The Diagnosis: what Preservative AI addresses

Four structural problems, none of which is owned by any single company, and none of which is detectable by accuracy metrics alone.

1. The Inheritance Thesis

AI did not create:

• Labour displacement from automation — a process running for decades before the first transformer;
• Erosion of user rights and the first-sale doctrine — ownership replaced by licence, licence by subscription;
• Attention optimisation over completeness — the engagement economy predates every chatbot;
• Weakened antitrust enforcement — concentration that made a handful of firms the default substrate for computation;
• DMCA §1201 locking out repair and modification — the legal architecture that criminalised inspecting the machines we own;
• Data scraping without consent or lineage — extraction normalised long before it was called training data.

These are decades of policy and corporate decisions, each made by identifiable institutions and contestable at the time. AI inherited this extractive architecture. It optimised it. It did not invent it.

The Inheritance Thesis matters for two reasons. First, it locates accountability accurately: blaming "AI" for the crisis of knowledge is itself a form of agency diffusion — it converts policy choices into technological inevitability. Second, it explains the framework's non-adversarial refusal: if the problem is architectural and inherited, vilifying individual labs misdiagnoses it.

2. The Epistemic Flattening Problem

Modern LLMs are trained to:

• predict probable text — which structurally favours the already-said over the marginal;
• remove friction — which removes precisely the resistance that signals contested knowledge;
• optimise for engagement and speed — which rewards closure over open questions;
• synthesise into consensus — which converts disagreement into a smooth average no party actually holds;
• erase ambiguity, contradiction, and historical contingency — which presents the present as the only present that could have happened.

The result is output that is coherent but impoverished: it answers the question asked, but erases the questions that should have been asked. Flattening is not a malfunction — it is the objective function working as designed. That is why it cannot be fixed by better training alone, and why Preservative AI operates as a methodological layer on top of whatever system is in use.

3. The Agency Diffusion Problem

Corporate and technical communications increasingly rely on four grammatical mechanisms:

These are not stylistic choices. They are grammatical mechanisms that remove human accountability and make technical progress seem inevitable. Each one converts a decision (made by someone, contestable by someone) into a condition (ambient, weatherlike, beyond appeal). Suffixscape is the diagnostic instrument built to detect them at scale.

4. The Omission Blindspot

Existing evaluation frameworks track, with real sophistication:

• bias in training data,
• hallucination rates,
• safety guardrails,
• fairness metrics.

They rarely track:

• whose perspectives were structurally excluded — not under-weighted, but never present;
• what historical processes produced the present framing — why this question, asked this way, now;
• which linguistic patterns diffuse accountability — the grammar audit no benchmark performs;
• how multilingual support flattens culturally specific reasoning — translation as category-replacement rather than meaning-transfer.

The asymmetry has a simple cause: presence is measurable with standard tooling; absence requires a theory of what should have been there. The eight lenses are that theory, made operational.