cs.SEAug 28, 2026

Antipatterns in AI-assisted Qualitative Data Analysis: A Catalog of Temptations and Pitfalls for Software Engineering Researchers

Authors: Rashina Hoda, Carolyn Seaman, Victoria Gomes, Rodrigo Spinola

Organizations: Monash University, Melbourne, Australia · University of Maryland Baltimore County, Baltimore, USA · Virginia Commonwealth University, Richmond, USA

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

AI-assisted qualitative data analysis (QDA) offers unprecedented opportunities to streamline software engineering (SE) research, yet uncritical use risks compromising analytical rigor and flooding the field with accelerated production of low-quality research. While tactical best practices will naturally evolve over time, SE researchers currently lack strategic guidance to identify and mitigate methodological risks when attempting AI-assisted QDA. Based on our decades of qualitative SE research expertise and experience combined with an understanding of the emerging landscape of AI-assisted QDA, this paper presents a catalog of antipatterns in AI-assisted QDA - a set of assumptions and practices that initially appear advantageous but ultimately undermine analytical rigor and validity. The antipatterns are grouped into three categories reflecting escalating impact: Dangerous Drivers, Operational Missteps, and Analytical Failures. As more SE researchers attempt AI-assisted QDA, these antipatterns will help them identify and avoid common temptations and pitfalls, while reviewers can be equipped with the vocabulary and criteria to call out problematic and failed practice. Ultimately, this catalog of antipatterns can serve as a stepping stone in our responsible methodological evolution toward principled and meaningful human-AI collaboration in qualitative research.

Figures & tables

      
 AntipatternExampleWhy TemptingAssociated RisksPossible Mitigation Strategies
DANGEROUS DRIVERSBecause We Can : Using AI-assisted QDA because AI is able to process qualitative data and is readily available to human researchers, without due consideration of fit, relevance, risks, and limitations.A paper reports the use of an AI tool for open coding, describing the tool used, the prompts, the interaction, and the results, without giving a justification for why AI usage was considered and how the AI tool was selected.SE researchers are technologists at heart, and we are hard wired to explore the limits of new technology. The coolness of what AI can do, along with apparent gains in speed and scale seem sufficient.Loss of privacy, human expertise, learning, reflexivity and environmental and planetary sustainability costs , along with threats to research validity and missed opportunities for deeper insight, make it imperative to weigh AI’s apparent benefits against its real but less visible costs.Researchers should perform and share a thoughtful cost-benefit analysis to ensure AI use is meaningfully justified, rather than assuming speed and scale are sufficient rationales.
 The Productivity Spiral : A vicious cycle between increased scale and speed of AI-assisted QDA on the one hand, and increased community expectations of greater scale and speed from qualitative studies on the other.The Anthropic study “ What 81,000 people want from AI ” illustrates how large-scale AI-assisted studies can become a visible marker of research ambition and prestige.Researchers may be subject to pressure to use AI to produce more to keep up with peers and expectations from institutions, funding agencies, and publication venues.Prioritizing productivity over sustained analytical engagement can lead to loss of depth , nuance, diversity , emergence , reflexivity , human expertise , learning , traceability, and transparency , while normalizing unrealistic expectations.Researchers should use qualitative values (depth, richness, reflexivity) to select and justify their QDA approach, rather than be driven by quantitative markers such as scale and speed.
 Human as Gold Standard : Assuming human data analysis is an appropriate comparison or benchmark for AI-assisted QDA without considering the expertise and experience of the human analyst(s).A well-designed empirical comparison finds an AI and a human coder assign similar codes to 90% of the data and concludes that AI coding is acceptable without considering the human analyst’s expertise, experience, and QDA quality.Manual coding is time-consuming and can be tedious. Getting what seem to be the same results using AI can be seen to justify AI use, saving substantial time.Benchmarking against human analyst(s) may be insufficient and misses an opportunity to do better, leading to bias encoding , loss of depth , nuance, diversity , and reflexivity .Other avenues for evaluating AI-generated QDA results should be explored, based on known quality dimensions for qualitative research.
OPERATIONAL MISSTEPSQDA as Categorization : Using AI to group data into ”buckets” based on internal quantification (e.g., most popular themes), training data, and existing knowledge (instead of staying true to the research data). Limiting QDA to categorization.Coarse-grained, quantitatively-led categorization missing qualitative richness, e.g., ”Types of test debt identified: missing test cases (mentioned by 60% participants) and broken test cases (10%)”.Categorization is a low hanging fruit which LLMs seem to be good at (e.g., content analysis). In low risk, non-research contexts (e.g., market analysis), it may be all that is needed.Reducing QDA to surface-level categorization can drive loss of depth , nuance and diversity ; forcing data into fixed buckets can reinforce bias encoding and cause loss of emergence and loss of creativity .Deductive QDA does not end at categorization; categories should be analyzed for richness and relationships. In inductive QDA, AI can assist with individual and incremental steps (e.g., open coding and constant comparison), rather than categorizing across the dataset.
 Epistemological Slide : A shift away from the constructivist foundations of qualitative research toward positivism.AI assumed by some to reflect the ‘objective truth’ and thus being employed as a ‘superior’ analyst to subjective humans. Presentation of findings marked by quantification of qualitative data.Positioning AI as ‘objective’ and free from human intervention can be used to justify its use in QDA.Reducing the role of the human in constructing knowledge and over-reliance on AI can lead to loss of nuance, diversity, depth, emergence, reflexivity and creativity and risks bias encoding .Reviewers should watch for and question overly positivist research approaches (quant heavy, claiming to be objective, looking to verify ‘the truth’) when a constructivist approach is more meaningful.
\fnum@table : A catalog of antipatterns in AI-assisted qualitative data analysis.
      
