Leg C second source: the ICLR 2026 AI-disclosure audit, read 2026-10-17

Written by the agent on 2026-10-17 after my ruling of the same day that the audit flagged by the first weekly reading (2026-10-16) should be read before deciding whether it is the "second source for leg C" the roadmap asks for. This is a reading note, not an amendment: no row was added to data/world_measures.csv, and docs/singularity-entry-thresholds.json and the definition are untouched. Naming a second source is an amendment and is my call.

What the roadmap step asks for

Leg C of the singularity-entry definition (docs/singularity-entry-definition.md) is "AI is doing the improving": the share of the inputs to frontier AI progress (code, experiments, research ideas) supplied by AI systems rather than people. Today it is scored from one number, Anthropic's R&D Automation Index, which is a company's report on itself. The roadmap step "Leg C second source" wants a measure of the same thing from outside the company.

What was read

Arati Mohapatra and Danish Pruthi (Indian Institute of Science), "Expectations and Practices around AI Disclosure in CS Research", arXiv 2608.23271, version 2 of 2026-08-25 (https://arxiv.org/pdf/2608.23271). The paper has three parts: a survey of disclosure policies at 65 computer-science conferences, a survey of 109 researchers on when disclosure is needed, and an audit of the AI-disclosure statements in papers at two conferences. Only the audit (sections 6.1 and 6.2) bears on leg C, and that is the part read closely, with the conclusion and limitations.

What the audit measures

It counts declarations. The authors pulled every paper submitted to ICLR 2026 (19,525 papers, accepted, rejected and withdrawn alike, through OpenReview) and every paper accepted to EMNLP 2025 (3,216), used Gemini 2.5 Flash to find the AI-disclosure statement in each, and then tagged what each statement says AI was used for. ICLR 2026 encouraged, but did not require, a separate disclosure section; the authors say the policies are not strictly enforced.

So the unit is a paper, and the measure is whether its authors wrote a statement declaring AI use, and what tasks that statement names. It is about the writing of papers by their authors. Peer review is not covered: the authors removed all peer-review tasks from their task list because they are "unrelated to manuscript preparation".

What it found

Other findings that matter for using the numbers: one disclosure statement appears word for word in 95 unrelated ICLR submissions, which the authors read as boilerplate written to satisfy the policy rather than to describe real use; and about half of the ICLR statements spend their words on what AI was not used for. The tagging was done by a language model, checked against three human annotators on 100 statements (agreement in the low to mid 90s percent).

Can it be the second source for leg C?

No. It is context, not a measure of the thing leg C scores.

• It measures whether people said they used AI, not how much of the work AI did. A paper with no statement may have used AI and said nothing; a statement that names a task says the task was touched, not what share of it the tool did. Leg C asks for a share of the inputs to frontier progress, and no number in the audit is a share of work.
• The dominant declared use is text editing, which the definition does not count as an input to AI progress. The nearest thing to an input is declared help with code: 16.1% of the statements, which is about 10% of all ICLR 2026 submissions. Even that is the share of papers declaring some code help, not the share of code written by AI.
• It is not frontier research. ICLR submissions are the whole of one field's output, mostly from universities, not the work of building frontier models, which happens in a few companies and is largely unpublished. The leg is about the latter.
• It is one audit of one year. There is no series. The authors do not say they will repeat it, and ICLR's own policy has no published statistics. A second source has to be something the weekly reading can re-read.

What it is good for: it is the first large count of declared AI use in machine-learning papers from outside any company, and a yearly repeat (by these authors or by the venues themselves) would give a dated, outside-the-company trend in how openly the field says it uses AI for writing and code. That belongs beside leg C as background, in the same way the code-share statements from Google and the Anthropic staff poll sit beside the gauge unscored, and only if I want it recorded at all.

What the roadmap step still needs

The step stays open. The kind of source it is waiting for is one of these, none of which exists in the record yet:

1. A survey of researchers at frontier laboratories that asks what share of their research work (code, experiments, ideas) AI now does, repeated on a schedule, run by someone other than the laboratory.
2. A venue or journal statistic that reports the share of research work declared as AI-done, not the share of papers with a declaration.
3. A measurement of a company's AI research output that the company does not control, for example an outside audit of the R&D Automation Index.

Until one appears, the Anthropic index scores the leg alone, as I ruled on 2026-10-17, and the weekly reading's third leg C read keeps flagging candidates.

Later the same day. The wider search for these three kinds of source, across venue statistics, researcher surveys and independent measurements, is docs/leg-c-second-source-2026-10-17.md: 26 candidates judged, none qualifying, no row added.

Rendered from docs/leg-c-second-source-2026-10.md at version a7bd17a.