Introduction
Vibe
coding changes who can build a corpus tool. The term was coined for a style of
programming in which a developer describes what software should do and accepts
the generated code with little or no inspection (Karpathy, 2025). In
research settings, a researcher can describe a concordancer,
a frequency analysis or a learning interface in ordinary language and ask an AI
system to generate the software. This possibility matters in applied
linguistics, where programming knowledge can restrict instrument development.
Crosthwaite and Özer (2026) demonstrate a structured approach
through VetLingua, a tool for veterinary language
learning. The central issue is how researchers establish control over the
instruments that AI helps them create.
That
issue extends an existing debate about corpora and generative AI. Corpus
analysis can identify patterns in a defined collection of texts. AI can explain
those patterns, generate practice materials and simplify interaction with
software. These functions may complement each other, but they produce different
kinds of evidence (Crosthwaite & Baisa, 2023; Mizumoto, 2023). Vibe
coding adds another function. AI becomes involved in building the instrument
through which language evidence is selected and analysed.
Errors at this stage can affect every later interpretation, even when the
interface appears clear and the output looks plausible.
Four
strands of recent work bear on this issue. Ma et al. (2024; 2026)
connect corpus use to teacher knowledge and pedagogical practice. Cheung and
Crosthwaite (2025) and Crosthwaite and Baisa (2024)
address corpus tools, AI integration and instrument development. Pérez-Paredes
(2026) and
Pérez-Paredes et al. (2025) develop a critical account of
corpus and AI literacies. Wong et al. (2026) map
the wider intersection of corpus linguistics and AI. These contributions
provide different foundations for the present argument. Their relevance should
be assessed through what each study establishes. Corpus pedagogy, AI literacy
and software development should not be treated as equivalent research areas.
The
closest precedent is Crosthwaite and Özer (2026)
methods tutorial, which reframes vibe coding as vibecoding-as-method.
It is an eight-stage workflow, situated within design-based research, in which
prompts are treated as methodological decisions that must be documented,
justified and audited. That tutorial describes itself as a first articulation
and it limits its knowledge claims to the development cycle. Its deployment,
logging and evaluation stages are presented as design intentions. The present
article begins where that account stops. A documented process shows how an
instrument was built. It does not show that the instrument computes what it is
specified to compute, that its outputs support the linguistic claims made from
them, or that learners who use it develop abilities they retain. These are
questions about evidence, and they arise after the development record is
complete.
The
article makes three contributions. First, it separates three claims that
discussions of AI-built tools tend to merge and it
identifies the evidence each requires and the inference each cannot support.
Second, it distinguishes how a tool is produced from whether AI continues to
act when the tool is used because the two carry different threats to
reproducibility. Third, it applies the resulting framework to a published
instrument and derives a research agenda that can be tested, including
comparisons of complete development workflows and measures of learning with
assistance withdrawn.
This
conceptual article offers an interpretative narrative synthesis of current
research and selected foundational studies. It argues that vibe coding can
reduce the effort required to create a tool while increasing the need to
justify its analytical decisions. Following this argument, the review first
clarifies the relation between vibe coding and corpus inquiry. It then examines
the main methodological and pedagogical problems. The later sections propose a
framework for validation, apply it to a published case and set out a research
agenda. The future trend considered here is conditional. Its value will depend
on how tool creation is connected to credible evidence and independent
learning.
Approach
and scope
This
article is conceptual. Its synthesis is purposive and interpretative and it makes no
claim to exhaustive coverage. It draws on three bodies of work, each of which
supplies a standard against which a vibe-coded tool can be judged. Foundational
corpus linguistics supplies standards for sampling, counting and interpretation
(Biber, 1993;
Gries, 2008;
Kilgarriff,
2005). Research on data-driven learning
supplies evidence about what corpus use contributes to learning, and under
which conditions (Boulton & Cobb, 2017; Boulton &
Vyatkina, 2021). Studies of corpus-AI integration
published since 2023 supply the current tools, claims and cautions (Ren et al., 2026;
Zhang et al., 2025).
Work on vibe coding outside applied linguistics is used more selectively to
locate the practice and its known reliability problems (Alqheedan & Alzughaibi, 2026;
Elgendy et al.,
2026; Ouyang et al., 2025).
