Acceleration and Recursion in AI-Mediated Qualitative Research: Discussing Adapted Criteria for Qualitative Inquiry in the Techno-Epistemic Age

Extended AI-Accountability Statement

The authors verified and are fully responsible for all content. No AI-system is listed as an author. Generative AI-tools (ChatGPT 5.0 Pro, Gemini 2.5 Pro) were used during four steps:

Step 1 (Sep 2025): The initial idea was based on our previous work (Meyer & Miggelbrink, 2013) at the crossroads of qualitative fieldwork and Hacking’s work (1995). We intersected that with reflection on the inclusion of (semi-)automated processes of data analysis in a discourse-theoretical project on populism (Nguyen & Meyer, 2026). After researching literature on epistemic risks, we triangulated this database with an LLM-assisted (ChatGPT 5.0 Pro) search for further literature, (I) expanding the literature base. Based on this expanded corpus, we refined our summary of epistemic risks, and proceeded with the conceptualization based on Hacking’s concepts of recursion and acceleration.

Step 2 (Oct 2025): We derived initially five, later four provisional, reportable evaluative criteria. We then used ChatGPT 5.0 Pro and Gemini 2.5 Pro to (II) review these criteria based on their applicability and feasibility. Based on these reviews, we revised the criteria. In this step, all literature references were re-validated.

Step 3 (Oct 2025): We created the full manuscript based on the previous work, including Pro- and Epilogue, and diving into Haraway (1985). After the full manuscript was devised, we prompted ChatGPT 5.0 Pro and Gemini 2.5 Pro to (III) act as scientific reviewers and thoroughly, critically, and constructively generate extensive reviews, in response to which we then revised the manuscript.

Step 4 (Oct 2025): ChatGPT 5.0 Pro was used to cross-check reference usage in the list of references and in the manuscript, as well as regarding the necessary formatting, (IV) serving as a formatting assistant. This was checked manually afterwards.

Prologue

Qualitative inquiry has always been an endeavour of crossings – between voices and codes, emotions and notebooks, lives and concepts. But the crossings have changed. Circuits hum inside our methods now. They creep inside through their pervasiveness in our daily lives. Once sound and ink were patiently waiting for the interpretative labour we so ambitiously executed from the black boxes of our human subjectivity. This has become silicon and immediacy.

Who dares to doubt our capabilities of transference? The cyborg age no longer sits at a distance; it threads through our prompts, our internet searches, our private conversations. This information accelerates and loops back into what we hope to remain our real lives. Outputs re-enter as inputs; drafts become datasets; templates harden into habits.

Change at machine speed? Whilst some, nevertheless, attempt to retreat from it, we cyborg apologets have to reflect consequences and ponder new progressive futures from what could easily mirror our most regressive nightmares. Rather than retreat from this condition, we ask what AI changes in the classificatory loops through which we produce and evaluate knowledge.

“By the late twentieth century, our time, a mythic time, we are all chimeras, theorized and fabricated hybrids of machine and organism; in short, we are cyborgs.”

1. Introduction

Since the 1990s, data-, software- and hardware-driven innovations have been re-shaping interpretative qualitative research (QR) (Flick, 2019; Williams, 2024) – approaches whose goal is to “reconstruct fixed structures and meanings that actors follow and that can, at best, be interpreted by actors” (Knoblauch, 2008, p. 210) and approaches that assume a “fundamental scope for subjective interpretations” that “is not absorbed into meanings and structures” (Knoblauch, 2008, p. 210). Early on, these impulses came from utilizing software to assist in lexicometric and qualitative-analytical research processes such as content analysis. In the 2020s, new techno-epistemic assemblages have emerged and allowed new stimuli: Big Qualitative Data (BQD) (meaning: large, often textual corpora processed, e.g., with topic modeling, Computer-Assisted Qualitative Data Analysis Software (CAQDAS), or LLM assistance) (Chandrasekar et al., 2024), Synthetic Data in model training and evaluation (SDT) (the growing presence of synthetic or model-amplified text in training/evaluation to further expand its capabilities) (Shumailov et al., 2024), and AI-mediated Qualitative Analysis (AIQA/AIaQA) (making use of LLMs, embeddings, and auto-coding, integrated into software analysis pipelines ¹ ) (Mayring, 2025). These stimuli expand what and how much can be analysed and how quickly; they also alter how qualitative classifications are formed and circulated (e.g., Foraker et al., 2025; Loni et al., 2025; Pink et al., 2025).