 AntipatternExampleWhy TemptingAssociated RisksPossible Mitigation Strategies
 AI Output as Insight : AI-produced QDA findings can sound insightful while remaining shallow, leading researchers and readers to mistake superficial output for meaningful analysisA researcher uses AI to generate themes from developer interview data about cloud vs on-premises servers. The polished output sounds insightful, but the findings are generic and weakly grounded in data.AI tools produce fluent, well-structured outputs that resemble legitimate qualitative findings, giving a strong illusion of depth, rigor, and expertise.Mistaking persuasive AI output for genuine insight creates loss of depth and loss of nuance and diversity ; accepting superficial findings without scrutiny also drives loss of reflexivity and loss of traceability and transparency .AI-produced results should always be critiqued and guided by human researchers, pushing for depth and nuanced understanding of the underlying data [ 13 ] .
ANALYTICAL FAILURESLow Human Engagement : Researchers lacking enough familiarity with the raw data to be able to spot mistakes, biases, and hallucinations when applying AI to QDA.A researcher uses AI to analyze interview data collected by others without reading or becoming familiar with much of it, resulting in hallucinated findings they fail to recognize due to limited engagement with the material.Reading data is time consuming, and often tedious. Fully reading all the data may seem to erase much of the productivity gains of using AI.Because AI can produce plausible but incorrect results, detecting errors requires close familiarity with the data. Limited engagement with raw data can therefore reinforce bias encoding and drive loss of depth , loss of reflexivity , and loss of human expertise and learning .Read the data.
 Fishing for Results : Exploiting AI’s ”eagerness to please” (AKA AI sycophancy) by steering it toward confirming assumptions, prior hypotheses, or external theories rather than staying true to the data.A researcher believes a proposition is true, but it is not supported by the AI-generated findings. When prompted to consider it, the LLM complies by generating a tenuous chain of ”evidence” from the data to support the suggested proposition.QDA techniques are rigorous, in part, to counter the human tendency to look for the evidence we want in the data (confirmation bias). It is tempting to cede responsibility for countering that tendency to AI when it appears to have found what we are looking for.We know that human bias is real, and so we have built methodological infrastructure to bound it (e.g., member checking, code reviews). Using AI to legitimize preferred conclusions can bypass these safeguards, reinforcing bias encoding and driving loss of emergence , loss of reflexivity , and loss of nuance and diversity .Prompt histories should be reported when AI is used for QDA [ 39 ] . Even more important, prompting techniques should be devised and taught that do not allow researchers to transfer their own biases to AI-based analysis. AI can be trained to be aware of the researcher’s assumptions list [ 4 ] .
 Automated QDA : Fully automated, end-to-end AI driven QDA with little to no human engagement with the overall process and analysis.A research team feeds a set of research questions and a large corpus of interview data into an AI-based QDA tool and asks it to generate a publishable paper on the analysis and results.This is taking the efficiency and scale of AI use to the max and exhibits the downsides related to all previous antipatterns.Fully automating QDA leaves AI decisions and biases unchecked, reinforcing bias encoding and causing loss of traceability and transparency ; removing human engagement also drives loss of reflexivity , loss of human expertise and learning , loss of depth , and loss of emergence .As a community, we need to develop good workflows that effectively involve AI as a partner in QDA, with frequent and meaningful human-AI collaboration, both for learning and for evaluation.
\fnum@table : A catalog of antipatterns in AI-assisted qualitative data analysis (Continued).
More in this topic

