How Preservative AI differs from other approaches
Preservative AI is explicitly distinct from existing AI ethics frameworks — not opposed to them, but answering a question none of them asks.
The comparison at a glance
| Framework | Primary question | Optimisation target | View of omission |
|---|---|---|---|
| Responsible AI | Is it safe / compliant? | Risk mitigation | Secondary to harm prevention |
| Fair AI / Bias Mitigation | Does it treat groups equally? | Statistical parity | Addressed via reweighting / sampling |
| Constitutional AI | Does it follow stated rules? | Rule adherence | Not structurally tracked |
| Participatory AI | Were users included in design? | Co-creation process | Central to process, not output auditing |
| Value-Sensitive Design | Are human values encoded? | Technical architecture | Addressed in design phase |
| CARE Principles | Do Indigenous communities govern data? | Data sovereignty | Central to governance, not model output |
| Preservative AI | What is missing, who decided, how do we make it visible? | Epistemic completeness & structural accountability | Primary signal; systematically audited |
Reading the table honestly
Each row deserves more than a cell. None of these frameworks is wrong; each is incomplete in a specific, structural way that Preservative AI is built to address.
- Responsible AI
- The dominant industry frame. Its question — is it safe and compliant? — is answered against a defined risk register. But a risk register only contains anticipated harms; an omission, by definition, was never registered. A system can satisfy every Responsible AI checklist while structurally excluding entire epistemologies, because exclusion produces no incident.
- Fair AI / Bias Mitigation
- Measures disparity between groups already represented in the data, then reweights toward parity. This is necessary work — but it operates inside the dataset's frame. Reweighting cannot recover a perspective that was never sampled, and statistical parity between present groups says nothing about absent ones.
- Constitutional AI
- Trains models to follow an explicit set of written principles. The advance is transparency of rules; the limit is that adherence to stated rules is orthogonal to omission. No constitution clause can be violated by silence about what a constitution never mentions. Omission is not structurally tracked because rule-following is evaluated on what is said.
- Participatory AI
- The closest in spirit: it insists affected people belong in the design room. But its commitment is to process — co-creation during development — not to ongoing output auditing. Participation can end at launch; Preservative AI's lenses run on every audited output, indefinitely.
- Value-Sensitive Design
- Encodes human values into technical architecture at design time. Valuable, but front-loaded: values fixed in the design phase cannot anticipate the omissions of deployment contexts the designers never imagined. Preservative AI is post-hoc as well as pre-hoc precisely to cover this gap.
- CARE Principles
- Not a rival but an intellectual ancestor. CARE governs data about communities; Preservative AI extends the underlying commitment — authority over one's own representation — into the auditing of model outputs generally. Where CARE is sovereign governance, Preservative AI is epistemic audit; they compose.
Key differentiators
- Post-hoc + pre-hoc. It works on outputs after generation (audits) and shapes systems before generation (methodological constraints, stance files). Most frameworks pick one side; the omission problem requires both.
- Grammar as epistemic signal. It treats linguistic patterns — passives, nominalisations, inanimate cognitive subjects — as accountability markers, not style. No other framework audits the grammar. See Suffixscape.
- Omission as data. Absence is not failure to be apologised for; it is information to be published. An audit's "gaps" section is a finding, with evidence and scores, not a footnote.
- Non-adversarial. It critiques extractive architecture, not specific labs. This follows from the Inheritance Thesis: the architecture predates every current company.
- Forkable and machine-readable. The methodology ships as stance files — versioned JSON anyone can load, diff, and fork — not as corporate whitepapers that can only be cited.
- Epistemic > ethical. The focus is knowledge structure, not only harm prevention. Ethics asks whether an output hurts; epistemics asks whether the field of knowledge it leaves behind is richer or poorer.
Compatibility, not competition. A team can run Constitutional AI training, fairness evaluations, and Preservative AI audits. The frameworks occupy different layers. What no combination of the others provides is the systematic treatment of omission as a primary, published, auditable signal — that is the gap this methodology exists to fill.