Undoubtedly, AI-tools can be a force for increased productivity in QR, interdisciplinary innovation, and for a more widely and readily available circulation of knowledge. Furthermore, they can allow us to enhance existing analytical approaches beyond much current capabilities. However, much enthusiasm around AI in QR implicitly imports quantitative ideals into interpretative practice, such as representativeness/generalization, reliability/reproducibility, and objectivity – each with discipline-specific limits in qualitative inquiry (Braun & Clarke, 2025; Davison et al., 2024; Nguyen & Welch, 2025). AI-mediated methods fuel hopes for an expansion of interpretative methods in a way that gives them the same (assumed) robustness as is assumed for quantitative methods. From this, one could deduce that qualitative research can be enabled to meet the same quality criteria as are customary for quantitative methods.

• Representativeness/generalization: Enlarged corpora can broaden coverage and rhetorical reach, yet text corpora rarely rest on stable sampling frames. In discourse/content analysis, statistical representativeness is doubted as a suitable measure. Instead, the relevant standard is analytic generalization/transferability and transparent corpus construction (Bednarek, 2024; Keller, 2011; Stoltz & Taylor, 2024). For instance, in discourse/content analysis, “representativeness” is often meant to be semantic, meaning a proper coverage of positions/meanings, rather than a statistical measure. Corpus size widens coverage, but inclusion/exclusion remains theory-led and must be documented (e.g., via corpus boundaries, search strategy, timestamps) (Bendel Larcher, 2023; Gür-Şeker, 2014). ²

Given AI’s increasing push into QR, and building on Hacking’s (1995) looping effects of human kinds, we argue that AI does not merely speed up analysis (Chubb et al., 2022); it reconfigures the mechanics of qualitative classification. This results in various effects, of which we specifically want to highlight two: As classifications circulate through tools and publics, they feed back faster into talk, practice, and thus data – resulting in acceleration (Rosa, 2013). Evidence from human-AI interaction shows feedback loops that can amplify or skew judgments (Brown et al., 2022; Glickman & Sharot, 2025; Pagan et al., 2023; Perdomo et al., 2020). This concerns a second shift – recursion: when pipelines (partly) re-use their own outputs or synthetic surrogates as inputs (e.g., auto-coding templates, prompt libraries, synthetic corpora), loops iterate without renewed contact with lived worlds, closing the evidential circuit and risking evidence erosion, synthetic drift of meaning, or even model collapse (Foraker et al., 2025; Loni et al., 2025; Shumailov et al., 2024).

Hacking’s original looping effect concerns the circulation of classifications through institutions and public life, their uptake, resistance, or inhabitation by the people classified, and the subsequent revision of those classifications through renewed encounters with practice. We extend this account in two analytically distinct directions. Acceleration names the compression of the interval between classification, uptake, and reclassification. Recursion names the re-entry of earlier outputs into later analytical stages without renewed empirical contact. Acceleration changes the tempo of the loop; recursion changes its evidential architecture.

This article develops a conceptual framework and proposes four provisional, reportable evaluative criteria for AI-mediated qualitative research. Specifically, it speaks to researchers using these tools in processing qualitative data who seek methodological reflections as well as methodologists reflecting, re-negotiating and reframing conditions and criteria for evaluating qualitative research methods. It extends existing debates on reflexivity and transparency by identifying acceleration and recursion as specific techno-epistemic dynamics. These criteria are not intended as fixed metrics or a closed checklist, but as points of orientation that make AI-mediated interpretative work visible, reviewable, and contestable. In this sense, the paper serves both as a theoretical extension of looping effects and as a methodological guideline for documenting AI-mediation in interpretative research.

