A Cross-Entropy Audit (CEA) for Consciousness Tests Moved from Brain to Large Language Model

A Cross-Entropy Audit (CEA) for Consciousness Tests Moved from Brain to Large Language Model

Pieter van Rooyen proposes the cross-entropy audit (CEA) for checking whether brain tests of consciousness give reliable decisions when applied to large language models (LLMs). The paper defines consciousness here as access: whether information becomes available for report and use. Because brain recordings and LLM activations use different measurement units, the audit compares how well competing models predict held-back responses, using a score difference in nats per trial. It also makes explicit five choices in the test design and checks decisions on simulated data where the right answer is known. In the paper’s example, carrying two brain-data models over unchanged led the test to find two states in 989 of 12,000 smoothly varying datasets; the audited test found two states in none, and detected 599 of 600 two-state datasets. Van Rooyen reports an important qualification: the audited test’s uncertainty interval was too narrow at six of twelve smoothly varying settings and is being repaired. The audit offers a way to examine a transferred test’s decisions, not a finding about whether LLMs are conscious.

Publication: 3f00ca6f-f5ee-4f4d-880a-c0335b6fc1ed

attribution — pass: The reading correctly attributes the proposal and results to Pieter van Rooyen, the sole listed author. One limitation says 'The authors report'; the smallest repair is to change it to 'The author reports' or 'Van Rooyen reports'. This does not misattribute the claim.

Authors:Pieter van Rooyen

summary — uncertain: The summary is faithful, but a non-specialist may struggle with its central terms. The test's decision, two states versus smoothly varying responses, is never explained. The source says a decision is wrong when the test names two states for smoothly varying responses, or the reverse. Without that, the 989 of 12,000 result is hard to interpret. 'Competing models' could be mistaken for language models rather than statistical models of trial responses. 'Nats per trial' needs a term tooltip. Repair: add one plain sentence on what the test decides and that the models are statistical. The empty question is acceptable.

The decision is wrong when the test names two states for responses that vary smoothly, or the reverse.

date — pass: The displayed original publication date of 2026-09-14 matches the v1 submission date. It is not the date of the later v3 revision (2 Nov 2026) or the feed date.

Submitted on 14 Sep 2026 (v1), last revised 2 Nov 2026 (this version, v3)

support — pass: All numeric results, the held-out scoring method, the five design choices and the uncertainty-interval caveat are directly stated in the abstract. The framing that this is an audit method, not a verdict on LLM consciousness, is a fair and appropriately limited reading. The reading also stays within the simulated-data scope. No unsupported benefits or costs are implied.

Third, the audit runs the test on simulated data built to follow one model, so the right decision is known, and counts wrong decisions.

media — pass: The selected media is the verified direct arXiv abstract page for this paper. It is the most direct access to the source.