GATOS qualitative analysis workflow

A research workflow for using open-source generative AI and machine learning to support inductive qualitative codebook development.

Built on our work in AI for qualitative research: Using Large Language Models and Generative AI to Scale Qualitative Data Analysis.

Research prototype, in development, packaging the published method for release.

The problem

Large qualitative corpora are difficult to analyze rigorously without hiding methodological decisions or overwhelming human coding capacity.

The workflow connects model-assisted theme generation, researcher validation, and documented codebook decisions so AI support remains inspectable rather than replacing qualitative judgment.

For qualitative researchers, engineering education researchers, research methods collaborators.

Research basis

GATOS stands for Generative AI-enabled Theme Organization and Structuring. It was introduced and validated in Humanities and Social Sciences Communications, then applied at scale in the Journal of Engineering Education. The workflow is a published method first and a piece of software second, which is why the method is citable today while the packaged tooling is not yet public.

What it does today

Summarizes each unit of raw text against the researcher’s stated research question, producing atomic summary points analogous to observation memos.
Embeds those summary points, reduces their dimensionality, and clusters them so semantically related observations group together.
Generates candidate codes from each cluster and organizes those codes into higher-level themes.
Keeps every intermediate step (summaries, clusters, candidate codes) open to inspection, so the path from raw text to theme can be audited rather than taken on trust.
Runs on open-source models, so a corpus never has to leave the researcher’s own environment.

What it does not do

There is no installable package or hosted service yet. Using the method today means implementing it from the papers.
It does not ingest data, manage projects, or produce analysis reports for you.
It has no interface. The published work describes a workflow, not an application.

What it does not show yet

Validation to date rests on three synthetic datasets whose themes were known in advance, plus one applied study. That is real evidence for recovering known structure, not a general validity claim across arbitrary corpora or domains.
It is decision support for qualitative researchers, not a replacement for interpretive judgment. The published method assumes human review at the points where it matters.
Agreement between the workflow and a human coder has not been established as a reliability statistic that would satisfy every methodological tradition.

Outcomes

Anchored in published work on open-source generative AI for thematic analysis.
Designed as a bridge from peer-reviewed methods to reusable research infrastructure.

Limitations and responsible use

Researchers remain responsible for the ethics and permissions covering their own corpus; the workflow makes analysis faster, not consent broader.
Model temperature is set to zero for determinism, but generative models remain probabilistic and outputs should be treated as candidates for review.

Evidence

Method validated against known themes: Across three synthetic datasets built so their underlying themes were known in advance, the workflow generated themes closely matching most of the original sub-themes, and produced progressively fewer new codes as it processed more clusters rather than inventing one per cluster. As of 2026.
Applied at scale to a real corpus: Used to analyze more than 10,000 Reddit posts about why people leave computer science, published in the Journal of Engineering Education. As of 2025.
Funded work: Developed under a $10,000 Virginia Tech Academy of Data Science Discovery Fund award, 2024–2025. As of 2025.

Maintenance

Maintained by: Andrew Katz
Status: Packaging the published method for release
Last reviewed: August 28, 2026

All tools