Limitations & Self-Critique

A framework that audits omission must publish its own. This page is not a disclaimer; it is the framework applied to itself, and it is load-bearing — several of these limitations actively shape the methodology's rules.

The seven acknowledged limitations

1. Cultural positionality

Preservative AI is rooted in Western critical humanities — archival theory, standpoint theory, media ecology — even as it advocates plurality. Its categories (the lens divisions themselves, the very idea of an "audit") are located, not universal. The framework cannot escape this by declaration; it mitigates it structurally, by being forkable: regional adaptations with authority to change the lenses, not just translate them, are the test of whether the plurality commitment is real.

2. Language bias

Primary analysis is in English. Suffixscape's diagnostic patterns are patterns of English grammar — nominalisation, passive voice, and inflation work differently in other languages, and some of the evasions that matter most elsewhere have no English analogue at all. Non-English audits therefore require community co-auditors, not machine translation of the English method — translation of the method would itself be the flattening the method exists to detect.

3. LLM dependence

The audit engine is itself an LLM, subject to hallucination and training bias. The instrument is made of the material it measures. This is why audits must be anchored to quotable evidence (the Evidence requirement), why human review is mandatory rather than optional, and why a published audit found to contain an invented element must be corrected in public version history (Refusal 2). The dependence is disclosed, not solved.

4. Labour-intensive by design

The critical-social-theory analysis is automated: ACST — the eight-lens power-knowledge audit (Automated Critical Social Theory) — runs continuously and at scale, which is what lets the method audit tens of thousands of communications rather than a handful. The analysis proposes. What is not automated is the judgement — which surfaced absence is real, which matters, how to weigh it — and that is held human by design. This is not a failure to automate: automating judgement would centralise it, and centralised judgement is precisely the failure mode the framework exists to dissolve. Cognition here is distributed — human and machine in the loop together, not lifted out of human hands into a device. The cost is real — distributed, reflexive, compensated judgement is slower and more expensive than the extractive automation it critiques — but it buys plural, accountable judgement rather than the absence of it.

5. Not a substitute

Preservative AI cannot replace lived experience, Indigenous governance, or subject-matter expertise. An audit can show that a voice is missing; it cannot be that voice (Refusal 1). The framework's outputs are maps of absence, and a map of absence is worthless — or worse, alibi-producing — if it is treated as the presence it points toward.

6. Risk of performativity

Every accountability practice can be hollowed into checkbox compliance, and this one is no exception: a lab could publish a stance file, run perfunctory audits, and wear the vocabulary as reputation laundering. The structural defences are the public version history (hollow audits are visibly hollow over time), the right-of-reply record, and the forkability that lets communities run independent audits against an organisation's published stance. But the limitation stands: the framework can be performed, and detecting the performance requires exactly the human judgement it does not automate, by design.

7. Temporal lag

Audits capture snapshots; rapid controversies may outpace publication. A careful omission map of last week's framing may arrive after the framing has already done its work. The methodology accepts this trade deliberately — speed-optimisation is on the refused list — but accepting a trade does not erase its cost.

The doorway, not the room

“The framework points toward the room. The room is not in the framework.”

Preservative AI is a doorway, not a destination. It makes absences inspectable. Filling them requires human courage, institutional change, and centring excluded voices — none of which any methodology can supply. The most honest statement of the framework's scope: it can tell you, rigorously and reproducibly, who is not in the room and how the grammar kept them out. Everything after that is politics, in the oldest and most respectable sense of the word.