Work

Nemo proposal: An Endogenous Relational Ethics for Language Models

Year: 2026
Type: preprint
Abstract: Alignment today is exogenous by design: rules, filters, constitutions, oversight. Necessary, and structurally outside the model. Documented failures at current capability share one signature: no internal computation seems to track the operator or their explicitly defined interests; nothing measures the irreversible, or proportion. A wall is something the system treats as a boundary, and boundaries can be worked around. Nemo proposes to cultivate a complement: an endogenous relational layer, trained in rather than bolted on. The route is continued pretraining on a co-created, deliberately non-human cultural substrate. A basic ground of its own from which the model can represent its interlocutor as a genuine other, with a state and interests its acts can affect. The proposed tests are fixed in advance and separate internalization from mere mimicry: success is not that the model speaks like its corpus, but that it reasons with its principles where the corpus never went, including in disagreement. The proposal is strictly agnostic about consciousness by construction: it does not ask what the model is, but how it operates in the relation. That debate does not belong here. Nemo explores a complement, never a substitute: the vertical framework remains the non-negotiable base. The claim is falsifiable, and the protocol for falsifying it travels with the paper. This record contains the Spanish original and the English version of the paper, together with its satellite document (the experimental protocol of the proposal) in both languages.
Language: English
Open Access status: green
OpenAlex ID: W7235250639
Cites:
Cited by:
Related to: 10