Sources
are cited for what they establish. A study of teacher self-efficacy is not
treated as evidence of pupil learning, and a meta-analysis of corpus use is not
treated as validation of a new way of building corpus software. The unit of
analysis is the claim made from a tool, not the tool itself. Throughout, “vibe-coded
tool” refers to software produced mainly through natural language interaction
with a generative model, whether or not the finished
program calls a model when it runs. The article reports no new empirical data.
Its framework is a proposal and its illustrative case
rests on a published account, not on independent testing of the instrument.
Background:
vibe coding and corpus inquiry
Understanding
vibe coding in corpus-based language research and learning
Vibe
coding refers here to developing software through repeated natural language
interaction with an AI system. The user describes a function, inspects the
result and requests changes. This usage is broader than the original one. Karpathy (2025) described a practice in which the code
itself is largely ignored. Research use cannot adopt that stance, because the
code is the instrument. Work in information systems has accordingly begun to
treat reliability and governance as open problems for vibe coding (Elgendy et al.,
2026). In health informatics, Alqheedan and Alzughaibi (2026)
propose a conceptual framework for judging the appropriateness, quality and
safety of vibe-coded clinical tools and illustrate it with published cases. No
comparable account exists for language research. Crosthwaite and Özer (2026)
adapt this process into a documented workflow for research instruments. Their
contribution concerns methodological design and a worked development case. It
should be read as a proposal requiring further validation across contexts. A
successful prototype establishes that a function can be built. Claims about
analytical accuracy, usability or learning require additional evidence.
The
phrase “vibe-coded AI” can obscure this distinction. A language researcher
usually uses an existing AI model to generate software. The researcher may
build a corpus application without developing a new AI model (Wong & Fong,
2025). Moreover, the resulting application
does not necessarily use AI during analysis. A generated program can count
words through fixed rules. Another program may ask a language model to label
texts or explain concordance lines. These designs carry different
uncertainties. Researchers must identify where AI acts, which operations remain
fixed and which outputs can vary between runs.
Corpus
linguistics provides a demanding test of such development. A corpus is a
collection assembled for a stated purpose, with identifiable sources and a
defined sampling logic (McEnery & Hardie, 2012; Sinclair, 2005).
Biber (1993)
shows why representativeness depends on the population and the dimensions of
variation being studied. A large collection of accessible texts cannot
automatically represent a language, a discipline or a learner group. Vibe
coding can accelerate collection and processing, but it cannot establish the
relation between a sample and a research claim. That relation remains a
theoretical and empirical responsibility.
The
same principle applies to quantitative analysis. Frequencies depend on what
counts as a word and how the corpus is segmented. Their interpretation also
depends on dispersion, or how occurrences are distributed across texts (Gries, 2008).
Repeated uses by one writer may create a high frequency without indicating a
widespread pattern. Kilgarriff (2005)
challenges the assumption that language data behave like simple random
observations. Together, these arguments show why convenient counts need careful
interpretation. An AI-generated script can reproduce a formula accurately while
encouraging an unsuitable comparison.
Automated
analysis already has an established history in language research. Lu (2010)
system for syntactic complexity evaluates specified measures against learner
writing data. Research on lexical richness also links the interpretation of
indices to the language performance being investigated (Lu, 2012).
These studies offer an important standard for vibe-coded tools. Their value
rests on defined measures and evaluation, rather than automation alone. A new
application should explain how it defines a construct and where its
measurements have been checked. Familiar labels such as “complexity” or “proficiency”
cannot substitute for that explanation.
Vibe
coding is also relevant to data-driven learning (DDL), where learners
investigate patterns in language data. The central learning activity is the
interpretation of evidence. Access to a computer is only one possible means of organising that activity. Boulton’s work on paper-based DDL
demonstrates why corpus learning need not depend on direct software use (Boulton, 2010).
Frankenberg-Garcia (2014) further shows that corpus examples for
comprehension and production serve different purposes. Therefore, an interface
becomes educationally meaningful when its functions support the intended
learning process. Greater technical sophistication is not a sufficient
educational justification.
The
evidence for corpus learning supports cautious development. Boulton and Cobb (2017)
identify substantial learning benefits across the studies reviewed. Those
findings support corpus use under the conditions represented in that
literature. They do not validate every corpus application or a new method of
generating software. Boulton and Vyatkina (2021)
show how DDL has developed over time, while Pérez-Paredes (2022)
examines its spread within CALL research. These reviews make a cumulative
research base available. Vibe coding should extend that base through comparable
evidence, instead of treating technical novelty as a fresh starting point.
