Sources doubted in AI planning text
How they appear in two preserved research compilations
Published 11 September 2026 · Version 1.2
Abstract
Two preserved AI research compilations contain visible planning text that questions sources cited in the completed answers. This analysis compares their presentation within each file. The primary population comprises 140 instances, each combining a source, a type of doubt and a file; these cover 98 distinct canonical sources. The retained model coding fixes 114 instances: the source is absent from every answer in 48, always qualified in 18, and unqualified at least once in 48. Another 26 remain disputed, giving a conditional range of 48–74 instances (34.3–52.9 per cent) with an unqualified appearance. The answers contain added detail in 42–68 instances. These ranges assume the retained coding and membership decisions are correct; the planned human sample check was not performed. The files do not establish which planning session produced which answer, so the comparison describes co-occurrence within the compilations. It cannot measure doubt disappearing during generation. The practical distinction is between the form of a citation and evidence that its particular claims have been checked.
Suggested citation
Foster-Fletcher, R. (2026). Sources doubted in AI planning text: How they appear in two preserved research compilations. MKAI, working paper, version 1.2, 11 September 2026. https://mkai.org/studies/sources-doubted-in-ai-planning-text
PDF size: 175 KB
Record metadata
- Type
- Working paper
- Theme
- AI capability and evidence assessment
- Status
- Published working paper
- Author
- Richard Foster-Fletcher
- Publisher
- MKAI
- Version
- 1.2
- Research classification
- Descriptive coding of preserved material
Methods
- Descriptive coding of source appearance, qualification and added detail
- Amended model review with conditional ranges retaining disputed cases
- Separate comparison with a frozen research ledger
Evidence base
Two preserved compilations containing ten completed answers and eight planning blocks. The primary population contains 140 source-doubt instances covering 98 canonical sources; 21 abandoned instances are reported separately.
Evidential limit
The ranges assume retained model coding and source membership are correct. The original human sample check was not performed. The files do not establish planning-to-answer session identity, so the results do not measure causal doubt loss or a general model error rate.