Our focus here is interpretative QR on text- and transcript-based materials that we regard typical resources for QR single-method as well as mixed-methods projects (interviews, documents, online sources). We do not treat multimodal/sensory ethnography or visual analytics in depth. Our claims concern analytical practices. This framework is written primarily from and for interpretative, constructivist, and critical traditions of qualitative inquiry, where meaning, situatedness, and reflexive accountability are central, and where quality is not reducible to statistical representativeness or procedural standardization. Accordingly, the risks we foreground (e.g., over-smoothing, authenticity drift, performativity amplification, and externalized reflexivity) are framed as threats to interpretative validity and to the politics of classification. We nevertheless expect the framework to travel and change into more post-positivist qualitative or mixed-methods contexts: here, the criteria can be used as transparency and governance add-ons. However, we do not claim that these criteria replace validity- or reliability-related procedures typically emphasised in positivist paradigms; rather, they complement them by making AI-mediated decisions, feedback loops, and normative assumptions visible and contestable.

As a theory-led response: (i) we extend Hacking (1995) by defining acceleration (compressed loop time) and recursion (output re-entry) for AI-mediated QR (section 2); (ii) we provide a concise overview of the key technical developments of AI and their impacts on QR (section 3); (iii) we diagnose four groups of resulting epistemic risks (section 4); (iv) we propose provisional, reportable evaluative criteria for the AI use and concomitant risks (section 5).

“A cyborg is a cybernetic organism, a hybrid of machine and organism, a creature of social reality as well as a creature of fiction.”

2. Conceptual Background: Qualitative Epistemology and Hacking’s Legacy

Hacking’s account of so-called human and interactive kinds and their looping effects remains an influential framework for understanding how categories and the people classified by them coproduce one another. In the classic formulation (Hacking, 1995, 2006), expert classifications of people (often value-laden) circulate through institutions, media, and everyday practice. People recognize, resist, or inhabit these ascriptions, and consequently, their behaviour and self-descriptions change. Classifications develop a social life of their own by being reproduced, confirmed, criticized, and contested. The category is then re-made in light of those changes by scientific or diagnostic inquiry. Hence, human kinds are “moving targets” because the very act of classification alters what is being classified (Hacking, 1999/2000, 2006).

Qualitative inquiry has long addressed the performative circulation of classifications by foregrounding situatedness (Haraway, 1988), thick description (Geertz, 1973), abduction (Timmermans & Tavory, 2012), negative-/deviant-case analysis via constant comparison (Becker, 1998; Glaser & Strauss, 1967), reflexivity (Macbeth, 2001), or dynamics such as retro-action and counter-performativity (Hostniker & Meyer, 2024). QR has therefore been a significant source and target of criticism of reified and naturalized classifications as well as of classificatory infrastructures.

Subsequent authors have operationalized Hacking’s work. Contributions on classificatory infrastructures show how standards enact worlds (Bowker & Star, 2000). Works on quantification demonstrate that turning qualities into metrics reshapes behaviour and organizations (Espeland & Sauder, 2007; Espeland & Stevens, 1998; Mennicken & Espeland, 2019). Studies on peripheralisation illustrated how spatial ascriptions may influence behaviour in shrinking regions (Meyer & Miggelbrink, 2013). Economic studies interrogated markets and models as engines rather than cameras (Burrell & Fourcade, 2021; MacKenzie, 2006). Clinical studies reveal epistemological injustices developing in interactions between patients and clinicians (Gauld et al., 2025), and historiographies of objectivity locate neutrality claims in specific technologies, disciplines, and methods (Daston & Galison, 2007). Taken together, labels, rankings and algorithms do not simply discover social kinds; they re- and co-produce them (Fourcade & Healy, 2017a; Fourcade & Healy, 2017b) – a dynamic reflected and willingly utilized in modern marketing (Paulson & O’Guinn, 2018), self-optimization or plastic surgery (Elliott, 2019).

