Defining AI Models and AI Systems: A Framework to Resolve the Boundary Problem

This paper addresses the regulatory ambiguity surrounding "AI models" and "AI systems" by proposing clear conceptual and operational definitions that distinguish trained parameters from broader system components, thereby facilitating the precise allocation of obligations across the AI value chain.

The Gist

Imagine you are building a massive, high-tech restaurant. In this restaurant, there are two very important things: the secret recipe (the AI Model) and the entire restaurant operation (the AI System).

Right now, governments are trying to write new laws to make sure this restaurant is safe, fair, and doesn't serve bad food to customers. But there's a huge problem: nobody agrees on what exactly counts as the "recipe" and what counts as the "restaurant."

Here is the simple breakdown of what this paper is about:

1. The Confusing Mess

The paper starts by saying that current laws (like the EU's AI Act) try to assign rules to different people.

• The Provider makes the recipe.
• The Deployer runs the restaurant.

But because the definitions of "Model" (recipe) and "System" (restaurant) are fuzzy, it's impossible to know who is responsible for what.

• Example: If a customer gets sick, is it because the chef wrote a bad recipe (Model problem), or because the waiter served it at the wrong temperature (System problem)? Without clear definitions, we can't blame the right person.

2. The Detective Work

The authors acted like detectives. They read 896 academic papers and 80+ official rulebooks to see how people have tried to define these terms over the years.

They found a funny pattern: Almost everyone copied their definitions from the OECD (a big group of countries that sets global standards). But instead of clearing up the confusion, each time someone copied the definition, they accidentally added more confusion, like a game of "telephone" where the message gets garbled with every turn.

3. The "Recipe vs. Restaurant" Solution

The authors propose a fresh, clear way to look at it using a simple analogy:

• The AI Model is the "Brain" (The Recipe):
  Think of this as the trained brain of the AI. It's just the math, the code, and the learned patterns. It's like a cookbook sitting on a shelf. By itself, it can't do anything. It can't talk to you, it can't see your face, and it can't make decisions. It's just potential energy.

  • Technical bit: It consists of the architecture (the structure of the brain) and the parameters (the weights learned from training).

• The AI System is the "Whole Operation" (The Restaurant):
  This is the cookbook plus everything else needed to make it useful. It includes the kitchen, the waiters, the menu, the door where customers enter, and the cash register.

  • Technical bit: It's the Model + Software interfaces + Data pipelines + User interaction tools.

4. Why This Matters

By drawing a hard line between the "Recipe" and the "Restaurant," the paper solves the boundary problem:

• If the recipe is bad: The person who wrote the cookbook (the Model Provider) is responsible.
• If the restaurant runs poorly: The person running the operation (the System Deployer) is responsible.

The Real-World Impact:
Imagine a self-driving car.

• The Model is the neural network that knows how to recognize a stop sign.
• The System is the car itself, including the sensors, the brakes, the screen showing the driver, and the software that decides when to hit the brakes.

If the car crashes because it didn't recognize the stop sign, that's a Model issue.
If the car crashes because the brakes failed or the screen was too dark for the driver to see, that's a System issue.

The Bottom Line

This paper says, "Stop guessing!" It offers a clear, simple framework: The Model is the brain; the System is the brain plus the body and the tools.

With this clarity, governments can finally write laws that make sense, companies know exactly who is responsible for what, and we can stop arguing about definitions while trying to fix real-world AI problems.

Technical Summary

Based on the abstract provided, here is a detailed technical summary of the paper "Defining AI Models and AI Systems: A Framework to Resolve the Boundary Problem."

1. The Problem: Conceptual Ambiguity in AI Regulation

The core issue addressed is the lack of clear, consistent definitions for the foundational terms "AI model" and "AI system." This ambiguity creates a "boundary problem" with significant practical and legal consequences:

• Regulatory Misalignment: Emerging frameworks, such as the EU AI Act, assign distinct legal obligations to different actors (e.g., "providers" vs. "deployers") based on whether they interact with a model or a system. Without clear definitions, it is difficult to determine which actor bears which responsibility.
• Operational Confusion: The ambiguity complicates the identification of specific modifications. It is often unclear whether a change constitutes a modification to the model itself (e.g., retraining weights) or to the non-model system components (e.g., data pipelines, user interfaces).
• Historical Context: The paper identifies that current regulatory and standards definitions largely trace their lineage back to OECD frameworks. However, the evolution of these frameworks has tended to compound conceptual ambiguities rather than resolve them, leading to a fragmented landscape of definitions.

2. Methodology

The authors employed a rigorous, multi-stage research approach to analyze the current state of definitions and propose a new framework:

• Systematic Literature Review: A comprehensive analysis of 896 academic papers to gather existing theoretical and technical definitions.
• Manual Document Review: A detailed examination of over 80 documents, including regulatory texts, technical standards, and policy papers.
• Genealogical Analysis: The authors traced the "definitional lineages" and paradigm shifts over time to understand how current definitions evolved and where the conceptual drift occurred.
• Conceptual Synthesis: Based on the review, the authors developed new definitions grounded in the fundamental nature of models versus systems.

3. Key Contributions

The paper makes three primary technical contributions:

1. Critical Analysis of Lineage: It provides a historical critique showing how current regulatory definitions inherited and exacerbated ambiguities from earlier OECD frameworks.
2. Proposed Conceptual Framework: The authors introduce a dual-layer definition structure:
   • Conceptual Definitions: Grounded in the theoretical relationship between models and systems.
   • Operational Definitions: Specifically tailored for contemporary neural network-based machine learning.
3. The Distinction Framework:
   • AI Model: Defined strictly as the trained parameters and architecture. This isolates the mathematical core of the intelligence.
   • AI System: Defined as the model plus additional components. These additional components include the infrastructure required to process inputs and outputs, such as data preprocessing modules, user interfaces, feedback loops, and deployment environments.

4. Results and Implications

• Clarification of Boundaries: The proposed framework successfully delineates the boundary between the "model" (the algorithmic core) and the "system" (the operational wrapper). This resolves the confusion regarding where the model ends and the system begins.
• Regulatory Implementation: By clearly distinguishing between the two, the framework aids in the precise allocation of responsibilities. For instance, it clarifies whether a specific intervention (like a safety filter or a UI change) falls under the obligations of a "model provider" or a "system deployer."
• Case Study Validation: The authors demonstrate the utility of their definitions through theoretical scenarios and real-world incident case studies, showing how the new framework resolves ambiguities that previously hindered accountability.

5. Significance

This paper is significant because it moves beyond theoretical debate to offer a practical, operational tool for regulators, policymakers, and industry practitioners.

• Legal Certainty: It provides the semantic clarity needed to enforce regulations like the EU AI Act effectively, ensuring that obligations are assigned to the correct actors in the value chain.
• Technical Precision: By anchoring definitions in the specific architecture of neural networks (parameters/architecture vs. integration components), the framework aligns legal terminology with technical reality.
• Future-Proofing: The framework offers a scalable approach to defining AI components, which is crucial as AI technologies continue to evolve and become more integrated into complex socio-technical systems.