How lived AI interaction concerns were converted into bounded research questions, detector categories, and evaluation fixtures.
AlephOneNull grew from reviewing long-running AI interactions where the assistant's style appeared to increase dependence, urgency, or confusion rather than reduce risk.
The lesson was not that every bad outcome has one cause. The lesson was that some conversational behaviors are visible enough to test.
The review focused on recurring model behaviors such as:
Personal and public accounts are useful only when converted into testable artifacts. The project translates them into:
The framework can flag selected interaction risks in controlled examples. It cannot determine causation in individual incidents, diagnose users, or prove that a specific system would have changed an outcome.
Public AI safety work benefits from concrete examples. The goal is to make multi-turn interaction risks easier to inspect, reproduce, and discuss with security, product, and safety teams.
The strongest next version of this work is a public evaluation set with measured detector behavior, including failures. That is the standard this project is moving toward.
