The
authors
describe
how
their
Generative
Artificial
Intelligence
(GAI)-assisted
qualitative
research
project
failed
to
produce
publishable
results.
Based
on
this
experience,
they
argue
for
the
value
of
embracing
and
reflecting
failure
in
GAI-assisted
research.
To
frame
argument,
draw
two
theories
generative
failure:
failing
forward,
which
uses
failures
iterate
designs
meet
existing
criteria,
sideways,
reconsiders
criteria
success.
Using
a
fail-forward
perspective,
might
revise
methods
data
preparation,
process
documentation,
task
delegation
create
more
reliable
Then,
using
fail-sideways
reexamine
results
reimagine
study
fundamentally.
Frontiers in Public Health,
Journal Year:
2025,
Volume and Issue:
13
Published: Feb. 19, 2025
Psychosocial
autopsy
is
a
retrospective
study
of
suicide,
aimed
to
identify
emerging
themes
and
psychosocial
risk
factors.
It
typically
relies
heavily
on
qualitative
data
from
interviews
or
medical
documentation.
However,
research
has
often
been
scrutinized
for
being
prone
bias
notoriously
time-
cost-intensive.
Therefore,
the
current
investigate
if
Large
Language
Model
(LLM)
can
be
feasibly
integrated
with
procedures,
by
evaluating
performance
model
in
deductively
coding
coherently
summarizing
interview
obtained
autopsy.
Data
38
semi-structured
conducted
individuals
bereaved
suicide
loved
one
was
coded
researchers
server-installed
LLAMA3
large
language
model.
The
evaluated
three
tasks:
(1)
binary
classification
segments,
(2)
independent
using
sliding
window
approach,
(3)
summarization
data.
Intercoder
agreement
scores
were
calculated
Cohen's
Kappa,
LLM's
summaries
qualitatively
assessed
Constant
Comparative
Method.
results
showed
that
LLM
achieved
substantial
(accuracy:
0.84)
task
0.67).
had
variability
across
codes.
rich
enough
subsequent
analysis
researcher,
around
80%
rated
independently
two
as
'adequate'
'good.'
Emerging
assessment
included
unsolicited
elaboration
hallucination.
State-of-the-art
LLMs
show
great
potential
support
complex
data,
which
would
alleviate
investment
time
resources.
Integrating
models
procedures
facilitate
near
real-time
monitoring.
Based
findings,
we
recommend
collaborative
model,
whereby
deductive
complemented
review,
inductive
further
interpretation
researcher.
Future
may
aim
replicate
findings
different
contexts
evaluate
larger
context
size.
Frontiers in Social Psychology,
Journal Year:
2025,
Volume and Issue:
3
Published: Feb. 21, 2025
Large
language
models
(LLMs)
are
being
used
to
classify
texts
into
categories
informed
by
psychological
theory
(“psychological
text
classification”).
However,
the
use
of
LLMs
in
classification
requires
validation,
and
it
remains
unclear
exactly
how
psychologists
should
prompt
validate
for
this
purpose.
To
address
gap,
we
examined
potential
using
classification,
focusing
on
ways
ensure
validity.
We
employed
OpenAI's
GPT-4o
(1)
reported
speech
online
diaries,
(2)
other-initiations
conversational
repair
Reddit
dialogues,
(3)
harm
healthcare
complaints
submitted
NHS
hospitals
trusts.
Employing
a
two-stage
methodology,
developed
tested
validity
prompts
instruct
manually
labeled
data
(
N
=
1,500
each
task).
First,
iteratively
three
types
one-third
coded
dataset,
examining
their
semantic
validity,
exploratory
predictive
content
Second,
performed
confirmatory
test
final
remaining
two-thirds
dataset.
Our
findings
contribute
literature
demonstrating
that
can
serve
as
valid
coders
phenomena
text,
condition
researchers
work
with
LLM
secure
semantic,
predictive,
They
also
demonstrate
rapid
cost-effective
iterations
over
big
qualitative
datasets,
enabling
explore
refine
concepts
operationalizations
during
manual
coding
classifier
development.
Accordingly,
secondary
contribution,
enable
an
intellectual
partnership
researcher,
defined
synergistic
recursive
process
where
LLM's
generative
nature
facilitates
checks.
argue
may
signify
paradigm
shift
toward
novel,
iterative
approach
improve
operationalizations.
ChatGPT
represents
a
groundbreaking
AI
application
that
has
garnered
significant
attention
since
its
inception.
However,
despite
promising
potential,
ethical
implications
have
sparked
considerable
debate.
This
study
aims
to
examine
the
key
concerns
surrounding
governance
of
by
conducting
bibliometric
analysis
and
cluster-based
content
relevant
scientific
literature.
The
identifies
influential
authors,
countries,
pivotal
publications,
revealing
three
primary
categories
issues
associated
with
ChatGPT:
human-related
ethics,
academic
integrity
technical
literacy,
artificial
intelligence
(AI)
technology
ethics
derived
concerns.
Additionally,
further
refines
these
synthesizing
frequently
occurring
keywords.
Building
on
this
framework,
provides
comprehensive
discussion
major
challenges
faced
ChatGPT,
as
well
outlining
future
research
priorities.
Furthermore,
investigates
knowledge
base
underlying
ChatGPT's
governance,
exploring
high-citation
high-link-strength
literature
through
co-citation
analysis,
thereby
mapping
landscape
highlighting
areas
growing
scholarly
interest.
offers
valuable
insights
for
policymakers,
researchers,
practitioners,
emphasizing
need
more
stringent
policies,
guidelines,
robust
design
in
development
similar
technologies.
Journal of Technical Writing and Communication,
Journal Year:
2024,
Volume and Issue:
unknown
Published: June 11, 2024
Previous
literature
has
shown
that
generative
artificial
intelligence
(GAI)
software,
including
large
language
model
(LLM)
chatbots,
might
contribute
to
qualitative
research
studies.
However,
there
is
still
a
need
examine
the
relationships
between
researchers,
GAI
technologies,
data,
and
findings.
To
address
this
need,
our
team
conducted
thematic
analysis
of
reflexive
journals
from
an
LLM
chatbot-assisted
project.
We
identified
four
roles
researchers
adopted:
managers
closely
monitored
LLM's
work,
teachers
instructed
on
theories
methods,
colleagues
openly
discussed
data
with
LLM,
advocates
worked
improve
user
experiences.
Planning
for
playing
multiple
also
helped
enrich
process.
This
study
underscores
potential
using
conversational
as
framework
support
reflexivity
when
working
technologies
research.
