Humanist Discussion Group, Vol. 40, No. 98.
        Department of Digital Humanities, University of Cologne
                      Hosted by DH-Cologne
                       www.dhhumanist.org
                Submit to: humanist@dhhumanist.org




        Date: 2026-07-24 07:27:19+00:00
        From: 이용욱 <icerain@jj.ac.kr>
        Subject: Does AI-assisted humanities research need a methodology of its own?

Dear colleagues,

I would like to raise a methodological question that has emerged from my
recent experiments with AI-assisted humanities research:

Can a language model support humanities scholarship without allowing fluent
prose to conceal unfinished research?

Large language models can now produce academic-looking text with remarkable
speed. Yet the apparent coherence of that text may exceed the actual state
of the research. A research question may remain vague; the boundaries of a
concept may not have been established; a source may exist without
supporting the claim attributed to it; an interpretation may be presented
as fact; or a manuscript may appear complete before its central objections
have been seriously considered.

For this reason, I do not regard the problem primarily as one of prompting.
Better prompts may improve individual outputs, but they do not by
themselves provide a durable account of when evidence is sufficient, when
uncertainty must remain visible, when a researcher must intervene, or when
an apparently successful workflow should be stopped and returned to an
earlier stage.

I am therefore asking not whether AI can write humanities papers, but
whether AI-assisted humanities research requires a methodology of its own.

As one practical experiment in response to this question, I have developed
an open-source framework called Humanities Superpowers:

https://github.com/icerain-cmd/humanities-superpowers

Humanities Superpowers is not an autonomous research agent and not a
paper-writing system. It is a structured workflow intended to support
scholarly judgment when researchers work with AI agents.

The framework currently contains thirteen research skills and one routing
component. The skills address such tasks as formulating a research
question, defining the boundary of an argument, tracing a concept’s
lineage, organizing a dialogue with previous scholarship, performing close
reading, planning and stress-testing an argument, auditing citations,
maintaining terminological consistency, reviewing a manuscript, responding
to peer review, and verifying whether a submission is actually ready.

At designated points, the workflow requires an explicit decision: PASS,
CONDITIONAL PASS, or FAIL. A failure is not treated as a technical
malfunction. It indicates that evidence, verification, conceptual
clarification, or human judgment is still missing. When a later problem
reveals an earlier weakness, the workflow can return to the relevant prior
stage rather than polishing the downstream text.

The central premise is deliberately modest. The framework does not claim to
prevent hallucinations, guarantee citation accuracy, automate
interpretation, or produce publication-ready scholarship. Its purpose is to
make distinctions visible that fluent AI output can easily blur: verified
evidence and unverified material, interpretation and factual claim,
inference and hypothesis, completion and premature closure.

The first version has been tested with Claude Code and OpenAI Codex. It is
released under the MIT License, and its documentation is available here:

https://icerain-cmd.github.io/humanities-superpowers/

I am sharing the project with Humanist not simply to seek users, but to
invite criticism of its underlying assumptions.

In particular, I would be grateful for responses to the following questions:

1. Are there important stages or forms of humanities research that this
workflow fails to represent?

2. Which scholarly decisions cannot—or should not—be formalized as
procedures or quality gates?

3. Does the framework risk translating interpretive work into an overly
technical model of research?

4. Can explicit stopping and rollback procedures protect scholarly
judgment, or might they create a false impression that research quality can
be mechanically certified?

5. How would the framework need to change for different disciplines,
languages, source traditions, or research cultures?

My own position is that interpretation itself should not be automated away.
What may be made more explicit, however, are the conditions under which
interpretations are advanced, the evidence on which they depend, the
uncertainties they preserve, and the decisions for which the researcher
remains responsible.

Whether this distinction can be sustained in practice is an open question.
I would therefore sincerely welcome methodological objections, disciplinary
counterexamples, and suggestions for revision.

With best regards,

Lee Yong Wook
Professor of Korean Language and Literature
Jeonju University
Republic of Korea

Email: icerain@jj.ac.kr
GitHub: https://github.com/icerain-cmd


_______________________________________________
Unsubscribe at: http://dhhumanist.org/Restricted
List posts to: humanist@dhhumanist.org
List info and archives at at: http://dhhumanist.org
Listmember interface at: http://dhhumanist.org/Restricted/
Subscribe at: http://dhhumanist.org/membership_form.php