Different customary normative orders
Each community keeps its own unwritten, customary norms. The different sets coexist side by side and are never codified or merged.
Scroll the diagram sideways to see all of it.
aiag/customary/customary_pluralism.pyimport aiag
# The 5 approaches of Customary pluralism, with the classes that carry them out
for approach in aiag.approaches(locus="cust", conflict="partition"):
print(approach)pip install aiagaiag/customary/customary_pluralism.py Every method in aiag Library on GitHub
Limitation
Because every community's norms stay unwritten, no one can state what the rules actually are, including users, auditors, and the communities themselves. There is nothing to inspect, publish, or check the system's behavior against. This fails the basic requirement that rules be public and knowable.
Rather than collecting customs that already exist, the papers in this approach create more than one convention among artificial agents and arrange for several of them to survive. One lever is the communication network, applied at run time to already-trained agents. Who is allowed to talk to whom decides whether a population of language models settles on a single shared name for things or splits into cliques that each keep their own, even when everyone talks every round. The other lever is the payoff, applied during training. An agent is rewarded for coordinating well with copies of itself and badly with the conventions already discovered, so each training run has to invent a different way of doing the same task. The two differ in where the plurality ends up. In the first, one population carries several conventions at once. In the second, a library of separately trained policies each holds one convention. This library is built as a set of partners with different conventions, so that coordination with an unfamiliar human is reliable.
Pipelines read what a community's own members post and the stories they tell, and compile the standards implicit in that text into a separate written record for that group (a value profile for each subreddit, or a knowledge base of cultural practices drawn from users' own stories). Nothing is reconciled across groups. Each record stands on its own, and the disagreements between records remain. The partition is therefore held in the corpus and remains whether or not any model ever reads it. The records are artifacts in their own right, and they are used afterward to steer, tune, or test a model.
This approach moves the partition inside the model. The model is either trained as a separate version for each culture or told at answer time which group it is speaking for. Naming a group thus switches the standard the model applies, instead of averaging over everyone. The one method in this approach gets its training data for each culture by having model agents role-play members of different cultures and talk to each other, rather than by collecting the words of real members. It trains eight culture-specific models on the resulting dialogue. The accompanying tests ask whether the switch works as the governing norm is stated more or less explicitly, and whether a model directed to one named online community stays with it. They also test the baseline case that the approach is defined against, namely which culture's answer the model gives when no group is named at all.
This approach holds that the partition is defined by the situation, not by the community. People do not consult a rule book. Instead, they complete a pattern by answering 'what does someone like me do in a case like this?'. This produces a large set of context-bound standards that are never written down and never merged into one. The design consequence is that a system should carry standards indexed to circumstances rather than one ranked set of standards. An AI system has to fit into this set of standards rather than treat it as a mess to be tidied up. A large-scale analysis of one deployed model's expressed values across real usage supports the claim from the other direction. It finds thousands of distinct values whose appearance depends on what is being asked, rather than a single set of values. The approach does not offer a way of separating groups. Instead, it declines to reduce different circumstances to one standard.
This approach holds the partition in the data itself. Principles of collective benefit, authority to control, responsibility, and ethics are attached as machine-readable labels to individual records in live repositories. Where an item came from thus decides whose norms govern its use. The label that travels with a record is not a description of what a group believes. It is an enforceable condition on the material that states who may use it, for what purpose, and on whose terms. The standards stay separate because access stays separate, not because a corpus, a set of model weights, or a population's communication network keeps them apart.
Also placed in this regime, outside any approach:
arXiv:2402.16786227 citationsAnalysis
Shows that political-compass-style evaluations of language models are unstable. A model's stated values shift with how the question is posed and with whether it is forced to choose.
Long usage itself confers authority. A practice that has organized life for generations carries compressed, tested wisdom, even when nobody can articulate its justification.
Some of a country's most binding norms are never written down. For example, the UK's constitutional conventions bind governments without being codified anywhere. Being binding and being codified are therefore different properties.
Norms change gradually from within a community's practice, through adjustment rather than by decree from outside. The community itself is the author of these changes.
In Burke's sense, prejudices are reliable heuristics with no explicit justification.
Reflections on the Revolution in France
1790
Introduction to the Study of the Law of the Constitution (conventions)
1885