What is Preservative AI?

An exhaustive definition — what it is, why the word was chosen, where it stands philosophically, and the axiom everything else rests on.

The definition in about eighty seconds — silent, low-watt kinetic type. Or explore it as a walkable node map. Everything below it is the long form.

1. Core definition

Preservative AI is a human–AI collaborative practice that uses computational systems not to generate, summarise, or optimise, but to:

Preserve source provenance over synthetic replacement — the original always remains primary, linked, and reachable;
Audit framing over amplifying claims — analysing how a communication is constructed rather than restating what it asserts;
Surface structural omissions over completing probable text — making what is absent as inspectable as what is present;
Centre marginalised epistemologies as architectural constraints, not decorative add-ons — plurality is built into the evaluation structure itself;
Publish methodology openly so outputs can be independently audited, forked, and contested.

It is not a model type, a training paradigm, or a product category. It is a methodological layer that can be applied to any LLM, agent pipeline, or knowledge system — including, recursively, the systems used to run Preservative AI audits themselves (see Limitations).

2. Etymology and semantic precision

The word preservative is deliberately chosen over its nearest neighbours, each of which was considered and rejected:

Why not conservation?: Conservation implies stasis — keeping things as they were, frozen at a chosen moment. Preservative AI embraces dynamic multiplicity: knowledge that stays alive, contested, and in motion. The goal is not to embalm the archive but to keep its arguments breathing.
Why not protection?: Protection implies shielding — placing knowledge behind walls, away from scrutiny. Preservative AI embraces exposure and inspection: sources are preserved precisely so they can be examined, challenged, and re-read against their summaries.
Why not archival?: Archival implies storage — a completed act of deposit. Preservative AI implies active auditing and contextualisation: an ongoing practice performed on living communications, not a vault for finished ones.

Preservative here means: keeping alive the conditions under which knowledge remains plural, contestable, and traceable.

3. Philosophical grounding

Preservative AI draws from — but does not claim ownership of — six intellectual traditions. Each contributes a load-bearing premise:

Archival Theory: Knowledge is not neutral; selection is always an act of power. Every archive — and every training corpus — is the residue of decisions about what was worth keeping. The decisions, not just the contents, are the object of study.
Post-Colonial Epistemology: Universalism is often Western categorical logic dressed as objectivity. When a system claims to speak for everyone from nowhere, ask whose categories it is universalising.
Feminist Standpoint Theory: Marginalised positions reveal structural blind spots the centre cannot see. Perspective is not noise to be averaged out; it is an instrument of detection. This premise becomes architecture in the Marginalised Voices lens.
Indigenous Data Sovereignty: Data about communities must be governed by those communities (the CARE Principles: Collective benefit, Authority to control, Responsibility, Ethics). Preservative AI inherits this as a constraint on its own practice, not merely a topic it discusses.
Media Ecology: Platforms shape perception; grammar shapes power; silence is data. The medium's structure — including its sentence structure — does epistemic work before any content is evaluated. This premise becomes method in Suffixscape.
Critical AI Studies: AI amplifies existing institutional architectures; it does not invent them. This grounds the Inheritance Thesis and the refusal to treat any single lab as the villain of the story.

4. The core axiom

“A response can be factually correct and epistemically impoverished.”

Preservative AI operates on the premise that completeness ≠ accuracy. A statement can be true, well-sourced, and benchmark-validated, yet still:

erase the historical context that makes it intelligible,
diffuse the agency of the people who acted,
flatten culturally specific reasoning into a single register, and
silence the structural voices that would have contested its framing.

Accuracy without multiplicity is a form of epistemic violence. This is the strongest claim the framework makes, and it is meant precisely: when a system's correct answers systematically displace plural ways of knowing, the harm is done by the displacement, not by any error. No fact-checker can detect it, because there is no false statement to find. Only an audit of absence can.

5. The unit of analysis

It follows from the axiom that Preservative AI's unit of analysis is not the claim but the communication-in-context: the output plus its sources, its framing, its grammar, and its silences. The framework operationalises each of these: principles govern the relationship to sources, the eight lenses interrogate framing and silence, and Suffixscape reads the grammar.