What is new in the 2020s is not looping per se but that AI modifies the temporal and evidential conditions of the loop. First, acceleration: AI-mediated coding, retrieval and summarization shorten the interval between classification, uptake, and reclassification. We term this acceleration a compression of loop time, consonant with macro-accounts of acceleration of social relations (Rosa, 2013). Second, recursion: synthetic data and pipeline re-use allow outputs to re-enter as inputs, so loops iterate without renewed contact with lived worlds and empirical verification. Recent evidence shows model behaviour drifting when trained on synthetic/recursively generated data (sometimes termed model collapse), potentially resulting in evidence erosion and drift of meaning in machine-mediated systems (Resnik et al., 2025; Shumailov et al., 2024; Taori & Hashimoto, 2023). A loop can be accelerated without being recursive, and recursive without major temporal compression; the two therefore should not be collapsed into one another. Both acceleration and recursion change fundamental mechanics of generating knowledge by altering reflection and empirical correction of knowledge.

In Hacking’s original account, looping still depends on renewed empirical encounters with people and practices. Our extension identifies two ways in which AI-mediated research modifies that structure. Acceleration compresses the time between classificatory acts and their uptake. Recursion, by contrast, allows classifications, summaries, prompts, labels, or synthetic surrogates to re-enter later stages of analysis as inputs. The novelty of recursion is therefore not simply greater speed, but a partial displacement of empirical correction by pipeline-internal circulation.

This has profound ramifications: In literature on machine learning, some models show performative prediction: they don’t just predict the world; their predictions change the world they then learn from (Brown et al., 2022; Perdomo et al., 2020). Pagan et al. (2023) offer a taxonomy of feedback loops (who/what feeds back, with what bias effects). Contemporary philosophy recently has revisited Hacking’s theses under these conditions (Tsou, 2025; Vesterinen, 2021). In short, looping may shift from slow, dialogical circuits anchored in empirical data to machine-assisted, high-frequency, partially self-referential circuits.

Hacking’s original account largely presumes that empirical correction happens through renewed encounters with people and practices (Hacking, 1995). Today, classification pipelines can self-validate, in the shape of prompt-library inheritances, fine-tuning on analyst-approved outputs, preferences for prior labels, and synthetic corpora that encode earlier manual classifications. The result is a lack of renewed empirical checks – a loop that tightens on its own outputs (Pagan et al., 2023; Shumailov et al., 2024). Similar dynamics appear across diverse disciplines, e.g., in applied domains where disclosing algorithmic scores induces self-fulfilling prophecies (Bauer & Gill, 2024), or when clinical practice co-produces diagnostic kinds (Gauld et al., 2025; Lindholm & Wickström, 2020).

Without explicit attention to acceleration, recursion, and updated standards of documentation of the research process, Hacking’s framework even under-describes the role of AI-mediated QR. We propose a formal extension by differentiating acceleration and recursion. Based on these insights, the following sections explore current AI-dynamics and their effects on QR.

3. Technological Vectors and the Techno-Epistemic Effects on QR

In the following, we will descriptively cover the spectrum of technological developments in the AI-sector and structure them in relation to their relevance for QR. As of Oct 2025, we have identified a total of 11 technological developments, referred to below as A to K. We have grouped these into three categories based on their existing or anticipated relevance to qualitative research processes. The first category comprises developments with immediate, high-prevalence influences (A to D). The second category comprises developments with near-term, field-dependent influences (E to H). The third category comprises developments with emerging influences (I to K). Table 1 gives an overview of these developments, current practices and relevance for QR.

OPEN IN VIEWER

Table 1.

