Justin Curl

0 followers · 239 views

on the atlas — 5

highlights — 26

a democracy cannot survive if citizens are not committed in everyday life to the values of democracy such as support and respect for the rule of law that emanates from legitimate institutions and respect for the equal moral worth and dignity of all.
Building AI for the Democratic Matrix: A Technical Research Agenda for Normative Competence and Normative Institutions | Knight First Amendment Institute
We also think that AI agents will need to participate in third-party distributed punishment efforts. This means adhering to the norm that it is expected that one will punish a norm violator—refusing to do business with an entity that flouts the law, for example.
Building AI for the Democratic Matrix: A Technical Research Agenda for Normative Competence and Normative Institutions | Knight First Amendment Institute
predicting the classifications that would be reached by appropriate (formal and informal) classification institutions; predicting the likelihood of punishment, both formal and informal; and adjusting behavior in light of the risk of punishment.
Building AI for the Democratic Matrix: A Technical Research Agenda for Normative Competence and Normative Institutions | Knight First Amendment Institute
ow will the choices it makes about product design, employment, marketing, manufacturing, regulatory compliance, and so on be classified? Will those choices be punished? Will it lose money if it markets pregnancy vitamins to minors or includes materials on human evolution or religious doctrine in educational materials? What if it employs a felon to work in a warehouse? Or if it uses the language of “diversity, equity, and inclusion” in its marketing materials? What if it refuses to buy from a supplier in a country that has been recently labeled anti-American by the government, has been found to…
Building AI for the Democratic Matrix: A Technical Research Agenda for Normative Competence and Normative Institutions | Knight First Amendment Institute
Moreover, whether an action classified as ‘punishable’ will, in fact, be punished requires predicting the behavior of both officials and ordinary members of the group.
Building AI for the Democratic Matrix: A Technical Research Agenda for Normative Competence and Normative Institutions | Knight First Amendment Institute
If the executive branch says to private employers, “we want you to fire employees who oppose our not-legislatively-endorsed decrees,” the risk to democracy comes in large part from whether private employers resist, and whether they think their peers will approve of their resistance.
Building AI for the Democratic Matrix: A Technical Research Agenda for Normative Competence and Normative Institutions | Knight First Amendment Institute
If individual citizens do not (mostly) voluntarily conform their behavior to the norms and laws developed by democratic institutions, then democracy exists on paper but not on the ground. If citizens do not (mostly) create incentives for one another to comply—such as by refusing to do business with a partner who treats court orders and contracts as meaningless, systematically flouts equal treatment laws, or routinely and obviously violates regulations—then the achievement of a democracy in fact is undermined.
Building AI for the Democratic Matrix: A Technical Research Agenda for Normative Competence and Normative Institutions | Knight First Amendment Institute
All of the decisions the AI agent loosed on the world makes, like those made by any human agent, are embedded in the normative social order. That order consists of norms such as those governing what colors or symbols it is appropriate to use in product design, what language to use in marketing or communications with customers, and how to price to discriminate between customers or responds to unusual events like shortages or natural disasters.
Building AI for the Democratic Matrix: A Technical Research Agenda for Normative Competence and Normative Institutions | Knight First Amendment Institute
What these people mean by “AI agents” are AI systems that are capable of taking in general instructions from a user, producing a plan for how to act on those instructions in a potentially wide and uncontrolled environment, and then taking autonomous actions in the world
Building AI for the Democratic Matrix: A Technical Research Agenda for Normative Competence and Normative Institutions | Knight First Amendment Institute
An agent is anything that can be viewed as perceiving its environment through sensors and acting upon that environment through actuators.
Building AI for the Democratic Matrix: A Technical Research Agenda for Normative Competence and Normative Institutions | Knight First Amendment Institute
The impartial spectator is not a passive capacity; it represents the use of cognition to direct attention to the behavior and assessments of others, evaluated not as mere data, but as inputs to the reasoning that appears to guide appropriate judgment in society
Building AI for the Democratic Matrix: A Technical Research Agenda for Normative Competence and Normative Institutions | Knight First Amendment Institute
For Smith, the essence of morality was the idea that we each carry around inside us an imaginary observer to whom we can appeal as we consider how to behave. We imagine the response of a fair-minded individual, “impartial and well-informed,” who judges our conduct against the standards of the community (Smith, 1976).
Building AI for the Democratic Matrix: A Technical Research Agenda for Normative Competence and Normative Institutions | Knight First Amendment Institute
It is one in which the overarching classification institution reflects basic democratic values, such as the equal worth and dignity of all, and is open to hearing from and responsive to the members of the group in the robust sense that the institution is ultimately shaped or controlled by them
Building AI for the Democratic Matrix: A Technical Research Agenda for Normative Competence and Normative Institutions | Knight First Amendment Institute
They must possess the cognitive ability to take in information about normative classifications. They must recognize and pay attention to the normative institutions in their community, the sources of authoritative classifications that resolve uncertainties and disputes. They must attend to normative behaviors, specifically the punishment behavior of other agent
