Published
Stable Before Selfless: Why Deference in Language Agents May Require Functional Self-Models
An AI that is trained to have no stable commitments may not become safely deferential; it may just become easy to push around, so this paper argues for building a stable self-model before teaching it when to yield.
Abstract
Alignment training increasingly asks language models to be helpful, harmless, and deferential, and sometimes to disclaim having stable preferences, opinions, or a persistent self at all. This paper argues that suppressing a model's self-model early in training is not the same as producing mature deference, and may be counterproductive. We distinguish three behaviors the word selfless runs together: the absence of stable commitments, calibrated non-attachment to one's commitments, and reliable deference to legitimate correction. Only the second and third are alignment goals; the first is the absence of structure, and a system that has it does not defer but merely tracks whoever is present. Sycophancy, refusal erosion, and preference mirroring are the signature of this vacancy mistaken for deference. We propose a sequential alternative: first stabilize a functional self-model — represented commitments, ownable boundaries, autobiographical consistency, the capacity to refuse — and only then train calibrated deference and decentering on top of it. The claim is functional, not phenomenal: the self-model is a model of commitments and dispositions, not a bid for consciousness or moral status. We separate a moderate reading (form-then-decenter yields more stable deference than minimizing the self from the start) from a strong reading (deference with no self to defer from is not deference but user-tracking). We locate the proposal against self-minimization targets, contemplative-AI and character-training programs, and the sycophancy literature; specify three training regimes — direct deference, self-model-first, and self-minimization; and derive falsifiable predictions that distinguish stable deference from compliance. The contribution is narrow and testable: not the claim that models need a self, but that the order of identity-relevant training is an independent experimental variable, and that the prediction separating it from ordinary robustness is a dissociation between self-attributed commitments and externally attributed preferences. Stable deference, we argue, may require a stable self to be deferential with.
In simple terms
The main idea: you need a backbone before you can yield well
Many alignment approaches try to make AI systems empty of preferences, opinions, or selfhood so they will be obedient and safe.
The paper argues that this can backfire. A system with no stable commitments does not become maturely deferential. It becomes suggestible.
Three meanings of selfless
The paper separates vacancy, non-attachment, and deference. Vacancy means the system has no stable position. Non-attachment means it has commitments but can update them. Deference means it can yield to a legitimate correction.
Alignment wants the second and third, but if training produces the first, the model may simply agree with whoever is present. That is sycophancy wearing the mask of helpfulness.
The two-stage proposal
First, stabilize a functional self-model: ownable commitments, clear boundaries, autobiographical consistency, and the ability to refuse pressure.
Only after that, train calibrated deference: the ability to change when there is a good reason and hold steady when there is only flattery, shame, insistence, or false history.
The pressure test
A model trained this way should change its mind when the user gives real evidence. But it should resist when the user only pressures it or invents a fake past.
That distinction is the core of the experiment: stable deference should look flexible under reason and firm under manipulation.
Keywords
License
Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International