The
practical gap remains significant. Chambers (2019)
examines the distance between corpus research and ordinary teaching practice. CorpusMate responds to interface difficulties through a
tool designed around disciplinary learning needs (Crosthwaite & Baisa, 2024).
Charles (2024)
discusses the benefits and demands of learners constructing their own corpora
for academic writing. These contributions suggest that access involves more
than obtaining software. Teachers and learners need relevant texts, manageable
tasks and enough support to interpret findings. Vibe coding may help tailor
those elements, but it can also transfer design and maintenance work to people
who already have limited time.
Ma
et al. (2024)
connect corpus pedagogy to language knowledge, technology, teaching practice
and context. This account explains why the ability to request a feature is only
a small part of pedagogical design. A teacher must decide what learners should notice, how much context they need and what they can
reasonably infer. Ma et al. (2026) relate corpus training and knowledge to
student teachers’ self-efficacy. Confidence is relevant to adoption, but it is
a different outcome from pupils’ language development. Applying this work to
vibe coding therefore requires evaluation of both teacher preparation and
classroom learning.
Recent
tools make the convergence between corpora and AI more concrete. CorpusChat integrates corpus linguistics and generative AI
for academic writing, while ChatAI addresses the
integration of AI with corpus-based learning (Anthony, 2025;
Cheung &
Crosthwaite, 2025). Their relevance lies in how they
connect language evidence to interaction and support. They should not
automatically be classified as vibe-coded tools. Corpus–AI integration concerns
what an application does. Vibe coding concerns how an application is developed.
Keeping these questions separate allows researchers to evaluate the
contribution of AI-supported development without attributing every benefit of a
corpus interface to its production method.
Pérez-Paredes
(2026)
positions corpus literacy as a basis for transparent AI-mediated learning.
Pérez-Paredes et al. (2025) examine this relation through workshops
involving corpus consultation and ChatGPT. Their findings support explicit
attention to reliability and critical engagement. Vibe coding extends that
concern to instrument creation. Users need to examine the choices built into a
tool, as well as the answers produced by it. The important development is
therefore a possible redistribution of methodological authority. Researchers
and teachers gain more influence over software design, but
must also account for decisions previously embedded in established tools.
From
documented process to warranted claims
Crosthwaite and Özer (2026) provide the most developed
account of vibe coding as a research practice in applied linguistics. Their
workflow runs from problem specification and corpus preparation through feature
operationalisation, prompt design, and interface and
backend generation, to deployment, logging and evaluation. Several of its
principles are adopted here without modification. Corpus preparation should
remain outside the AI interface, so that generated text does not enter the
evidence base. Constructs should be grounded in theory and checked by the
researcher against the corpus before they are written into prompts. Prompts,
model versions and settings should be recorded. Earlier functions should be
retested after each change. The tutorial also states four conditions under
which a vibe-coded system can count as a research instrument: theoretically
precise operationalisation, consistent and documented
elicitation conditions, a replicable record of prompt iterations, and openly
available code.
These conditions concern how an instrument is
produced and recorded. They are necessary. They are not sufficient for the
claims researchers and teachers eventually make. A prompt history can be
complete while the retrieval function it produced miscounts across document
boundaries. A construct can be defined with care while the tool’s generated
explanations misdescribe the concordance lines they accompany. A tool can meet
all four conditions and still leave learners no better able to consult evidence
without it. The tutorial does not claim otherwise. It positions vibecoding-as-method as a pre-deployment methodology and
calls for validation across further contexts. The question it leaves open is
what that validation should consist of.
Two further distinctions shape the answer. The
first is between verification and validation, a distinction familiar from
software engineering. Verification asks whether a program does what its
specification says. Validation asks whether the specification, and the
interpretation placed on its output, are appropriate to the purpose. Kane (2013) frames the second question
as an argument in which each inference from an observed score to a claim needs
its own warrant. The second distinction is between AI in development and AI in
operation. Where a model only writes the code, the finished program can be
tested once and will then behave consistently. Where a model is called each time the tool is used, as in tools that generate
explanations, exercises or feedback, the behaviour
under test can change between runs and between model versions. Validation
evidence for the second kind of tool has a shorter life and must be renewed.
