The Unintended Straw Man

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Even when using accurate sources, language models can alter the substantive content of a reference text. This article examines how unintentionally distorted versions can arise through the omission of crucial conditions or the addition of unsupported assumptions, and why such distortions may recur even after correction. To this end, the concepts of the "principle of maximally omissible steps" and "negative compression" — along with the distinction between addressability and graspability — are developed and linked to current findings regarding overgeneralization and self-correction in language models. In legal contexts, a particular risk arises because the use of a correct source and citation does not guarantee that the statement derived from them remains supported by that source. Based on this analysis, the article formulates testable predictions and offers practical guidelines for verifying AI-generated assessments.