Source
TL;DRThe tendency of LLMs to reproduce dominant, statistically probable narratives — flattening the interpretive diversity that qualitative research depends on.

What it means

LLMs are trained to generate the most probable continuation of a text sequence. This structural feature means they are systematically biased toward dominant patterns in their training data. When used for qualitative analysis, this creates a specific risk: the AI will identify and foreground themes that are common, well-represented, and culturally dominant — while suppressing marginal, context-specific, or counter-hegemonic meanings.

(brailas-ai-qualitative-research-2025) calls this “epistemic flattening” — the AI doesn’t just fail to find certain meanings, it actively makes them less visible by organizing the data around what is statistically most likely.

Why this matters more than reliability

The usual metrics for evaluating AI-assisted qualitative research — intercoder-agreement, precision, Jaccard index — measure consistency and accuracy against a reference standard. They cannot detect epistemic flattening because:

  1. The reference standard itself may reflect dominant perspectives
  2. High agreement between AI and human coders can still systematically miss minority voices
  3. The problem is not that the AI is wrong — it’s that it’s limited to what’s already known

Qualitative research’s distinctive value is discovering unexpected patterns, surfacing contradictions, and representing perspectives that aren’t dominant. Epistemic flattening directly threatens this.

The alternative

(brailas-ai-qualitative-research-2025) proposes using AI abductively — asking it to surface contradictions, silences, and departures from expected patterns rather than confirming what is probable. (anis-french-ai-qualitative-research-2023) makes a related point via the “explicatory” argument: AI failures (algorithmic cases that don’t fit the coding scheme) are analytically valuable precisely because they flag the unusual.

(carlsen-ralund-computational-grounded-theory-2022)'s CALM framework addresses this at the methodological level: human immersion in the data — not model-selected paradigmatic cases — is what qualifies a researcher to interpret meaning within a specific community.

See also