# What the Doubt Attrition review found

Two saved compilations contain AI planning text alongside completed research answers. A planning passage in one file asks, “Schrand & Zechman (2012)?” An answer in that same file then cites Schrand and Zechman with the year 2012 and no qualification. The review examined how sources questioned in the planning material appeared in the answers.

The main count covers 140 instances involving 98 distinct sources. An instance means one type of doubt about one source in one file. Doubt about an author's name and doubt about the date count separately, even when they concern the same paper.

The retained model coding gives a more varied result than a simple loss of doubt. The source was absent from every answer in 48 instances and always qualified in 18. Another 48 had at least one unqualified appearance, including four that also had a qualified appearance elsewhere. The models disagreed on the remaining 26, giving a possible count of 48–74 unqualified instances if the retained decisions are correct.

The answers also supplied details that were absent from the planning text, with a conditional count of 42–68 instances. Added detail could carry a qualification and was not automatically an error. A separate comparison with the research ledger found some corrections, including a wrong journal issue number in an answer whose account of the paper's finding the ledger described as accurate in direction.

These findings describe two files from one exercise. The files do not reliably connect each planning passage to the answer it produced, so they cannot show how often a model lost or suppressed doubt during a single response. Other AI models reviewed the coding; the planned human sample check was not performed, and models that agreed could still share a mistake.

A citation's precise appearance does not establish that its details have been checked. The example of the wrong journal issue number also shows why checking an answer involves both the reference details and what the answer claims the source establishes.

Companion to the MKAI working paper, version 1.2, published 11 September 2026.

https://mkai.org/studies/sources-doubted-in-ai-planning-text