AI-Influences Affecting Qualitative Research (Oct 2025)

Tier	Influence	Explanation	Current practices	Relevance for QR
1 A	AI-assisted qualitative analysis (AIQA/AIaQA)	Large language models and embeddings now support every step of qualitative analysis - from transcription and retrieval to first-pass coding, memoing, summarizing and drafting.	In day-to-day projects, teams use prompt libraries and workflow templates; retrieval often resurfaces earlier labels, which nudges analyses toward the same frames.	This is the most common AI-touchpoint; it shifts reflexive labour, introduces prompt/path dependence, and raises transparency expectations.
1 B	Provenance, licensing & traceability	Provenance means documenting where data, labels and tools come from and how each analytical step was produced, including synthetic exposure.	Journals and funders increasingly ask for licenses, attribution and pipeline records; many teams add provenance capture to CAQDAS/LLM workflows.	Credibility now depends on such records; they underpin Provenance & Recursivity Accounting/Reflexive AI-Audit and let reviewers see how AI shaped interpretations.
1 C	Retrieval-augmented generation	Retrieval-augmented generation combines search with generation to find, cluster and summarize materials from the (changing) web.	Researchers rely on live indices and proprietary corpora when scoping literatures or coding document sets, then paste summaries into memos.	Because the web is volatile, RAG changes what counts as evidence; time-stamping and freeze-framing are needed for saturation and reproducibility.
1 D	Human-in-the-loop alignment & instruction-tuning	Alignment pipelines train models with human preference ratings and rubrics (RLHF) and, increasingly, AI-mediated feedback (RLAIF); instruction-tuning mixes public prompts, curated tasks and crowd annotations.	These trainer decisions and rubrics are embedded in general-purpose assistants that later help with coding and summarizing.	Trainer subjectivities can appear as “neutral” suggestions; qualitative studies should disclose alignment regimes and triangulate outputs.
2 E	Big Qualitative Data (BQD)	Very large text corpora analysed with CAQDAS+LLM features enable batch coding, semantic search and fast spread of “themes.”	Codebooks and auto-rules are reused across datasets; themes move quickly between related projects.	Scale aids coverage but strains negative-case work; treat genre/platform as variables and re-ground codes across sites.
2 F	Synthetic Data in model training and evaluation (SDT)	Synthetic text now appears upstream (in model training/evaluation) and downstream (as augmented inputs in studies).	Projects combine real and synthetic materials; prior model outputs may be re-used as “data.”	Alters the evidential mix and demands explicit provenance and authenticity checks.
2 G	Multimodal AI (audio/video/image)	Audio, video and image tools (diarisation, translation, gesture extraction) are entering qualitative and mixed-methods pipelines.	Teams generate automated transcripts, add speaker turns and timing, and sometimes code paralinguistic cues.	Widens what can be analysed but raises consent, fidelity and interpretability issues.
2 H	Low-resource/non-English LLMs	New models expand support for non-English and low-resource languages, though quality and bias vary.	Researchers run multilingual retrieval and coding and, where possible, cross-check with native speakers.	Improves inclusion and comparability but requires culturally aware prompting and validation.
3 I	Agentic/autonomous research assistants	Task-chaining agents coordinate retrieval, coding, memoing and drafting with limited supervision.	Analysts launch multi-step runs that produce labels and summaries; authorship and accountability can blur.	Saves time on routine work yet complicates attribution and auditability.
3 J	Utilizing on-device/edge LLMs	Models run locally on secure devices, reducing reliance on the cloud.	Fieldworkers process sensitive materials offline and sync later.	Supports privacy-preserving work but can limit reproducibility and requires clear consent/security language.
3 K	Affective/voice & adversarial/poisoning risks	Some tools claim to infer emotion or sentiment; open pipelines face data-feedback loops and poisoning risks.	Sentiment/affect add-ons are bolted on; datasets from the open web may be contaminated or drift over time.	These risks threaten corpus integrity and category stability; provenance checks are needed.

