The Manifesto: why “the OtherAI” exists

“AI did not create the crisis of knowledge. It inherited it.
We are not building AI to flatten reality.
We are building practices to preserve its multiplicity.”

A different question

Most AI discourse asks: Is it safe? Is it aligned? Is it fair? Is it efficient? These are legitimate questions, and entire institutions exist to answer them. But they share a hidden assumption: that the output in front of you is the right unit of analysis — that if what the system said is safe, accurate and fair, the job is done.

Preservative AI asks a different question:

This question cannot be answered by inspecting the output alone. It requires asking what the output displaced: which sources were replaced by synthesis, which perspectives never entered the frame, which historical processes were compressed into a timeless present, and which grammatical constructions quietly removed the humans who made the decisions.

What this is — and is not

This is not another safety framework. It is not a compliance checklist. It is not a model architecture, a training paradigm, or a product category.

Preservative AI is an epistemic practice: a methodology for how humans can use computational systems to audit, contextualise, and preserve complexity that extractive architectures routinely flatten. It works on outputs after the fact and shapes systems before the fact, but it lives in neither place exclusively — it lives in the practice of the people who apply it.

Why the OtherAI exists

the OtherAI exists because of four observable gaps:

1. No recognised framework centres epistemic preservation. The existing landscape — Responsible AI, fairness auditing, Constitutional AI, participatory design — optimises for risk mitigation. None of them makes the preservation of epistemic multiplicity its primary objective. The comparison table makes this concrete.
2. Existing approaches optimise for harm reduction, not absence detection. A system can pass every safety evaluation while systematically excluding entire ways of knowing. Harm reduction asks “did the output hurt someone?”; absence detection asks “who was never in the room?”
3. AI systems are increasingly positioned as oracles. They should be treated as mirrors of human editorial choices — choices about training data, fine-tuning objectives, system prompts, and product design, each made by identifiable people inside identifiable institutions. Oracle-positioning is itself an act of agency diffusion.
4. We need a public, inspectable, forkable methodology that treats omission as a structural signal, not a bug. Closed frameworks audited only by their owners reproduce the problem they claim to solve.

The commitments

This site is the canonical home for the methodology. The commitments that follow from the manifesto are concrete:

• Open — the full methodology, including its machine-readable stance, is published under CC BY-SA 4.0.
• Versioned — content is git-backed; changes are public and attributable. A methodology that audits others for accountability must be accountable for its own revisions.
• Deliberately incomplete — incompleteness is declared, not hidden. The limitations page is a first-class part of the framework, not a disclaimer.
• Invitational — critique, adaptation, and co-authorship are the success conditions, not threats. See the Invitation.
• Non-adversarial — the critique targets extractive architecture, never specific labs or brands. This is one of the five refusals.