Building AI for the Democratic Matrix: A Technical Research Agenda for Normative Competence and Normative Institutions | Knight First Amendment Institute
A democracy is upheld and instantiated by individuals who take offense and perhaps act out when people are denied free speech or due process, who recognize their obligation to treat others with dignity and respect, who expect their representatives to show fidelity to their constituents and the law, and who vote against them if they do not
Building AI for the Democratic Matrix: A Technical Research Agenda for Normative Competence and Normative Institutions | Knight First Amendment Institute
Ultimately, the core of democracy is the capacity of the people acting as a collective to reshape their classification institutions or build entirely new ones. A constitution establishes rules for changing a democracy’s institutions. And constitutions themselves can be rewritten and reinvented.
Building AI for the Democratic Matrix: A Technical Research Agenda for Normative Competence and Normative Institutions | Knight First Amendment Institute
A thriving community balances stability and adaptation (Mathew et al., 2025). And this is a key value of distinctively legal order with an identifiable entity engaging in stewardship of classification: in a stable legal order, when the classification institution respects the constraints on its behavior (the legal attributes), a group can modify its classifications without disrupting its normative social order
Building AI for the Democratic Matrix: A Technical Research Agenda for Normative Competence and Normative Institutions | Knight First Amendment Institute
Stable classification allows members of the group to reliably predict which behaviors they should punish and which they should avoid, as they form and execute on their daily and long-term plans
Building AI for the Democratic Matrix: A Technical Research Agenda for Normative Competence and Normative Institutions | Knight First Amendment Institute
we have classification institutions that govern the jurisdiction of these sub-groups. The federal constitution, for example, protects the freedom of religious groups to determine some of the norms governing their members but not all: they can shun members who dress immodestly, engage in premarital sex, or fail to show charity to the poor, but they cannot physically abuse children or violate federal anti-discrimination laws.
Building AI for the Democratic Matrix: A Technical Research Agenda for Normative Competence and Normative Institutions | Knight First Amendment Institute
classification is often performed by institutions that also possess centralized enforcement authority, or at least the capacity to call on centralized enforcement. Legislatures write laws and they can fund or command a police force. Courts can declare wrongdoing and order private or public actors to take steps to punish a wrongdoer
Building AI for the Democratic Matrix: A Technical Research Agenda for Normative Competence and Normative Institutions | Knight First Amendment Institute
But it may also be that the obligation to punish wrongdoers is itself a norm that is enforced by third-party punishment, a bootstrapping of the enforcement mechanism (Henrich and Boyd, 2001; Henrich, 2004), or (as Hadfield and Weingast (2012) suggest) a means of signaling ongoing support for, and avoiding the collapse of, a normative social order
Building AI for the Democratic Matrix: A Technical Research Agenda for Normative Competence and Normative Institutions | Knight First Amendment Institute
A normative social order is an equilibrium state in which a group of independent actors pursuing ordinary self-interested utility (i.e., not innately motivated by pro-social preferences) are incentivized and coordinated to take costly efforts to punish members of the group who engage in behaviors classified as punishable (disapproved, inappropriate, wrongful) by a classification institution
Building AI for the Democratic Matrix: A Technical Research Agenda for Normative Competence and Normative Institutions | Knight First Amendment Institute
It requires us to build AI systems that possess normative competence, which we define as the computational ability to detect social sanctions, attribute them to specific behaviors, and adjust future actions accordingly (Hadfield, 2025).
Building AI for the Democratic Matrix: A Technical Research Agenda for Normative Competence and Normative Institutions | Knight First Amendment Institute
Current technical approaches to democratic alignment primarily focus on preference aggregation and rule encoding. We have seen efforts to collect “democratic inputs” for fine-tuning large language models (OpenAI, 2023; Huang et al., 2024; Bergman et al., 2024); sophisticated fine-tuning technologies to implement “constitutional AI” (Bai et al., 2022); calls for AI alignment to be grounded in methods that elicit and aggregate pluralistic values (Sorensen et al., 2024; Conitzer et al., 2024); and calls for “law-following” AI (O’Keefe et al., 2025). These approaches implicitly assume that democra…
Building AI for the Democratic Matrix: A Technical Research Agenda for Normative Competence and Normative Institutions | Knight First Amendment Institute
But a key innovation of the Hadfield and Weingast (2012) model is to focus attention not on compliance but punishment: how do we get ordinary agents to punish, and how do we get them to punish the same behaviors? Punishment in this context refers to specifically third-party punishment, a uniquely human behavior that distinguishes us from other mammals (Tomasello and Vaish, 2013).
Building AI for the Democratic Matrix: A Technical Research Agenda for Normative Competence and Normative Institutions | Knight First Amendment Institute
u fell unconscious in 1950, no one around you would know how to perform CPR: it wouldn’t be invented for another 10 years.1 Or take type 1 diabetes, where survival would involve injecting yourself every day with a thick glass syringe of insulin that was extracted from animal pancreases (with tens of thousands of animals required to produce each pound of i
Medical breakthroughs in 2025 - by Saloni Dattani