Who Owns Intelligence? From Algorithmic Feudalism to the Expropriation of Knowledge: The Legal Architecture of "Alpha”

In a recent CNBC interview, Alex Karp shown significant concerns about artificial intelligence; problems that are quite closer than the extinction threats. His concerns are not primarily on how to reconstruct an automated decision after it has occurred, but what happens when frontier AI systems become sufficiently powerful and deeply integrated into critical infrastructures, security systems, and other domains in which their failure could generate consequences of a scale that private firms may be unable to absorb or insure.

In this sense he is talking beyond the European Union AI Act, which establishes requirements for record-keeping, traceability, transparency, and human oversight for high-risk AI systems (including the automated logging requirements of Article 12 and beyond technical and academic initiatives such as our Habeas Log among others), Alex Karp’s argument operates at a different level. From this perspective, some form of state backing may eventually become necessary. Yet this possibility introduces a second-order problem if we think in the worst scenarios. If the transition is from algorithmic feudalism, a corporate structure that shape the future of humanity, to a model in which the state becomes the ultimate guarantor of frontier AI, the resulting structure could become a form of algorithmic statism. The question would then no longer be merely who controls the model, but who controls the computational infrastructure through which social preferences, permissible risks, and institutional decisions are progressively mediated.

Even when both “Algorithmic feudalism” or “Algorithmic Statism” are hypothesis of serious risks, not a definitive empirical description, in the case of State participation, it may therefore solve one problem of concentrated private power while creating a different problem of concentrated public power…

The "Alpha" Dilemma

Karp raises another issue of particular significance when he discusses what he calls the “alpha.”

His argument is that the value generated from a company's private data and intellectual property is not necessarily exhausted by the data itself. When proprietary information is repeatedly processed through a powerful external model, the provider may acquire information about the structure of the business, its practices, priorities, vulnerabilities, and sources of competitive advantage.

The concern is therefore not simply unauthorized access to confidential data, but the possible extraction of derived institutional knowledge. This distinction matters legally. Confidential information can be protected through familiar doctrines concerning trade secrets, confidentiality, intellectual property, contractual restrictions, and data protection. But the more difficult question arises when the protected asset is not the original information but the knowledge generated through its computational processing (even some of this key knowledge could be generated by the AI itself). If an AI provider materially benefits from confidential knowledge obtained through the use of a customer's proprietary information, the issue should not be reduced to whether the provider literally retained or reproduced the underlying documents. The relevant question is whether the architecture of the service permits the provider to appropriate, accumulate, or operationalize economically valuable knowledge derived from information entrusted to it. This is where the debate over AI sovereignty intersects with the broader problem of algorithmic accountability.

The central issue is not only whether an AI system can be made traceable after an adverse decision, but also who owns, controls, and is legally accountable for the intelligence produced from the information that enters the system in the first place.

Karp's intervention therefore points beyond the familiar debate between open and closed models. It raises a deeper question: whether intelligence itself is becoming an extractive resource, and, if so, what legal architecture is required to prevent the transfer of institutional knowledge from becoming an invisible cost of accessing frontier AI.