4. Epistemic Risks of Integrating AI Into QR

We address four risks that we consider most important for interpretative validity in qualitative inquiry. These risks are specific to how AI intersects with everyday qualitative practice, not generic AI-failure modes. For each risk, we show how it arises from the vectors in Section 3 and why it matters for interpretation. Again, we summarise our observations in Table 2.

OPEN IN VIEWER

Table 2.

Overview of Epistemic Risks of Integrating AI Into QR

​	Risk	Explanation	Evidential status
1	Over-smoothing and novelty loss	• Risk of preserving existing categories and overlooking new developments	• documented tendency in adjacent literature
		• Risk of generating an outcome by predicting the most statistically probable sequence of data based on the prompt, input, and contextual information	• plausible QR consequence
		• Risk of overlooking subjective or deviant positions
2	Self-validation and authenticity drift (toward fossilisation)	• Risk of (falsely) stabilising existing knowledge through outputs that re-enter the pipeline as inputs	• documented recursive degradation/synthetic drift;
		• Ethical risks by inadequate use of synthetic data	• potential QR consequence
		• Risk of losing context sensitivity
3	Performativity amplification, lexical anticipation, and outlier suppression	• Risk of further privileging hegemonic positions through research based on statistical probabilities, while opposing positions no longer appear	• documented feedback-loop effects
			• plausible QR consequence
4	Externalized reflexivity, de-skilling, and governance-by-pipeline	• Risk of delegating or externalising not only key decisions in the data evaluation process, but also in reflecting on the quality of these decisions.	• primarily methodological/normative concern
			• supported by emerging practice-related literature

The risks discussed below do not all have the same evidential status. Some are directly documented in adjacent empirical literatures, such as over-smoothing in computational text analysis, feedback effects in human-AI interaction, or degradation under recursive synthetic training conditions. Others are plausible methodological risks for AI-mediated qualitative research inferred from those findings and from common workflow features. A third subset is forward-looking and normative: concerns not yet systematically demonstrated in qualitative research, but important to articulate because they bear on emerging standards of interpretative accountability.

5. Provisional, Reportable Evaluative Criteria for QR in the Age of AI

The following reportable criteria are proposed as provisional evaluative anchors for AI-mediated qualitative research. They are not intended as universal thresholds, nor as a closed checklist that could be applied uniformly across qualitative traditions. Rather, their function is to make AI-mediated interpretative work reportable, inspectable, and contestable: they ask whether claims remain empirically refreshed, whether provenance and recursive output re-entry are visible, whether heterogeneity and negative cases remain analytically available, and whether the conditions of AI-mediation are reflexively disclosed. In this sense, the criteria structure justification and critique without foreclosing debate about their future refinement.

The criteria extend classic commitments to credibility/trustworthiness (Lincoln & Guba, 1985) or reflexivity/transparency (Macbeth, 2001), adapting them to AI-mediated workflows. Because this article is a conceptual intervention, these criteria are deliberately proposed in provisional form. Their purpose is not to close discussion, but to open a more explicit debate about how AI-mediated qualitative work should be described and evaluated as practices, tools, and standards evolve.

Precisely, Tracy (2010, p. 840) identifies eight QR-criteria related to the topic, rigor, sincerity, credibility, resonance (with an audience) and contribution (to a state-of-the-art), research ethics, and meaningful coherence. Interestingly, whilst we consider our suggestions consequential, they also seem to stray from qualitative criteria due to the need to cover the hybrid regime of AI-mediated QR.

By proposing the following criteria, we aim, firstly, at introducing potential additional criteria for those research designs employing AI-mediated practices. Yet, secondly, we aim at triggering a deepened reflection with the potential ramifications of AI-mediated qualitative research. Given the rapid advancements in AI-technologies and their implementation into research designs, our criteria can only serve as a temporary snapshot of such reflections to deliberately allow for an ongoing, critical, and restless debate in the face of current and future technological advancements. With further progress, these criteria need to be doubted, substituted, and/or refined to match upcoming innovations and developments.