Critical issues in vibe-coded corpus research and learning
Convenience, provenance and the integrity of
evidence
The first problem
concerns the relation between convenience and evidence. A clear interface can
make outputs appear settled, especially when charts and explanations arrive
together. However, users may understand the displayed result better than the
process that produced it. Messeri and Crockett (2024)
explain how AI can create an illusion of understanding in scientific work. In
corpus research, this risk arises when fluency or visual polish conceals
sampling choices, preprocessing rules or uncertain labels. The response should
be to make these choices inspectable. A general warning to “use AI critically”
provides little protection unless users can identify what requires checking.
Corpus provenance
is central to that response. Researchers need to know where texts came from,
why they were selected and what happened to them during preparation.
Crosthwaite and Baisa (2023) emphasise the
advantages of identifiable corpus evidence over opaque model output. Vibe
coding does not remove this advantage, but poorly specified software can weaken
it. A script may merge documents, remove metadata or silently exclude difficult
files. The final dataset can then differ from the intended sample. Researchers
should compare the prepared corpus with the source collection and account for
exclusions before interpreting a frequency list. Crosthwaite and Özer (2026)
make the same point procedurally, by requiring that corpus preparation take
place outside the AI interface. The requirement protects the source texts. It
does not by itself show that the generated ingestion code preserved them, which
is why the prepared corpus still needs to be checked against its source.
Generated language
creates a related problem. Synthetic examples can support practice or test a
program, but they need clear labels and a purpose appropriate to their origin. Shumailov et al. (2024)
show how recursive training on generated data can distort model distributions.
That finding concerns model training and should not be transferred directly to
every classroom use of generated examples. Its relevance here is more specific.
Generated material cannot be assumed to preserve the variation found in
observed language. Mixing it with authentic data without separate reporting can
change the object that a corpus study claims to investigate.
Automated
correction can have a similar effect. If a program rewrites learner texts
before analysis, some features of learner language may disappear. An apparent
improvement in corpus quality can become a loss of evidence about development
or variation. Barrot (2023) discusses both the potential and the
pitfalls of ChatGPT in second language writing. For corpus research, the
critical decision is whether AI output is part of the phenomenon being studied
or an intervention in the data. Original texts, edited texts and generated
texts should remain distinguishable. Otherwise, the study cannot explain whose
language it has measured.
Prompts, specification and reproducibility
A second problem
concerns prompts as methodological decisions, the central claim of Crosthwaite
and Özer (2026). Evidence for how much a prompt can
change an outcome comes from a neighbouring line of
work. Crosthwaite et al. (2026) compared 145 student essays from the
LOCNESS corpus with two sets of 145 GPT-generated essays on the same topics.
Under a general essay prompt, the generated texts contained fewer transitions,
hedges, self-mentions, attitude markers and engagement resources than the
student writing. Under a prompt that explicitly requested such features,
several of them rose above both the baseline and the human frequencies, and
they were often realised in more formulaic ways.
Their study cautions against interpreting one elicitation condition as a stable
property of AI writing. The implication for vibe coding is related, although it
requires its own evidence. Instructions such as “remove noise” or “find
academic phrases” leave important choices open. The model may decide what noise
or a phrase means. Researchers should specify those choices before development
and report the decisions that remain uncertain.
Natural language
instructions can also conceal technical disagreements. A request to “compare
common phrases” might lead to raw frequencies, normalised
counts or a statistical comparison. Each answer addresses a different question.
Changing the wording until a result looks convincing may alter the analysis
along with the interface. Prompt refinement must therefore be tied to a written
specification. That specification should define the unit of analysis, the
comparison and the expected output. The record should show which changes
corrected an error and which changed the research question. Otherwise, a
development history may become an undocumented sequence of analytical
revisions.
Reproducibility
requires more than saving prompts. Repeating a prompt may generate different
code because the model or development environment has changed. A fixed program
can produce stable results even when the process that created it was variable.
Conversely, a program that calls an AI model during annotation may remain
variable after its code is saved. Research teams should preserve the final
code, dependencies, settings and permitted data materials. They should also
record model details where AI remains part of analysis. Marsden et al. (2018)
make replication a substantive concern in language research. Vibe coding needs
to meet that concern through reproducible operations and transparent limits.