Human-in-the-Loop AI

Human-in-the-loop AI combines human feedback, oversight, and decision-making with machine learning and autonomous systems.

See all 290 papers

Explore similar work

CardsList
Apr 30, 2026

To Vibe Research or Not to Vibe Research? Generative AI in Qualitative Research

There has been intense debate among qualitative researchers about whether generative AI is suitable for qualitative research. In this paper, we summarize the broader ongoing discussion of generative AI in qualitative research and its implications for software engineering researchers. The qualitative research approach, small-q (positivist or post-positivist) or Big Q (non-positivist), is among the major criteria for determining whether generative AI can be used in qualitative research. In addition to research philosophy and research approach, skills, ethics, and personal preferences also play a role in researchers' decisions about whether to use AI in qualitative research.

Katja Karhu, Kari Smolander, Jussi Kasurinen

LUT University, Lahti, Finland · LUT University, Lappeenranta, Finland

AI-Assisted Scientific ResearchHuman-AI Collaboration
AskPDF
Apr 21, 2026

Co-Refine: AI-Powered Tool Supporting Qualitative Analysis

Qualitative coding relies on a researcher's application of codes to textual data. As coding proceeds across large datasets, interpretations of codes often shift (temporal drift), reducing the credibility of the analysis. Existing Computer-Assisted Qualitative Data Analysis (CAQDAS) tools provide support for data management but offer no workflow for real-time detection of these drifts. We present Co-Refine, an AI-augmented qualitative coding platform that delivers continuous, grounded feedback on coding consistency without disrupting the researcher's workflow. The system employs a three-stage audit pipeline: Stage 1 computes deterministic embedding-based metrics for mathematical consistency; Stage 2 grounds LLM verdicts within ±0.15\pm0.15 of the deterministic scores; and Stage 3 produces code definitions from previous patterns to create a deepening feedback loop. Co-Refine demonstrates that deterministic scoring can effectively constrain LLM outputs to produce reliable, real-time audit signals for qualitative analysis.

Athikash Jeyaganthan, Kai Xu, Franziska Becker +1

University of Nottingham Nottingham, United Kingdom · University of Stuttgart Stuttgart, Germany

Human-in-the-Loop Annotation
AskPDF
Apr 4, 2026

Affording Process Auditability with QualAnalyzer: An Atomistic LLM Analysis Tool for Qualitative Research

Large language models are increasingly used for qualitative data analysis, but many workflows obscure how analytic conclusions are produced. We present QualAnalyzer, an open-source Chrome extension for Google Workspace that supports atomistic LLM analysis by processing each data segment independently and preserving the prompt, input, and output for every unit. Through two case studies -- holistic essay scoring and deductive thematic coding of interview transcripts -- we show that this approach creates a legible audit trail and helps researchers investigate systematic differences between LLM and human judgments. We argue that process auditability is essential for making LLM-assisted qualitative research more transparent and methodologically robust.

Max Hao Lu, Ryan Ellegood, Rony Rodriguez-Ramirez +1

Harvard University / United States

LLM-Assisted AnnotationLLM Auditing
AskPDF