6. Discussion and Conclusion

Qualitative interpretation depends on situated judgment, responsiveness to difference, and the possibility of being corrected by the material. In contrast to Jowsey et al. (2025), we do not generally doubt the potential of AI for QR. Yet, AI-mediated QR must confront a structural tension: current models generate statistically probable continuations, whereas interpretative inquiry seeks to reconstruct meaning, contradiction, and context in ways that cannot be reduced to probability alone. AI may therefore assist interpretative work, but it does not remove the researcher’s responsibility for reflexive judgment.

Our claim is therefore not that AI creates looping from scratch, but that it intensifies one dimension of looping (tempo) and transforms another (evidential re-entry). Long before LLMs, CAQDAS (e.g., Atlas.ti, NVivo, MAXQDA) already sped up and supported coding, retrieval, and cross-project reuse. Likewise, recursive dynamics were discussed in non-AI settings under performativity, retroaction, and counter-performativity (see Hostniker & Meyer, 2024). What is distinctive today is (i) the magnitude and speed of these dynamics in AI-mediated pipelines, and (ii) their partial decoupling from renewed field encounters – through output re-entry via prompt libraries, fine-tunes, and synthetic data.

Our critique rests on a normative judgment, not an empirical claim: researchers often treat some degree of acceleration and recursion as acceptable when it is human-driven, and problematic when it is AI-mediated. The boundary is fragile and contingent, and overlaps with debates about how much human control or “teaming” with AI is desirable or possible (Tsamados et al., 2025). Noting this boundary surfaces a hidden premise: we want humans to remain agents of social change, or at least accountable stewards of the classifications we circulate, as “[o]ur own mythology consists in imagining ourselves as radically different, even before searching out small differences and small divides.” (Latour, 1993, p. 116)

As Latour argues, modern social change is better understood by attending to mediations (the hybrids of nature and society) rather than presuming a purified separation between them (Latour, 1993). In this perspective, actants are any human or nonhuman entities that participate in networks of association; agency is not possessed in isolation but emerges from relations among these actants (Latour, 2005). This shift from purified domains to mediating networks helps to explain the nature of change – and, to some extent, the core conflict of paradigms at hand. As a matter of fact, most foundation models are produced and governed by large vendors; thus, vendor defaults can steer qualitative methods and topics. Keeping accountability with the researcher requires disclosure and, where possible, auditable open pipelines.

Although we sympathise with Jowsey et al. (2025), change is already underway, across fields and within qualitative research, as research teams adopt AI-tools and practices. Consequently, QR needs to adapt. This descriptive claim simply notes that the widespread circulation of large, industry-built foundation models is creating a new state of the art, one that will shape what counts as evidence, how we analyse it, and the conclusions we draw.

Since widely accessible large language models entered practice only in late 2022, the field is changing faster than its standards. The abundance of new capabilities risks rendering established assumptions, procedures, and quality heuristics obsolete. We therefore call for explicit normative commitments for knowledge production using AI, detailing how evidence is collected, traced, audited, interpreted, and governed – both in general and for qualitative research in particular in this new age of techno-epistemology.

Epilogue

Should Qualitative Inquiry dream of Eden? Is there any garden to return to outside the circuitry of our tools and the endless cycles of progress, mediation, and regress?

Our criteria are not commandments but commitments that keep methods porous to the lives they read. We do not need to quietly inherit the hierarchies of technological corporatism. The criteria we offer will not open the loop; they render it visible. They stage conditions under which error can be seen, contested, and repaired. They are cyborg practices for a cyborg craft: neither anti-machine nor credulous, but accountable.

And accountability, in the end, is less a rule than a posture, and the readiness to be addressed by what we study and by those who will live with our classifications.