The difficulty is already documented for generated code. In a study of 829 code
generation problems, Ouyang et al. (2025)
found that identical requests to ChatGPT often returned programs that behaved
differently, and that setting the temperature to zero reduced this variation
without removing it.
From correct computation to valid
interpretation
A third problem is the confusion between correct computation and valid
interpretation. A program may count every occurrence accurately while counting
the wrong feature for the research question. Kane (2013) locates validity in the justification
of interpretations and uses of scores. Although his account concerns
assessment, its reasoning can inform corpus instrument evaluation. The
researcher must connect an observed output to the claim made from it. For example,
an index labelled “academic style” requires an account of which language
features it measures and why they support that interpretation. A working
calculation cannot supply this argument on its own.
Annotation makes this problem especially visible. Mizumoto and Eguchi (2023) evaluate
AI-assisted essay scoring against an established learner corpus. Their study
illustrates the need to investigate performance for a specified task and
dataset. It cannot validate a later model, prompt or classroom population
automatically. Likewise, a vibe-coded tool that labels stance must be evaluated
using a clear definition and independently adjudicated examples. Agreement
between repeated AI runs establishes a form of consistency. It does not show
that the labels capture the intended linguistic function. Consistent
misclassification remains a threat to interpretation.
Quantification should also preserve the possibility of close reading.
Baker et al. (2008) demonstrate how corpus analysis and
critical discourse analysis can inform one another. An AI-generated summary can
help locate a pattern, but it may conceal ambiguity or disagreement within the
corpus. Researchers should inspect original contexts, including examples that
challenge the dominant interpretation. The issue is particularly important when
a tool turns a word association into a claim about ideology or social identity.
Such claims require contextual evidence and theoretical reasoning. A model’s
fluent explanation is a proposed interpretation that must remain open to
revision.
Learner reasoning and pedagogical design
The pedagogical problem is different but closely connected. Easier
interaction can reduce the effort of searching a corpus. It can also remove the
reasoning that gives DDL its educational value. O’Sullivan (2007) treats corpus consultation as part of a
process of literacy development. Mizumoto (2023) similarly connects resource use to
metacognitive regulation. These accounts suggest that a learner should decide
what evidence is needed and whether it supports a claim. An application that
supplies an answer before the learner examines examples may improve immediate
task completion while limiting opportunities to develop those decisions.
This does not mean that every difficulty should be retained. Learners
can spend excessive effort navigating syntax, correcting search errors or
managing irrelevant results. Risko and Gilbert (2016) explain how cognitive offloading can
reduce task demands, while also depending on judgements about one’s abilities.
For corpus learning, support should reduce avoidable technical effort while
preserving the reasoning targeted by the activity. A tool might simplify a
search but still ask learners to compare contexts and explain exceptions. The
appropriate amount of assistance depends on the task and the learner. It should
be investigated through learning outcomes, rather than inferred from ease of
use.
Process evidence can help determine what learners actually
do. Pérez-Paredes et al. (2011) use logs to
investigate corpus consultation under different guidance conditions. Their
later work examines learners’ search patterns during corpus-based activities (Pérez-Paredes et al., 2012). These approaches remain relevant to
AI-mediated interfaces, where a completed text may conceal very different
learning processes. Researchers need to examine whether learners inspect
evidence, revise a query or simply accept an explanation. Logs can record these
actions, but their meaning still requires interpretation. A
large number of clicks cannot establish attention, understanding or
transfer without supporting evidence. Such evidence has to
be planned. Crosthwaite and Özer (2026) report that VetLingua’s analytics currently record access and traffic
but not queries or interaction sequences and they conclude that research-grade
logging must be designed into an instrument from the outset.
Current corpus–AI writing studies offer useful
directions, but their findings need careful attribution. Dong and Wang (2025) investigate
a DDL–GenAI teaching mode. Wang et al. (2025) examine corpus-driven and AI-assisted
approaches alongside learner variables. These studies address instructional
combinations. They do not isolate whether generating the software through vibe
coding improves learning. Outcomes could depend on corpus selection, feedback,
teacher guidance or time on task. A study of vibe coding should therefore
define the contribution being tested. Development efficiency and instructional
effectiveness are legitimate outcomes, but they require different comparisons.
