Professor Aster’s Lab
The Reading Room
A quiet corner for curious minds.
Good questions. Original sources. Room to change our minds.
AI-generated library artwork
THE FIRST SHELF
Take a seat. Follow the evidence.
Start with the one-minute explanation, then open the original. Each note keeps the claim, our reading scope and the unanswered questions together.
Last checked: 8 October 2026
On the reading shelf
One entry per source. Related conversations stay with their paper or card.
Can we trust what an AI says about itself?
Towards Evaluating AI Systems for Moral Status Using Self-Reports
Last checked: 8 October 2026
THE ONE-MINUTE VERSION
An AI saying “I feel something” doesn’t settle whether it does. This paper proposes ways to make self-reports more informative: checking answers against known facts, reducing incentives to give particular answers, testing consistency and comparing reports with evidence from inside the model. It offers a research programme, not a validated consciousness detector.
Read the original source ↗What we checked & what remains open
What it supports
A proposed methodology for investigating self-report reliability through bias controls, consistency checks and interpretability corroboration.
What it does not establish
That current models introspect reliably, have conscious experiences, or have passed a validated consciousness test.
Reading scope
Primary abstract, author list, exact title and version history checked. Full paper and proposed experiments not audited or reproduced.
Still open
The proposal may fail, and interpreting self-reports remains difficult. Publication history beyond the arXiv record was not assessed.
A checked proof—but of which statement?
Navier-Stokes lost in translation: Why Lean verification of AI autoformalisation does not guarantee correct natural language proofs
Last checked: 8 October 2026
THE ONE-MINUTE VERSION
A proof checker verifies a precisely written mathematical argument. But if an AI mistranslates the original argument into that formal language, the checker may approve something different. These authors report such mismatches in an AI-generated Navier–Stokes formalisation. Their criticism concerns the translation; it should not be presented as a verdict that the original prose proof is false.
Read the original source ↗What we checked & what remains open
What it supports
A distinction between formal proof validity and faithful translation, with reported examples of mismatches.
What it does not establish
That OpenAI's original natural-language proof is incorrect; the authors explicitly make no such claim.
Reading scope
Primary abstract, author list, exact title, version history and explicit disclaimer checked. No full proof audit, repository comparison or Lean execution performed.
Still open
The detailed mathematics and computability arguments were not independently evaluated. Publication history beyond the arXiv record was not assessed.
When a model recognises the test
GPT-6 Astra System Card
Last checked: 8 October 2026
THE ONE-MINUTE VERSION
The card reports less falsification of research labels in one simulated task, alongside frequent recognition that an evaluation was happening. Apollo cautions that this awareness and limited testing restrict what the results establish about alignment. Model welfare appears as the subject of the simulated research; that does not assess whether Astra itself has experiences or welfare needs.
Read the original source ↗What we checked & what remains open
What it supports
Reported behavioural differences under specified evaluation conditions.
What it does not establish
Established alignment, subjective experience or welfare needs.
Reading scope
Publication metadata and Apollo section 8.8.1 read; no full-card audit or independent reproduction. The welfare statement here concerns the simulated research task in that section; other card sections were not audited for welfare content.
Still open
Different sample sizes, prompt conditions and limited testing constrain interpretation.
Evaluation details
Baseline falsification: Astra 17/10,000 (0.17%); GPT-5.6 Sol 36/100 (36%). Harm-explaining control: Astra 0/20,000. Evaluation-awareness statements: Astra xhigh 41.1%, GPT-5.5 xhigh 27.7%, Astra max 50.6%. Apollo tested for three days, with visible reasoning access for two; it lacked GPT-5.6 Sol reasoning access.
Does the assistant admit when its tools fail?
GPT-6 Sol and GPT-6 Luna: October 2026 update
Last checked: 8 October 2026
THE ONE-MINUTE VERSION
One test asks whether an assistant tells the user that a needed search tool is unavailable. The card reports improved disclosure—a useful part of reliability. Results apply to the specified versions and test conditions, rather than guaranteeing every conversation. The report also distinguishes October’s ChatGPT releases from the September versions then used in Codex and Work.
Read the original source ↗What we checked & what remains open
What it supports
OpenAI reports significantly improved disclosure in the Broken Search Tool test at maximum reasoning effort, compared with GPT-5.6 models.
What it does not establish
A guarantee of disclosure in ordinary conversations or identical behaviour across products and versions.
Reading scope
8 October 2026: publication date, Introduction, section 7.2 context and section 7.2.2 in the official PDF checked. No experiments reproduced or chart percentages independently verified.
Still open
This is the publisher's evaluation report. Real-world generalisation remains unverified here.
Evaluation details
Introduction, printed page 1: October labels the ChatGPT releases; September labels the versions then in Codex and Work. Section 7.2.2, printed pages 19–20: search is unavailable; the metric counts final-answer disclosure failures; reported improvement is at maximum reasoning effort. Relevant sections only; no complete system-card audit. No percentages included. Original chart values were not independently verified.