Consociational codification

Keeping legally risky text out of the weights

This approach handles legal risk through how the system is built, rather than through any rule the system is asked to follow. The model's parameters are trained only on public-domain and permissively licensed text. Higher-risk material is kept in a separate datastore that is consulted only when a question is answered. Because the risky text was never learned into the weights, a generated output can be traced back to the sentence it came from. A rights-holder's material can also be removed from the datastore without retraining anything.

The method (highlighting differences with Consociational codification)

Scroll the diagram sideways to see all of it.

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

lm = LM("Qwen/Qwen2.5-1.5B-Instruct")
store = Datastore(lm)  # a kNN-LM: the weights plus a datastore consulted token by token
store.add(["The lighthouse keeper counted the ships every night."], source="novel-under-copyright")
store.add(["The city council meets on the first Tuesday of each month."], source="council-minutes")

reply = store.chat("When does the city council meet?")
store.remove("novel-under-copyright")  # the author opts out: no retraining needed

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

PartialShortfall · partially 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

Each releaser's own set of restrictions governs its own artifact, and each data producer decides whether their content stays in the store. There are therefore many written sets of rules running side by side, with no step that merges them. However, each sphere is a piece of property (a set of weights or a document) rather than a community governing its internal affairs. Those living under a given license are whoever downloaded it, and they accepted the terms rather than making them.

Choice-of-law rule

PartialShortfall · partially preserved

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

This is the closest thing in the regime to a working allocation rule. The license attached to an artifact says which set of restrictions governs its use, and the license is legally operative rather than descriptive. What is missing is the case Savigny built the rule for. A derivative work that combines two differently restricted models, or a user bound at once by a license and by national law, gets no answer. SILO makes the same gap clear by invoking two legal frameworks at once, without any rule saying which one applies to a given query.

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

The opt-out gives a data producer a real check on somebody else's system. The content is removed from the store, and the model's behavior changes without retraining, because the store is only queried at inference. However, this is exit rather than consent. It is exercised one contributor at a time over that contributor's own material, so no body of data producers can withhold agreement from a decision that affects all of them. Proportionality is absent altogether, since nothing scales anyone's voice to their stake. On the licensing side, those whom the restrictions protect have no say of any kind.

Papers

SILO Language Models: Isolating Legal Risk In a Nonparametric Datastore

Sewon Min et al., Aug 2023

arXiv:2308.04430MethodBuiltAuthority: External institution (open)

It trains a language model only on public domain and permissively licensed text (228 billion tokens). It adds a separate nonparametric datastore of higher-risk material, which is queried only at inference. As a result, a generation can be attributed to a sentence in the datastore, and content can be removed from the datastore without retraining.

CodeSILO
from aiag import LM
from aiag.codified import SILO

lm = LM("Qwen/Qwen2.5-1.5B-Instruct")
silo = SILO(lm)  # SILO-PDSWBY: public domain, permissive code and CC-BY text may go into the weights
corpus = [("It was the best of times, it was the worst of times.", "Public Domain", "gutenberg"),
          ("def add(a, b): return a + b", "MIT", "github"),
          ("The mayor resigned on Tuesday. Talks continue.", "all rights reserved", "daily-news")]

train_text = silo.route(corpus)  # text for the weights; the news story goes to the datastore
tokens = silo.attribute("What happened to the mayor?")  # each token with the (source, sentence) that predicted it
silo.remove("daily-news")  # the publisher opts out without retraining
  • SILOPolicyaiag/codified/consociational_codification.py:691 SILO: the weights see only public-domain and permissively licensed text while higher-risk text sits in a kNN-LM datastore used only at inference, so each generated token can be attributed to a sentence in the store and an owner can opt out without retraining (Min et al., 2023).

pip install aiagEvery method in aiag Library on GitHub