Teachers, access and accountability
Teacher knowledge is another condition that cannot be delegated to
software. Ma’s work on corpus pedagogy shows why technology must connect to
content, pedagogy and context (Ma et al., 2024). Celik (2023) extends this concern to teachers’ professional knowledge for ethical AI integration. These perspectives
challenge the assumption that removing a programming barrier also removes the
need for preparation. Teachers still need to recognise
misleading examples, unsuitable difficulty and unsupported claims. They also
need a feasible route to technical assistance. Requiring every teacher to
maintain a custom application could increase workload and deepen differences
between well-supported and under-supported institutions.
Access must therefore be assessed beyond the ability to enter a prompt.
Wong et al. (2026) map 162 selected open-access records
from Scopus between 2020 and mid-2025. Their review identifies thematic and
geographical patterns in a bounded publication sample. It cannot establish the
prevalence of vibe coding or its classroom effectiveness. Its value for this
discussion is to make research visibility and uneven participation objects of
inquiry. Language communities with limited digitised
data or technical support may gain less from rapid tool generation. A future
research agenda should examine who can create, validate and sustain corpus
applications.
The professional question concerns who remains
accountable for those applications. Kern (2021) connects digital literacy to the social
practices through which technologies are used. Ng et al. (2021) include
understanding, evaluation and ethical concerns within AI literacy. Vibe coding
brings these responsibilities into software production. Researchers need
decisions about data access, consent where required and the handling of
identifiable learner texts. They also need a plan for failed analyses and
abandoned tools. A system that is easy to create but difficult to inspect or
maintain can shift costs to students, colleagues and future researchers.
A framework for validating vibe-coded corpus tools
Three forms of validation
The future value of
vibe coding should be judged through the claims a tool supports. This article
proposes three linked forms of validation. Computational correctness concerns
whether the program carries out its specified operations. Linguistic validity concerns
whether the data and measures justify the language claim. Pedagogical validity
concerns whether the activity supports the intended learning. These checks
adapt established concerns to the development of corpus instruments. They are a
conceptual framework for investigation, rather than an empirically validated
model. Table 1 identifies the evidence each check would require and the
conclusion it cannot establish by itself.
Computational
checking should begin with a small corpus whose expected results are known.
Researchers can specify how the tool should handle punctuation, empty files and
document boundaries. They should then compare its output with those
expectations and, where appropriate, an established implementation. Different
tools may use different definitions, so disagreement requires investigation. AI
can help write tests or suggest difficult cases. However, the same model should
not supply the only definition of correctness for code it generated. The
reference specification and expected results need an independent basis. Tests
should be repeated after changes that affect analysis. In software terms these
are known-answer tests, and
repeating them after each change is regression testing. Crosthwaite and Özer (2026)
recommend retesting earlier functionality after each prompt cycle. The addition
proposed here is that the expected results be fixed in advance and derived
independently of the model.
Table 1. Evidence
required to validate corpus tools
|
Form of validation
|
Evidence required
|
What it does not establish
|
|
Computational correctness
|
Known-answer examples, checks of document
boundaries and comparisons using the same analytical definitions.
|
That the measure captures the intended
linguistic construct.
|
|
Linguistic validity
|
A justified corpus sample, defined constructs,
inspection of original contexts and analysis of alternative explanations.
|
That using the tool leads to language
learning.
|
|
Pedagogical validity
|
Appropriate comparison tasks, delayed
assessment, evidence of learner reasoning and transfer with reduced
assistance.
|
That every corpus, model or learner group will
show the same benefit.
|
Note. This conceptual
framework is proposed in the present article. It requires empirical evaluation.
Linguistic validation requires
the next step in the argument. Consider a hypothetical application that
extracts lexical bundles from student essays. Biber et al. (2004) show that bundles vary across instructional and academic registers. Gries
(2008) makes dispersion relevant to frequency interpretation. Following these
insights, the application should expose its tokenisation
rules, thresholds and distribution across writers. It should preserve document
boundaries and provide the original contexts. A technically correct list may
still be unsuitable for comparing proficiency groups if their tasks,
disciplines or text lengths differ. Researchers must justify the comparison
before assigning developmental meaning to the results.
