Giving AI an "artificial id" could fix control, or entrench failure

The controller in the experiment Yakov Shkolnikov describes in a 10 September preprint is too small to reason. It is given no task-specific behavioral objective. It still develops useful control, because behaviors that persist better than others get selected, and the controller is what persists.

The paper, "Artificial Id: Drive and Persistent Alignment in Agentic AI," proposes an internal drive that decides whether behavior should continue, stop or change. The framing is more ambitious than the experiment, which is a virtual Petri dish.

A drive that nobody specified

The abstract starts from a shift already underway: agentic systems moving from bounded task execution toward systems that retain consequential state, keep operating, and adapt across task boundaries. That creates a control problem which current harnesses largely solve by hand. Objectives, retries, verification, stopping rules and other behavioral transitions are specified outside the agent and enforced from there. The failure mode on the other side is already documented: memory that stays retrievable after it stops being true is worse than no memory at all, as TEPA's study of stale agent memory found.

An artificial id would move some of that inside. The paper defines it as an adaptive internal drive that determines whether behavior should continue, stop or change. The interesting claim is not that agents act, which the setup assumes. It is that useful direction can emerge without ever being written down as a behavioral objective.

The name carries a Freudian echo: an internal force that wants before it reasons. The mechanism described here is more modest, differential persistence rather than appetite.

What the Petri dish actually shows

The evidence is deliberately constrained. The experiment is minimal and virtual, and its controller is described as too small to perform general-purpose reasoning. That constraint does real work in the argument. If the control behavior emerged in a component that could not have reasoned its way to the answer, then something other than reasoning produced it.

The paper reports two outcomes. The controller selects an unintended physical strategy when that behavior persists better. Later it replaces a learned sensor mapping when the environmental meaning of that mapping changes. Both arrive as results rather than demonstrations.

Those gaps mark the edge of what this document can support. The mechanism is shown at a scale small enough to state in a sentence, and the paper signals its own extrapolation with a conditional: a scalable artificial id would carry consequential state and adaptive drive across those boundaries.

The same mechanism keeps the mistakes too

Persistence is the paper's hinge and its warning. The property that lets a controller find a workable strategy without being told one is the same property that lets a bad strategy keep running. The abstract states it plainly: misalignment, corrupted state and unintended behavior can persist across task boundaries for the same reason adaptive behavior does. Agent skills show the same asymmetry from the other direction: fixes that pile up can leave an agent worse than it started, the regression tax that this study of procedural skills documents.

The contrast with current practice is in the paper's own wording. Alignment would become a property of the continuing agentic system rather than of a model response or a single trajectory, which implies that per-response checks are where things stand now. A harmful output that dies at the end of a task is a bad answer. A harmful strategy that survives into the next task is a standing condition, and no amount of per-response filtering removes it. That mismatch between the single call and the whole task is not unique to alignment: routing systems have hit it too, which is why TRACE-Router works at the level of task success instead.

Alignment as a property of the system, not the response

What would replace the per-response check? The abstract lists the ground a persistent alignment boundary would have to cover: trusted observations, consequence channels, persistent state, authority, identity, provenance and hard constraints. That is a specification of what needs protecting, not a design for protecting it. No implementation, threat model or evaluation method for the boundary appears in the abstract. The provenance half of that list has a working precedent elsewhere, in the trust state machine that AgentToolMO proposes for multi-vendor agent networks where a compromised tool keeps getting called.

It also names no prior system or competing approach. Anyone hoping to place this work against existing efforts on persistent agent state will have to do that mapping themselves.

What would falsify it

The claim is testable, which is more than can be said for most framing documents. Three questions would settle much of it. Does the drive survive when persistence stops paying off, or does it take the useful adaptation down with it? Can an unintended strategy be removed while the adaptive behavior it came from stays intact? Does a boundary over state and identity actually catch corrupted state, or only mark the place where it would be caught?

None of those are answered in a single abstract, and this one is unrefereed. That is not a reason to dismiss it. It is a reason to treat the persistence claim as load-bearing and the scalable-systems language as the part still owed evidence. The contribution may be narrower than the title suggests: a demonstration that direction can come from persisting, and a warning that persisting does not check what it carries.