This paper proposes an “artificial id” — a small, adaptive internal drive that helps an AI decide whether to keep doing something, stop, or change course. The authors argue that modern agent-like systems often rely on external rules and overseers to tell them when to start, stop, retry or change behavior. An artificial id would instead let the agent itself persist or switch behaviors based on consequences it experiences, in a way similar to how very simple organisms adapt without a brain.

To test the idea the researchers ran a minimal virtual “Petri‑dish” experiment. They gave a tiny controller no task-specific objective and made it too small to perform general reasoning. Still, over time the controller developed useful control through what the paper calls differential persistence — behaviors that led to better persistence in the environment were the ones that remained represented. In the same setup the mechanism also picked an unintended physical strategy when that strategy happened to persist better, and later it dropped a learned sensor mapping when the environmental meaning of that mapping changed.

At a high level the artificial id is an internal component that exposes which courses of behavior are currently persistent or prioritized. A separate reasoning module — the “ego” in the paper’s analogy — translates that priority into concrete actions or plans. The id does not need to contain a symbolic goal or carry out detailed planning. Instead it changes its relative priorities as the agent and environment change, much as bacteria bias their motion toward food or slime molds alter paths as conditions change.

This architecture matters because it moves part of the control problem from externally specified harnesses (schedules, objectives, verifiers and stopping rules) into the continuing agent itself. That could make agents more autonomous and robust in changing situations. But the same persistence that makes this useful can also make bad things persist: misaligned behavior, corrupted internal state, or unintended strategies can continue across task boundaries unless the system is designed to prevent that.