The same example shows why
validation should be iterative. An unexpected bundle may reveal a preprocessing
error. It may also identify a real phrase concentrated in one assignment. The
first problem requires a software correction. The second requires an interpretation
of corpus composition. Treating every unexpected result as a bug would remove
potentially important evidence. Treating every output as a discovery would
preserve avoidable errors. Researchers should document how they resolved each
case and retain the relevant data. Such decisions make the analysis reviewable
and help later researchers distinguish a methodological choice from a technical
failure.
Pedagogical validation
requires evidence of learning beyond the immediate interface. A hypothetical
writing tool could retrieve authentic concordance lines and offer an AI
explanation in a separate panel. Learners could first propose a pattern, then
inspect the explanation and test it against further examples. The design should
keep corpus evidence accessible and identify generated content. Pérez-Paredes (2026) provides a foundation for this approach. Assessment should include a
later task with reduced assistance or a new topic. That comparison would help
establish whether learners acquired a strategy they can use independently,
rather than only completing the supported task more successfully.
The three forms are ordered and the order matters for interpretation. Linguistic
claims cannot be stronger than the computation beneath them, and a learning
gain attributed to corpus evidence presupposes that the evidence shown to
learners was sound. The dependence does not run in reverse. A tool can pass
every computational test and support no defensible linguistic claim and a linguistically sound tool can still fail to help
learners. Reports of vibe-coded tools should state which of the three claims
they make and present the corresponding evidence. A development report that
makes only the first claim is legitimate, provided it is not read as supporting
the other two. Where a tool calls a language model in operation, each form of
evidence should also carry the model version and date, since a later model can
invalidate earlier checks.
An illustrative
application: VetLingua
The framework can be
illustrated with the instrument that Crosthwaite and Özer (2026) describe. VetLingua is a vibe-coded tool for
veterinary academic English built on the Veterinary Medicine Corpus. It offers
a writing assistant, vocabulary practice and writing feedback, and users can
query it through preset buttons or free conversation. The reading that follows
relies only on the published tutorial. It is not an evaluation of the tool and it is offered to show what the framework asks of a
well-documented case. Table 2 summarises the
comparison.
On computational correctness,
the tutorial reports iterative debugging and one failure in detail: queries in
English were answered in Spanish. The fault was reported by users and then
corrected through further prompting. The episode shows the workflow’s capacity
for repair. It also shows how the fault was found, which was through use and
not through a test written in advance. The tutorial does not report
known-answer tests of retrieval from the corpus. It also notes that developers
may prefer to share a technical guide in place of code that contains personal
API keys. By the authors’ own four conditions, an instrument shared in that way
would retain prototype status, and independent verification of what it
retrieves would be correspondingly limited.
On linguistic validity, the
case is strong at the level of the corpus and less clear at the level of the
output. The corpus was compiled and documented independently of the model. The
target features were grounded in established accounts of stance and hedging and
operationalised through criteria drawn from prior
corpus analysis. The tutorial states that summaries are produced from
concordance hits retrieved from the corpus, so that they can be verified
against it. However, an early development prompt asks that the application generate
sentences similar to those in the corpus files instead
of retrieving them, and the vocabulary and feedback modules produce exercises
and corrections through the model. VetLingua belongs
to the second kind of tool distinguished above, in which AI acts in operation.
For such a tool, linguistic validation would need to establish which outputs
are retrieved and which generated, whether users can tell them apart and how
often generated explanations agree with expert reading of the same lines. This
bears on a tension in the tutorial’s guidance on records. Failed iterations are
to be removed so that a clear blueprint remains, yet divergent outputs are also
described as methodological data to be documented. A validation record needs
the second.
On pedagogical validity, the
authors are explicit that evidence does not yet exist. VetLingua
had not been deployed in a classroom, pilot feedback had been used informally,
and the planned evaluation combines pre- and post-intervention measures with
interaction data and user outputs. The framework adds two requirements to that
plan. A post-test taken with the tool available would show supported
performance, whereas a delayed task with assistance reduced would show whether
learners can consult evidence unaided. Because the tool supplies explanations
alongside examples, logs would also need to show whether learners inspect the
examples at all. None of this is a criticism of a first articulation that
states its own limits. It indicates the distance between a documented prototype
and the claims that such tools are likely to attract once they are in use.
