Pillar 2: Authority

When AI can’t tell who is allowed to speak, relevance replaces legitimacy

AI systems are fluent, fast, and increasingly embedded in decision making. But they share a structural blind spot: they cannot recognise authority. They can retrieve relevance, but they cannot evaluate legitimacy. They treat every piece of text as an equal contributor to an answer, regardless of who wrote it, what domain they belong to, or whether they have any right to define truth in that area.

This is not a limitation of any particular model. It is a structural property of vector embeddings, full text search, and large language models. These systems measure similarity, overlap, and pattern likelihood – but none of them contain a mechanism for understanding who is allowed to speak.

That failure becomes acute in financial institutions, where knowledge is not flat. It is structured by domain boundaries, seniority, jurisdiction, regulatory standing, and institutional hierarchy. Equity analysts speak for equities. Macro analysts speak for macro. Country specialists speak for their geography. Risk and compliance speak with overriding authority. These distinctions are not cosmetic; they are the foundation of how the institution manages risk.

AI systems erase them.

A retrieval engine can easily surface an equity analyst speculating on FX, a macro analyst drifting into commodities, a country specialist generalising about another geographic region, or a junior analyst contradicting a senior one. All of these may be semantically close to the question. None of them may be legitimate contributors to the answer. The system cannot tell the difference between commentary and expertise, adjacency and jurisdiction, relevance and authority.

The problem is deeper than individual analysts. Authority can sit at multiple levels: a publisher, a department, an author, or a source type. A central bank statement, a regulatory filing, a risk department’s guidance, a senior analyst’s note, and a speculative blog post are not epistemically equivalent. But to an embedding, they are just vectors. To full text search, they are just strings. To an LLM, they are just patterns. The authority structure behind them is invisible.

When authority is not explicitly modelled, the system invents its own hierarchy. It elevates voices based on stylistic confidence, frequency of mentions, embedding density, or accidental correlations. It suppresses voices that are institutionally critical but semantically sparse. It blends domains that should never be blended. It produces answers that are fluent and well structured – but shaped by the wrong speaker.

This is not a technical quirk. It is a governance failure. The institution loses control over who defines truth inside its own walls.

The implication that every AI system that synthesises information is also redistributing authority. If you do not define who is allowed to shape an answer, the model will decide for you.