The Truth Surface: Measuring how much an AI claims to know, not only what it says

Abstract

Evaluations of AI systems ask whether an answer is true. They rarely ask whether it claims to know as much as the evidence allows: no more, and no less. This paper argues that the second question is where many of the most consequential failures of AI answers now sit, and that it can be measured without settling any contested question of fact. Borrowing the formalism of subjective logic, each claim in an answer is represented as a point on a surface with two independent directions: verdict, from false to true, and height, from settled to unknown. An answer can be right in verdict and wrong in height. On this surface the paper defines two content-neutral measures. The height error separates false certainty from manufactured doubt. The collapse index detects credibility read from who speaks rather than from corroboration, which gives propaganda an operational definition. It shows that a reader’s prior enters only through height, so the evidence itself fixes how far honest readers may disagree. And it proposes that the measuring be done by a swarm of AI and human scouts whose trust is a public ledger, kept per domain and per model version, and whose quorum must cross lineages. Fifteen equations, eight falsifiable predictions and a reference implementation are provided. The paper does not solve the problem of truth. It makes two of its most dangerous symptoms measurable. What this paper claims Height as an object of evaluation and alignment. An answer’s stated uncertainty should be scored against the state of the evidence, not only its verdict against a fact or its confidence against its own accuracy. A symmetric definition of epistemic manipulation. Manipulation falsifies height, in two directions: false certainty pushes it down, manufactured doubt pushes it up. Propaganda as collapse. Credibility read from the speaker instead of from corroboration collapses a two-dimensional surface into a line. The collapse index measures it. The room for conviction. Readers who share the evidence and differ only in their priors can disagree by at most the height of a claim. Legitimate pluralism is bounded by what is not known. An accountability design. AI scouts with per-domain, per-version trust ledgers, a quorum that must cross lineages, revision receipts, and decay without deletion. Section 13 of the paper lists the prior work on which all five are built, and says what, to the author’s knowledge, is new. Files: the paper (PDF) and truth_surface.py, a dependency-free Python reference implementation that reproduces every number in the text. This paper was written in part with Claude (Anthropic), an AI model, which also wrote the reference implementation. The ideas, editorial decisions and responsibility for the content are the author’s; the division of contributions is set out in section 13 of the paper.