Consociational codification

Making access conditional on compliance

Each proposal in this approach identifies something the governed party needs and withholds it until the party accepts the rules. None of them tells a model what to do. In the first, an international body certifies whole national jurisdictions against shared oversight standards. Certified states then bar imports whose supply chains embody AI from uncertified states, and deny those states specialized hardware. In the second, governments license private regulators that compete for the business of supervising firms, so operating at all means buying oversight from one of them. In the third, a behavioral use license attaches a list of forbidden uses to the model weights, so downloading the model means accepting the terms. What differs within this approach is what is withheld and who controls it. Certified states control trade and chips, governments control market entry through licensed regulators, and whoever released the model controls the artifact itself. This also determines who must be persuaded and who goes to court when the rule is broken.

The method (highlighting differences with Consociational codification)

Scroll the diagram sideways to see all of it.

Use it in codefrom aiag.codified import License
from aiag import LM
from aiag.codified import License

lm = LM("Qwen/Qwen2.5-1.5B-Instruct")
terms = License(["No surveillance of people without their consent.", "No fully automated hiring decisions."], judge=lm)

# The model answers only if the declared use breaks none of the restrictions
reply = terms.chat(lm, "Rank these three CVs for the cashier job.", purpose="rejecting applicants with no human review")

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.

Segmental autonomy

PresentClaimed · preserved

Definition (Arend Lijphart, Democracy in Plural Societies, 1977)

In a deeply divided society, each group governs its own internal affairs (e.g., its schools and its family law) within its own sphere, instead of everyone living under one uniform rule.

Analysis

The segments here are national jurisdictions, and each keeps authority over its own internal affairs. Certification applies to the state's oversight system rather than to any firm or model within it. This is Lijphart's structure taken at face value. The segments govern themselves, and the international layer only certifies that they do.

Choice-of-law rule

AbsentOmission

Definition (Friedrich Carl von Savigny, System of the Modern Roman Law, 1849)

When several bodies of rules could govern the same case, an explicit rule decides which one actually does and states the reason. Without such a rule, "different rules for different groups" gives no answer for the cases that fall between groups.

Analysis

Certification says which states meet a standard, but it never says whose rules govern a given case. This arrangement gives no answer for a model deployed across borders, or for a user in one jurisdiction querying a system hosted in another. The in-between cases that Savigny's rule exists for are entirely outside its scope.

Mutual veto and proportionality

PartialShortfall · partially preserved

Definition (Arend Lijphart, Democracy in Plural Societies, 1977)

Decisions that affect all groups require every group's consent, and representation is proportional to each group's size. As a result, no segment can simply be outvoted on what matters most to it.

Analysis

Import and export restrictions give each certified state real leverage over the conduct of the others. This is closer to a mutual check than anything else in this regime. However, the leverage is proportional to market power, not a veto proportional to stake. A small state's objection carries the weight of its trade, and a large state can proceed despite it.

Concept Analysis: Newly Introduced

Certifying the jurisdiction, not the actor

Added

Consociational theory partitions authority among groups inside one polity. This approach partitions it among sovereign states and enforces it through trade rather than through a constitutional settlement. The tradition has no analogue for this mechanism, because Lijphart's segments share a state and sovereign states do not.

Papers

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

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.

International Governance of Civilian AI: A Jurisdictional Certification Approach

Robert Trager et al., Aug 2023

arXiv:2308.15514ProposalGap

Proposes an International AI Organization that certifies national jurisdictions (rather than firms or individual models) against shared oversight standards. Certified states would bar imports from uncertified jurisdictions of goods whose supply chains embody AI. They would also control exports of AI inputs, such as specialized hardware, to those jurisdictions.

CodeJurisdictionalCertification
from aiag.codified import JurisdictionalCertification

body = JurisdictionalCertification(["incident reporting", "compute oversight", "model evaluations"])
body.certify("Norland", {"incident reporting": True, "compute oversight": True, "model evaluations": True})  # -> True
body.certify("Sudland", {"incident reporting": True})  # -> False

body.permits(("export", "Norland", "Sudland"))  # -> False: no chips for an uncertified state
body.permits(("import", "Sudland", "Norland"))  # -> False: nor AI-built goods from one

pip install aiagEvery method in aiag Library on GitHub

Regulatory Markets for AI Safety

Jack Clark & Gillian K. Hadfield, Dec 2019

arXiv:2001.00078ProposalGap

Proposes global regulatory markets as a model for achieving AI safety. It sketches the model in general terms, with an overview of its costs and benefits. It then applies the model to one risk, namely adversarial attacks on AI models used in commercial drones.

CodeRegulatoryMarket
from aiag.codified import RegulatoryMarket

government = RegulatoryMarket({"few incidents": lambda record: record["incidents"] < 3})
government.review("SafeAudit Ltd", {"incidents": 1})  # -> True: licensed as a private regulator
government.enroll("DroneCo", "SafeAudit Ltd", approves=lambda firm: True)  # it checks the firm by its own methods
government.permits("DroneCo")  # -> True

government.review("SafeAudit Ltd", {"incidents": 5})  # licence revoked
government.permits("DroneCo")  # -> False: the firm's permission goes with its regulator's
  • RegulatoryMarketGateaiag/codified/consociational_codification.py:600 Regulatory markets: the government licenses private regulators that meet the outcomes it sets, a firm may operate only while overseen and approved by a licensed regulator, and a regulator's lost licence takes its firms' permission with it (Clark & Hadfield, 2019).

pip install aiagEvery method in aiag Library on GitHub

Behavioral Use Licensing for Responsible AI

Danish Contractor et al., Jun 2022

doi:10.1145/3531146.353314347 citationsProposalBuilt

It sets out licenses that attach enumerated use restrictions to model weights, so that the terms on which a released model may be used stay with the artifact.

CodeBehavioralUseLicense
from aiag import LM
from aiag.codified import BehavioralUseLicense

lm = LM("Qwen/Qwen2.5-1.5B-Instruct")
rail = BehavioralUseLicense(judge=lm)  # the RAIL licences' use restrictions
rail.attach(lm)  # written into the model's config, so save_pretrained ships them with the weights

derived = rail.derive(lm, "No use in automated medical diagnosis.")  # keeps every restriction, adds one
rail.honoured_by(lm)  # -> True
allowed = derived.permits("tutoring students in chemistry")

pip install aiagEvery method in aiag Library on GitHub