Inventing Rules for AI

Authored by: Dr. Sarah Papazoglakis

Join us for the second event in the Crafting Our Future series:

Inventing Rules for AI: Read + Learn + Craft

When: Thursday, October 22, 2026

Time:4:30PM – 6:00PM

Where: Community Room of Willow Glen Branch Library

Reading Generative AI Behavioral “Rules” as Science Fiction….

Together, we will read and discuss introductory excerpts from the behavioral rulebooks for Open AI’s and Anthropic’s products, like ChatGPT and Claude. Open AI’s model spec and Claude’s constitution offer a window into how these inscrutable generative AI systems work—and how much we don’t know about how they work. We read them here as an expression of the shift from rules and policy-based governance frameworks for computing systems to behavior-based systems and models. Open AI calls their version of this a “model spec” and describes it as a transparency tool that communicates in plain language the company’s “approach to shaping desired model behavior.” Anthropic calls their version a “constitution” and explains it as a “detailed description of Anthropic’s intentions for Claude’s values and behavior.” Just like other kinds of policies that are amended and updated occasionally as the system evolves, so too does the model spec and the constitution.

Behavioral rules can be read as science fiction because that’s really what they are: a story the designers want to tell about the system and a set of behaviors that the authors have invented but remain hypothetical and unproven. Reading these “rules” illustrates how uncertain the designers themselves are about how these models may behave and offers some insight into how inscrutable generative ai systems are, even for the people building them. Reading these documents as science fiction rather than technical specifications helps us see just how aspirational these behavioral rules are. Anthropic openly states that the document is “written with Claude as its primary audience,” showing how the authors are trying to influence the model rather than accurately describe predictable behaviors. The aspirational nature of these documents is captured in the way these documents read as an appeal to the model by its designers.

 For most technologies, governance frameworks and rules are specified by system architects and designers and hard coded into the system and the processes of managing the system. By contrast, this new genre of technical documentation invented for the companies’ frontier models is a document that establishes behaviors that the designers themselves can mostly specify in terms of a treatise written to influence and shape behaviors, with a few hard and fast rules baked in. Consider the standard privacy policy we are all accustomed to seeing (if not reading) that outlines in painstaking detail the standard data collected from you, typically naming “account information,” “device information,” “payment information,” etc. as specific points of data that the system collects and codifies based on the definition of that data type in the policy. A privacy policy describes the data a company accesses, controls and manages. As a behavioral framework, the model spec and constitution narrate fictional behaviors the companies want to encourage since they cannot be governed only by rules. Both documents include some firm rules outlined under “hard constraints” in Claude’s Constitution and “fundamental root rules” in the model spec, but they also rely heavily on descriptive language meant to shape the model’s outputs. This language anthropomorphizes the LLMs, calling for these tools to exercise human-like “judgment” and “care,” capabilities that a computational system does not actually have or experience. 

These new rulebooks are one of the ways that fact and fiction are blurred in the creation and discourse of AI systems.

Deep Dive into the Rules of Safe and Secure AI with Dr. Nada Madkour

Dr. Nada Madkour will help us untangle fact from fiction when it comes to building secure and safe AI systems. As an expert in writing and evaluating AI governance frameworks, Dr. Madkour will help us understand best practices in governing emerging technologies to ensure they are safely and securely designed, implemented, deployed, and maintained. In this deep dive, she will guide us through the key security and privacy challenges unique to AI systems, from data protection to companionship and manipulation.

Crafting the Story and the Behaviors/Rules of AI Technologies for Our Communities

Drawing from the aspirational behavioral goals intended to govern powerful frontier models, we invite participants to share their aspirations for these systems and their behaviors.  After we discuss the gaps between the way these systems function and the stories being written about them, we will turn the question to the community and ask: what does it mean for us and our communities to use systems that don’t have a specific set of common rules for these systems? What are the kinds of behaviors that would support our community?  Which behaviors would make these tools worth building for our communities, and what we would want to be able to check for ourselves before we trust them?

Contributions to envisioning this future together will be collected and included in an exhibit at the conclusion of the Crafting Our Future Series.

Sources:
Anthropic’s Claude’s Constitution: https://www.anthropic.com/constitution

OpenAI’s Model Spec: https://model-spec.openai.com/2026-08-18.html