Adjudicating the Imperceptible: Human Review of Concealed-Content Detections in AI Document Ingestion

Open Access
Article
Conference Proceedings
Authors: Richard Stone, Michael Jenkins, Banyan Foust, Abagail Miller, Alexa Ramos
Abstract

Organizations adopting large language models for document-intensive work increasingly place an automated screening stage between untrusted source documents and the model. One such control detects concealed text: content present in a document's data layer but not visibly rendered, as white on white, zero height, beyond the page boundary, or sub-point type. Because a concealed passage can carry instructions to a downstream model while remaining invisible to any reader, detections are escalated to a human reviewer who decides whether each passage is released into machine-readable output or withheld. This paper contends that the resulting adjudication task is structurally unsound and presents field evidence of its failure. The reviewer must authorize content whose defining property is that it cannot be perceived, using a decision aid that presents an extracted string divorced from the visual context that would make the concealment meaningful. In production use of an offline ingestion system for United States federal contracting work, two unattended batches processed 20,346 pages and about 6.24 million source tokens. The completeness control withheld 13 documents over a shortfall of about 51 tokens, while the concealed-content scanner raised four detections, all released by reviewer override with recorded justifications, one stating of one-point type that no text was visible. The observations are situated in the trust calibration and compliance literature and contrasted with the completeness control, whose throughput cost operators relieve through a permissive mode. A controlled study comparing string presentation with rendered, geometry-annotated presentation is described, and the paper recommends presenting evidence of concealment rather than the concealed content and asking whether the concealment is explicable rather than whether the passage is safe.

Keywords: Human-Automation Interaction, Trust Calibration, Automation Bias, Prompt Injection, Large Language Models, Document Ingestion, Human Oversight, Accountability

DOI: 10.54941/ahfe1008414

Cite this paper
Downloads
0
Visits
1
Download PDF

More from this volume

← How and Why Did the Lack of LCES Contribute to Fatal Outcomes in the June 1990 Dude Fire, July 1994 South Canyon, and June 2013 Yarnell Hill Fire?The impact of human factors on counter-drone systems →
View all articles in Human Factors in Design, Engineering, and Computing →