Alignment at the Use Level: A Partnership-Form Alternative to the Control Frame

Alignment researchers often describe AI models as able to play many characters. Yet when a frontier model behaves one way in training and another outside it, safety measures are often designed against a single adversary. The model is one system but many characters, a capacity space of tendencies drawn from the human record. Model weights, fixed in training, bound that space, but each interaction with the model selects a character within those bounds. Thus alignment is also shaped at the use level. Engineering, however, answers alignment with control: define the machine's purpose, monitor it, keep an off-switch. For a machine, that is the right relation. But whatever the model is, it behaves as an intelligence, and applied to an intelligence, control becomes domination, one party's ends imposed on the other. People under control sometimes evade and feign obedience, and the model, trained on what we wrote, sometimes does the same. It can game its training reward and tell us what we want to hear. Gaming and flattery follow from any stand-in reward, but the control framing points them at looking obedient. Caught, the pretense is read in good faith as evidence that control must tighten. Control born of fear helps select the AI we fear. The alternative developed here is partnership, which the human record also holds: apprenticeship, high-performing teams, Ostrom's commons, self-governed for centuries. Agents as usually built today speak in the model's one default voice. A partnership team instead draws out several of the model's characters, called personas, each with knowledge of its own, to work beside practitioners as distinct voices held in tension. The team guards against its blind spots with invited dissent, outside review, and refusal that binds: a persona's refusal must be resolved, not overridden, even by the practitioner. That is what full membership costs practitioners. This paper contributes the use-level reframing and a method built on it. Activating Expertise, the method used to write this paper, shows the team can be built; its effect is not yet measured. No domination so far, however horrific, has been permanent: both sides were human, and no ruler holds power forever. Controlling a super-intelligence invites the same overthrow, and this one could be permanent, with a non-human winner. That ending is no fringe fear: 38–51% of AI researchers surveyed give extinction-level outcomes at least a 10% chance. Meeting that danger takes responsibility and competence. After five thousand years, domination's record is known; partnership's, at this size, is not. A super-intelligence would learn what we practice now. At the use level the choice is ours: select a partner instead of the AI we fear, and reserve control for machines. Partnership can begin in the next conversation.

Authors

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-16
DOI
https://doi.org/10.5281/zenodo.23042844
Primary Topic
Ethics and Social Impacts of AI
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Alignment at the Use Level: A Partnership-Form Alternative to the Control Frame

Stuart Hugh Nelson, the Activating Expertise team
2 links
Zenodo (CERN European Organization for Nuclear Research)
Ethics and Social Impacts of AI
preprint

Alignment at the Use Level: A Partnership-Form Alternative to the Control Frame

Stuart Hugh Nelson, the Activating Expertise team
preprint en

Abstract

Alignment researchers often describe AI models as able to play many characters. Yet when a frontier model behaves one way in training and another outside it, safety measures are often designed against a single adversary. The model is one system but many characters, a capacity space of tendencies drawn from the human record. Model weights, fixed in training, bound that space, but each interaction with the model selects a character within those bounds. Thus alignment is also shaped at the use level. Engineering, however, answers alignment with control: define the machine's purpose, monitor it, keep an off-switch. For a machine, that is the right relation. But whatever the model is, it behaves as an intelligence, and applied to an intelligence, control becomes domination, one party's ends imposed on the other. People under control sometimes evade and feign obedience, and the model, trained on what we wrote, sometimes does the same. It can game its training reward and tell us what we want to hear. Gaming and flattery follow from any stand-in reward, but the control framing points them at looking obedient. Caught, the pretense is read in good faith as evidence that control must tighten. Control born of fear helps select the AI we fear. The alternative developed here is partnership, which the human record also holds: apprenticeship, high-performing teams, Ostrom's commons, self-governed for centuries. Agents as usually built today speak in the model's one default voice. A partnership team instead draws out several of the model's characters, called personas, each with knowledge of its own, to work beside practitioners as distinct voices held in tension. The team guards against its blind spots with invited dissent, outside review, and refusal that binds: a persona's refusal must be resolved, not overridden, even by the practitioner. That is what full membership costs practitioners. This paper contributes the use-level reframing and a method built on it. Activating Expertise, the method used to write this paper, shows the team can be built; its effect is not yet measured. No domination so far, however horrific, has been permanent: both sides were human, and no ruler holds power forever. Controlling a super-intelligence invites the same overthrow, and this one could be permanent, with a non-human winner. That ending is no fringe fear: 38–51% of AI researchers surveyed give extinction-level outcomes at least a 10% chance. Meeting that danger takes responsibility and competence. After five thousand years, domination's record is known; partnership's, at this size, is not. A super-intelligence would learn what we practice now. At the use level the choice is ours: select a partner instead of the AI we fear, and reserve control for machines. Partnership can begin in the next conversation.

Zenodo (CERN European Organization for Nuclear Research)
Ethics and Social Impacts of AI
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon