Introduction
As artificial agents—ranging from high-frequency trading algorithms to automated logistics systems—increasingly manage complex operations, they create a significant legal vacuum. Specifically, the "doctrine of attributed knowledge" dictates that if an agent (traditionally a human employee) knows a fact, their employer is legally deemed to know it too. This attribution is vital for establishing liability in cases of fraud, negligence, or contract disputes. However, applying this doctrine to software is difficult because traditional legal and philosophical frameworks define knowledge as "justified true belief"—a mental state that code arguably cannot possess.
This paper proposes a pragmatic, functionalist framework to bridge this gap, determining when an artificial agent "knows" something and when that knowledge should be legally attributed to its human or corporate principal.
The Epistemological Shift: From Belief to Function
The authors critique the classical "Justified True Belief" (JTB) model of knowledge. In philosophy, JTB has long been challenged by "Gettier cases," where an individual holds a true belief for the wrong reasons. In law, the JTB model is even less helpful when applied to artificial agents, as it requires an inquiry into the "internal state" or "consciousness" of a machine.
Instead, the authors advocate for a functionalist account of knowledge. By adopting the "intentional stance"—treating a system as if it has goals and knowledge to better predict its behavior—courts can bypass the "black box" problem. Under this view, knowledge is not a hidden mental state but a functional capacity.
A Pragmatic Test for Artificial Knowledge
The core of the proposed framework is a three-part test to determine if an artificial agent ($X$) knows a fact ($p$):
- Truth: $p$ must be true.
- Ready Access: $X$ must have ready access to $p$ within its data structures or environment.
- Use: $X$ must be able to make use of the informational content of $p$ to perform its designated functions.
This shift moves the legal focus from "did the bot tell the boss?" to "was the information available and actionable within the system?"
Legal Synthesis and the "Amazon Example"
The authors examine existing agency law in the US and British Commonwealth, specifically the "duty to communicate." Traditionally, knowledge is attributed if an agent has a duty to inform their principal. The authors argue this requirement is obsolete for AI.
Consider the "Amazon Example": If Amazon’s database contains a customer’s address and its software uses that data to facilitate a shipment, Amazon "knows" that address. If the software fails to ship because it "forgot" the address, the company cannot claim ignorance as a legal defense if the data was readily accessible. Attribution should be based on the availability of information within the principal's controlled systems, rather than the transmission of a specific message from the software to a human supervisor.
Historical Parallels: The Roman Law Analogy
Interestingly, the authors suggest a historical precedent in Roman law regarding the treatment of slaves. In the Roman legal system, slaves could enter into contracts and acquire knowledge for their masters without possessing full legal personhood. This provides a useful model for artificial agents: they can be seen as "legal instruments" that extend the principal's reach and knowledge without requiring the AI to be recognized as a person.
Limitations and Future Outlook
While the framework provides a roadmap, several challenges remain:
- The "Dead Data" Problem: There is a gray area regarding information that exists on a hard drive but is not easily accessible by the agent's current software. Determining the boundary between "stored data" and "known information" remains a technical hurdle.
- Jurisdictional Constraints: The analysis focuses on common law systems (US and British Commonwealth) and may not translate directly to civil law jurisdictions.
- Lack of Precedent: As of the paper's writing, there is a paucity of judicial pronouncements regarding AI knowledge. The framework serves as a normative proposal for how courts should handle these cases as they inevitably arise.
Conclusion
By redefining knowledge as "ready access and use," the authors provide a functional path forward for corporate accountability. As automation becomes the default for global commerce, the law must evolve to ensure that companies cannot hide behind the perceived "ignorance" of their autonomous systems.
Summary updated