Source
urlhttps://doi.org/10.1177/10778004251412871
rawraw/friese-2026-from-coding-to-conversation-a-new-methodological-framework-for-ai-assisted-qualitative-analysis.pdf

TL;DR: Friese makes the most radical structural argument in the corpus: traditional qualitative coding may be obsolete. CAAI (Conversational Analysis to the Power of AI) replaces segmentation-and-labeling with structured dialogic interpretation between researcher and LLM, grounded in hermeneutic epistemology and abductive reasoning. The framework is demanding but coherent, and its critique of “classification proxies” masquerading as qualitative analysis is the sharpest in the literature.

Problem

The dominant pattern in AI-assisted qualitative research is what Friese calls classification proxies: researchers use LLMs to assign predefined labels to text segments faster than humans can. This is often described as “AI-assisted coding,” but it preserves the surface form of coding while discarding what makes qualitative coding valuable — iterative immersion, theoretical positioning, abductive insight, and interpretive depth.

The problem is structural, not a matter of bad implementation. Qualitative coding in practice involves 80–250 distinct codes across 10–35 interviews, developed through repeated engagement with full transcripts, memoing, and reconfiguration of categories. Most “AI coding” studies use small, pre-segmented datasets, a handful of predefined categories, and chunk-level processing that destroys the continuity essential for interpretive leaps.

Friese identifies a disciplinary split: researchers from computer science and HCI treat qualitative data as text to be segmented and labeled; researchers from sociology, education, and health tend toward interpretive, dialogic engagement. Both groups claim to be doing qualitative analysis. Only the latter actually is.

Beyond the classification proxy critique, Friese raises a structural question: why are we trying to automate coding at all? Coding was invented as a systematic way to work with large text corpora before LLMs existed. Now that structured dialogue with an LLM can surface patterns across a corpus iteratively, the rationale for coding may have dissolved.

Approach

CAAI is a five-step framework:

1. Preparation. Data structuring, research question definition, epistemological positioning. The researcher must articulate their theoretical assumptions before dialogic analysis begins — not as a formality, but because the questions they ask will reflect and constitute those assumptions.

2. Exploration. Open-ended dialogue with the LLM about the corpus. The researcher asks pattern-surfacing questions: “What tensions run through these accounts?” “Which voices appear most and least?” This stage is deliberately inductive.

3. Deepening. Targeted dialogic exchanges using all three inferential modes — inductive (“What patterns do you see?”), deductive (“How does this relate to theory X?”), and abductive (“What would be a surprising interpretation here?”). The abductive move is CAAI’s distinctive epistemological contribution: using AI to surface the unexpected, not confirm the expected.

4. Synthesis. LLM-assisted integration of interpretations across the corpus. The researcher guides synthesis; the AI maintains coherence across a document set that exceeds any single conversation’s context window.

5. Validation and Reflexivity. Researcher review, documentation of interpretive decisions, and maintenance of an audit trail. Friese is explicit: this is not the same as intercoder reliability. Validity in CAAI is achieved through transparency and documented reasoning, not through agreement metrics.

The paper also distinguishes CAAI from existing CAQDAS integrations. MAXQDA’s AI Assist enhances traditional coding workflows while preserving methodological continuity. ATLAS.ti’s approach codes paragraph-by-paragraph without maintaining meaning across documents, generating hundreds of fragmented codes that require extensive cleanup. Neither is what CAAI proposes.

AI’s Role

AI is positioned as a co-analytic dialogue partner — contributing to the construction of interpretation through responsive engagement with the researcher’s questions, not as a coder producing outputs to be checked. The distinction is epistemological: in traditional AI-coding approaches, AI is an agent that produces results; in CAAI, AI is a medium through which the researcher develops understanding.

This is the most ambitious framing in the corpus. Where most papers position AI as a second coder, research assistant, or scale instrument, Friese proposes a genuinely different epistemological relationship — distributed cognition in which meaning emerges from the dialogue itself.

The paper is careful to note the risk: treating statistical vectorization (what LLMs actually do) as equivalent to conceptual coding is a category error. CAAI tries to use what LLMs are actually good at — responsive engagement with questions across large language spaces — rather than forcing them into a coding template they cannot fulfill.

Epistemological Stance

Hermeneutic, with explicit Big Q alignment. Friese situates CAAI within the tradition of interpretive inquiry: understanding emerges through iterative interpretation, not through classification. The hermeneutic circle — reading parts to understand the whole, understanding the whole to reread the parts — is the epistemological model that dialogic AI engagement actualizes.

CAAI integrates all three inferential modes: inductive openness, deductive structure, and abductive surprise. The abductive mode is philosophically significant: it positions the researcher as seeking the unexpected rather than confirming the expected, which is the opposite of the statistical tendency of LLMs toward dominant patterns (epistemic-flattening).

Rigor and Trustworthiness

Rigor in CAAI is achieved through dialogic transparency: the researcher’s questions, the AI’s responses, and the researcher’s interpretive moves are documented as an analytic record. This record functions as an audit trail analogous to memos in grounded theory.

Friese is one of the few authors who addresses the reproducibility problem directly. CAAI conversations are not reproducible in the same way quantitative analyses are — two researchers asking the same corpus the same questions will get different responses. This is not a bug but a feature: the goal is interpretive richness, not identical outputs. The quality standard is coherence and documented reasoning, not replication.

Limitations

CAAI requires a sophisticated researcher who can construct productive questions, recognize when AI responses are superficial, and document their interpretive reasoning. It is not accessible to researchers without substantial qualitative training — possibly more demanding than traditional coding.

The five-step framework is more sketch than protocol. Friese provides the epistemological rationale and the inferential vocabulary, but the practical guidance is thinner than, say, a grounded theory procedure manual. Researchers wanting to implement CAAI will need to develop their own operational procedures.

The paper does not engage empirically with how CAAI performs on specific datasets. The argument for CAAI is theoretical and methodological, not evidential. Whether it produces better analyses than coding-based approaches remains an open empirical question.

Connections

What links here