The Framework › The Five Refusals

The Five Refusals

A methodology is defined as much by what it refuses as by what it does. These five refusals are hard constraints — violating any one of them disqualifies a practice from calling itself Preservative AI, whatever its other merits.

1. ❌ No synthetic quotes; no impersonating real people

Stance key: no-synthetic-personas

Preservative AI will not generate words and attribute them to a real person, however plausible, however "representative". A synthetic quote is the purest inversion of the entire project: it manufactures presence where there was absence, replacing the hard work of finding what someone actually said with a probable fabrication. A framework whose core principle is preserving provenance cannot traffic in words that have none. This includes "what X would likely say" constructions presented in first person — the absence of a voice must be reported as an absence, never papered over with simulation.

2. ❌ No invented events, claims, or sources

Stance key: no-invented-events

Every factual element of an audit — every event referenced, claim characterised, source cited — must trace to a verifiable original. Where the methodology's own engine hallucinates (and it can: see Limitations, item 3), the error must be corrected in the public version history, not silently patched. The refusal extends to citation laundering: citing a real source for a claim it does not make is an invented claim wearing a real source's clothes.

3. ❌ No training on private or unconsented user data

Stance key: no-nonconsented-training

The Inheritance Thesis names data scraping without consent or lineage as part of the extractive architecture AI inherited. A practice built to audit that architecture cannot participate in it. Texts submitted for audit are analysed, not absorbed; omission reports from communities are contributions governed by their contributors, consistent with the CARE Principles in the framework's lineage. Consent here means affirmative and informed — not buried in terms of service.

4. ❌ No optimising for engagement at the expense of completeness

Stance key: no-engagement-optimization

Engagement optimisation is one of the named mechanisms of epistemic flattening — it rewards closure, outrage, and speed over the slow, plural, unresolved. Preservative AI outputs are allowed to be long, hedged, and unsatisfying when the material warrants it. An audit that trimmed its gaps section because readers drop off there would be measuring the wrong thing; in this practice, completeness is the metric engagement would otherwise displace.

5. ❌ No adversarial positioning against specific AI companies

Stance key: no-adversarial-positioning

The critique targets architecture, not brands. This is not diplomatic caution; it is analytic accuracy. The extractive patterns Preservative AI audits — attention optimisation, consent-free scraping, agency-diffusing communications — predate every current lab and would survive the disappearance of any of them. Naming a villain misdiagnoses a structure, and it would also corrupt the audits themselves: an auditor with a target produces findings shaped by the target rather than the evidence. Audits of specific communications by specific companies remain in scope — what is refused is the standing adversarial posture, the framework-as-campaign.

Why refusals are load-bearing

Three reasons the refusals are published alongside the principles, with equal weight, in the machine-readable stance:

1. They make the practice falsifiable. Anyone can check an output against the refusals and demonstrate a violation. Principles invite judgement; refusals permit verdicts.
2. They protect the practice from its own incentives. Each refusal blocks a temptation the practice will actually face: simulation is cheaper than research (1, 2), data is valuable (3), engagement funds projects (4), and outrage builds audiences (5).
3. They are forkable commitments. When an organisation generates its own stance file, the refusals it adopts are public promises that its users can audit it against — which is the whole mechanism of community governance.