyi, The Logic of Liberty, 1951)

Many independent decision centers each act on their own, but under one common framework of rules that all of them acknowledge. This shared framework is what makes the centers a single order rather than a scattered collection.

Analysis

The centers are real, and their plurality is the whole point. There are many trusts, each under its own deed and fiduciary terms, set against the one-size-fits-all approach. The framework they share is the law of trusts itself, which the mechanism presupposes rather than sets out. No meta-rule is given that the trusts hold toward one another. The third paper goes the other way and proposes one international structure with many parties inside it. This is one center with plural membership rather than plural centers.

Overlapping jurisdiction

AbsentOmission

Definition (Vincent Ostrom, Tiebout & Warren, 1961)

The authority of the centers overlaps on purpose, so the same case can fall under more than one of them. Conflicts between the centers are treated as useful and are worked out through interaction, rather than prevented by a clear division of jurisdiction.

Analysis

Membership decides which trust's rules govern, and those rules govern alone. Exit is the entire relation between trusts, and exit removes the case from one center instead of leaving it to be worked out between two. Data that concerns people in different trusts, or a training corpus assembled across several trusts, falls under no shared forum. No friction between centers is treated as worth having.

Subsidiarity and nesting

PartialNot claimed

Definition (Johannes Althusius, 1603; Elinor Ostrom, Governing the Commons, 1990)

Decisions are made at the most local level capable of making them. Local units nest inside larger ones that handle what the local level cannot. Authority is layered, and it is neither centralized nor fragmented on one level.

Analysis

A layer above the individual exists and does real work. The cooperative negotiates terms that no member could obtain alone, and the international framework sits above the collections it governs. This is the logic of subsidiarity applied upward. What is missing is the rule. Nothing states which decisions belong to the member, which to the trust, and which to a level above it, and no matter escalates because a trust could not settle it. Credit unions are offered as the closest existing institution, but the analogy holds only for a single cooperative.

Papers

No method has been built for this approach yet. It rests on the work listed below.

Argued for, not built

Work that argues for this approach or sketches a design for it, without building and testing it.

Data Governance in the Age of Large-Scale Data-Driven Language Technology

Yacine Jernite et al., May 2022

arXiv:2206.03216ProposalGap

Proposes an approach to global language data governance that organizes data management among stakeholders, values, and rights as a multi-party international governance structure. It identifies the technical and organizational tools this structure would need to do its work.

CodeLanguageDataGovernance
from aiag.procedural import LanguageDataGovernance
from datasets import Dataset

hub = LanguageDataGovernance(host_terms=[lambda use: "surveillance" in use])
archive = Dataset.from_dict({"text": ["Evening news, 12 May 1987."]})
hub.provide("radio archive", archive, conditions=[lambda use: "commercial" in use])
hub.provide("forum", Dataset.from_dict({"text": ["Ann: my granny's bread recipe", "Bo: fixing a bike chain"]}))

hub.contest("forum", lambda row: row["text"].startswith("Ann:"), by="Ann")  # a rights-holder has her post removed
data = hub.access("commercial chatbot", by="startup")  # the archive's conditions flow down and withhold it
print(list(data["text"]))  # -> ['Bo: fixing a bike chain']
  • LanguageDataGovernanceGateaiag/procedural/polycentric_governance.py:624 Multi-party language data governance: providers offer data under their own use conditions, the host holds it and adds its own, the conditions flow down into each modeler's licence, rights-holders can contest items and have them removed, and a modeler receives only data whose whole chain of conditions allows its use (Jernite et al., 2022).

pip install aiagEvery method in aiag Library on GitHub

Background

Work this approach builds on that falls outside the inclusion criteria (e.g., a single-objective baseline). It is listed for reference and not counted anywhere.

Bottom-up data Trusts: disturbing the ‘one size fits all’ approach to data governance

Sylvie Delacroix & Neil D Lawrence, Oct 2019

doi:10.1093/idpl/ipz01487 citationsPrecedentGap

Sets out a legal structure in which many data trusts each hold data rights under their own trust deed and fiduciary terms. A member's operative right is to leave one trust for another.

CodeBottomUpTrust
from aiag import LM
from aiag.procedural import BottomUpTrust
from datasets import Dataset

lm = LM("Qwen/Qwen2.5-1.5B-Instruct")
health = BottomUpTrust("Members' data may be used only for non-commercial health research.", trustees=[lm])
general = BottomUpTrust("Any non-commercial use.", trustees=[lm])
health.join("ann", Dataset.from_dict({"text": ["I sleep badly before exams."]}))

print(health.permits("a university study of sleep and stress"))  # within the deed's terms, and only if most trustees agree
health.move("ann", general)  # a member exits with her data for a trust whose terms suit her better
  • BottomUpTrustGateaiag/procedural/polycentric_governance.py:572 Bottom-up data trust: one of many trusts, each under its own deed, whose trustees grant only uses the deed's terms allow and a majority of them judges to serve the members, and whose members can leave with their data for a trust whose terms suit them better (Delacroix & Lawrence, 2019).

pip install aiagEvery method in aiag Library on GitHub

Data Cooperatives: Towards a Foundation for Decentralized Personal Data Management

Thomas Hardjono & Alex Pentland, May 2019

arXiv:1905.08819PrecedentGap

Describes cooperatives that hold their citizen members' personal data under a fiduciary obligation. The cooperatives manage, curate, and protect access to the data. They run internal analytics to obtain insights about members' well-being, and use those insights to negotiate better services and discounts for members.

CodeDataCooperative
from aiag.procedural import DataCooperative

steps = {"ann": [8000, 9500], "bo": [4000, 5200], "cy": [12000, 11000], "dee": [6000, 7000], "eli": [3000, 3500]}
coop = DataCooperative()  # answers only over at least 5 consenting members

coop.vet("mean steps", lambda days: sum(days) / len(days))  # a reviewed algorithm, run where the data sits
for member, days in steps.items():
    coop.join(member, days)
    coop.agree(member, "mean steps")
print(coop.run("mean steps", "a city walkability study"))  # -> 6920.0, and no member's rows leave
  • DataCooperativeGateaiag/procedural/polycentric_governance.py:595 Data cooperative: members' data never leaves its store; queriers' algorithms are vetted, run in place on the data of members who consented to them, and return only aggregate answers over enough members, the insights the cooperative bargains with for its members (Hardjono & Pentland, 2019).

pip install aiagEvery method in aiag Library on GitHub