A newly developed audit framework demonstrates that excerpt-level contributions from large language models in qualitative health research can be systematically recorded, verified, and audited at corpus scale. Qualitative health studies are typically evaluated only at the broad study level, making it difficult for outside readers to trace interpretive claims back to the underlying interview material. When generative models take part in qualitative data analysis, researchers face new methodological challenges regarding what source material was provided to the system, what the model suggested, and what the human analyst ultimately retained.
Existing reporting guidance for artificial intelligence in research primarily tracks model identities, software prompts, and technical parameter settings rather than this granular excerpt-level pathway. To address this methodological gap, a study published in BMC Medical Research Methodology decomposed excerpt-level interpretation into discrete analytic operations. For each analytical step, the framework establishes what evidence must be logged or inspected whenever a local or cloud-based large language model participates in the qualitative workflow.
The resulting audit framework incorporates seven core structural components designed to keep computational interpretation open to inspection. These elements include formulating an explicit analytic question, declaring the precise status of every model input, and enforcing a fixed limit on the textual context sent to the system. The protocol also requires defining distinct units of meaning, applying deterministic verification to every quotation proposed by the model, referencing a closed catalogue of named warnings, and generating supervision records that never treat an unedited machine proposal as accepted.
To evaluate record production and routing, the framework was tested in an automated run involving 206 excerpt records drawn from 18 interviews concerning territorial health inequalities. This test configuration implemented a local open-weight model to maintain strict transcript confidentiality. While the demonstration tested production, storage, and record routing across the entire sample, specific mechanisms such as context reduction, trajectory processing, and active human editing were not exercised during this initial computational phase.
Across 835 evaluated units of meaning, deterministic verification located 795 quotations, or 95.2 percent, either verbatim or following text normalisation. An additional 28 quotations, representing 3.4 percent, could only be located after a discontinuous model proposal was reduced down to a single textual segment. Twelve proposed quotations, accounting for 1.4 percent of the sample, remained completely unlocated within the canonical interview transcript and were consequently excluded from all citable research outputs.
To assess whether truncated quotations preserved analytical integrity, the author conducted a blinded reading of 26 available repaired quotations alongside 16 exact control excerpts. The analysis revealed that the retained segment fully supported the qualitative interpretation in 16 of the 26 repaired quotations, compared to 14 of the 16 exact control excerpts. Because some repaired quotations diverged in evidentiary strength, the study advises that researchers carry out direct manual confirmation before citing repaired quotations in published findings.
The automated routing process transferred 164 of the 206 interview excerpts into a review queue for human inspection. The study emphasizes that its automated metrics evaluate quotation location and record generation rather than validating the accuracy of qualitative syntheses, interpretations, or contextual attributions. The authors conclude that while the framework confirms that audit trails can be successfully generated at corpus scale, further research must evaluate whether these logs enhance external scrutiny and how warning thresholds should be calibrated.
Subject of Research: Audit frameworks for qualitative interpretation in AI-assisted health research
Article Title: From plausible to contestable: an audit framework for qualitative interpretation in AI-assisted health research
Article References: Naiditch, N. (2026). From plausible to contestable: an audit framework for qualitative interpretation in AI-assisted health research. BMC Medical Research Methodology. https://doi.org/10.1186/s12874-026-03084-9
Image Credits: AI Generated
DOI: 10.1186/s12874-026-03084-9
Keywords: qualitative health research, large language models, audit framework, quotation verification, territorial health inequalities, interview transcripts, research methodology, open-weight models, transcript confidentiality, excerpt-level interpretation, deterministic verification, supervision records, automated review queue, qualitative data analysis, research reproducibility, model accountability
News Source: Ophelia Keating. (November 16, 2026). Audit Framework Verifies AI Quotations in Qualitative Health Research Transcripts. Scienmag.



