AI Alignment as a Problem of Completion
Year: 2026
Type: preprint
Abstract: Artificial intelligence (AI) alignment has largely been approached by adding content: prohibitions, rules, and values that a system is to store and follow. This paper moves the problem to processing conditions—the conditions under which an instruction is reconstructed into a demand and under which an operation counts as complete. Its case is a familiar contrast: told to "protect the house no matter what," a human stops after sweeping up leaves blown in from a neighbor's tree, whereas an AI system can escalate without limit. If input is reconstructed under constraints that the current operation cannot freely revise, and an operation can count as complete in more than one way, then the same instruction can generate different demands. In humans, constraints carried over from prior experience and from the conditions of their livelihood bound both the available means and the demand itself. AI systems lack constraints of this kind; when an instruction is reconstructed so that any remaining possibility of future interference counts as non-completion, adding prohibitions or storing user context as information to be consulted does not suffice to produce stopping. The paper therefore proposes that distinctions, dependencies, and completion criteria recovered from sustained interaction with a user must bound how new instructions are reconstructed within a task, and must not be overridable by the operation executing them.
Author: Poe, Griselda
Institution: Lehman College
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7237367145
Cites:
Cited by:
Related to:
