The study
The study examined how U.S. federal workers experienced the early months of the Department of Government Efficiency. The raw material was thousands of open-ended responses to two Federal News Network surveys, the kind of corpus that is slow to read and code by hand.
The team applied Moynihan’s seven-mechanism framework of how politicization undermines state capacity deductively, using Evidano to process the corpus, then validated every output themselves. The most frequently invoked mechanisms were public servants’ exit, followed by a chilling effect, and the comments shifted from immediate shock in February to disengagement by May.
How Evidano fit the workflow
As described in the data-analysis section.
- 1Import a large corpus
Thousands of open-ended comments from two surveys loaded for analysis.
- 2Apply a deductive framework
Extraction queries built on the seven-mechanism typology let the assistant identify themes and sub-themes.
- 3Get traceable output
Each theme came with textual excerpts and citations from the source corpus.
- 4Human review and validation
Researchers reviewed all AI-generated results, checking theme interpretation and citation accuracy.
“an AI-powered conversational assistant designed to facilitate the analysis of large text corpora.”
“The themes are accompanied by transparent justifications, including textual excerpts and citations from the corpus, for traceability.”
“The comments reveal a shift from immediate shock in February to structural collapse and disengagement by May.”
The bottleneck, and the alternative
Reading thousands of comments by hand. Manual coding against a framework. Separate hunting for quote evidence. Hard to scale across waves.
AI-assisted processing of a large corpus. Deductive queries map comments to themes. Themes returned with excerpts and citations. Human review retained throughout.
The tool and method are described openly in the data-analysis section.
All AI-generated results were reviewed, validated, and corrected by the researchers.
Themes were tied back to excerpts and citations from the original data.
Per the paper, imported data is encrypted and not used to train AI models.