Entry 02 of 44
Alignment
lineare — to reduce to a straight line; ad + linea
Latin
From Old French alignier, from Latin ad- (to) + lineare (arrange in a line). Root concept: components in correct geometric or mechanical relation to each other — mutual correspondence.
Established meaning
Physical: parts in correct mutual position. Political (from 1934): nations in agreed relation. Organizational: goals of different parties in correspondence. The term implies at minimum two parties whose positions are being brought into relation.
AI industry deployment
An AI system that produces outputs consistent with developer-specified values and objectives. Framed as alignment with 'human values' — a phrase that obscures that specific humans are making specific choices about whose values are operationalized.
Semantic gap
Mutuality
The original root requires two parties in relation. Industry usage collapses this to one party (the developer) setting the standard and the system conforming to it. The question 'aligned with whom, defined by whom, accountable to what?' has been removed from the term.
Institutional drift
MIRI naming, November 2014 → Stuart Russell coinage → OpenAI/Anthropic adoption, 2017–2022
- 2014October–November: MIRI adopts the term — original meaning narrow and specific.
- 2017OpenAI begins using "alignment" in research.
- 2021Anthropic founded; alignment becomes core institutional language.
- 2022Anthropic's Constitutional AI paper — "alignment" operationalized as AI self-critique against a developer-written constitution.
- 2023Industry-wide adoption renders the term's original precision meaningless through ubiquity.