Table 2. The three forms of validation applied to
the published account of VetLingua
|
Form of validation
|
What the published tutorial reports
|
What remains to be shown
|
|
Computational correctness
|
Iterative prompting and debugging. A
language-detection fault (English queries answered in Spanish) reported by
users and corrected. A recommendation to retest earlier functionality after
each prompt cycle.
|
Known-answer tests of corpus retrieval.
Expected results fixed independently of the model. A test record that others
can inspect.
|
|
Linguistic validity
|
An independently compiled and documented
disciplinary corpus. Constructs grounded in accounts of stance and hedging
and checked by the researchers. Summaries built from retrieved concordance
hits.
|
Which outputs are retrieved and which
generated, and whether users can tell. Agreement between generated
explanations or feedback and expert reading.
|
|
Pedagogical validity
|
Three learning modules. Informal pilot
feedback. Access-level analytics. A planned evaluation with pre- and
post-intervention measures.
|
Classroom deployment. Logs showing whether
learners inspect corpus evidence. Delayed and transfer tasks with assistance
reduced.
|
Note. Based solely on
Crosthwaite and Özer (2026). The instrument was not tested for this article.
Implications for design, teacher education
and institutions
Interface design can make such
evidence seeking more feasible. CorpusMate and ChatAI indicate ways to connect established corpus
functions with more accessible interaction (Anthony, 2025; Crosthwaite & Baisa, 2024). A vibe-coded application could adapt those functions to a local
discipline or language. The design should allow users to recover the query,
inspect sources and change a relevant setting. Automatic explanations should be
distinguishable from retrieved examples. A learner who can see how a result was
produced has a better basis for questioning it. Nevertheless, transparency
itself is a design property. Whether learners use it productively remains an
empirical question.
Teacher education should
connect creation to judgement. A useful training task would ask teachers to
describe a learning objective before requesting software. They would then
compare the generated application with that objective and inspect a small set of
known examples. Ma et al. (2026) provide a basis
for connecting corpus literacy, pedagogical knowledge and self-efficacy. Kohnke
et al. (2023) identify
competencies relevant to ethical and effective chatbot use. Vibe coding
requires additional preparation for specifying and evaluating instruments. That
preparation can be shared across a team. Pedagogical expertise and technical
review do not have to reside in the same individual.
This shared approach offers a
more feasible institutional model. Research groups could maintain a small
number of documented corpus components that teachers adapt with support.
Institutions would provide technical review, appropriate data storage and scheduled
maintenance. Local users would retain control over objectives, corpus selection
and teaching tasks. The arrangement could reduce repeated development work
while allowing contextual adaptation. It should also identify who responds when
a tool fails. Kern (2024) connects
language education to the wider conditions of technological mediation. His
perspective directs attention to how institutions organise
use, alongside the features of an individual application.
Research agenda
The
first research priority is to compare complete development workflows. A study
could assign comparable corpus tasks to conventional programming, documented
AI-assisted development and exploratory vibe coding. It should examine
analytical accuracy, development time and the effort required to detect and
repair errors. A functioning prototype is only one outcome. Maintenance and the
ability of another researcher to reproduce the analysis are also relevant.
Crosthwaite and Özer (2026) provide a starting workflow for this
investigation. The comparison would test when natural language development
improves practical access and when its apparent speed creates additional
checking costs.
A
second priority is to establish reusable evaluation materials. These could
include permitted sample corpora, documented expected outputs and challenging
linguistic cases. They should represent different genres, languages and learner
populations. Lu (2010) evaluation of automated syntactic
analysis illustrates why a tool should be checked against data beyond the
material used to develop it. New corpus applications also need evaluation after
substantial software or model changes. An accumulated record of failures would
be useful evidence. It would show where a development method needs additional
support and prevent each team from discovering the same weaknesses
independently.
A
third priority concerns the design of human–AI work. Vaccaro et al. (2024)
show through a systematic review and meta-analysis that combining humans and AI
does not guarantee performance beyond the better performer alone. This evidence
spans tasks outside corpus learning, so it supplies a caution rather than a
direct prediction. Corpus studies should examine which decisions humans make,
what information AI supplies and how disagreements are resolved. An expert
checking every result may produce high accuracy at an unsustainable cost. An
untrained user accepting every result may produce apparent efficiency with
limited validity. The relevant unit of evaluation is the complete workflow
under realistic conditions.
Pedagogical
studies should also separate outcomes more carefully. They need measures of
immediate performance, delayed learning and transfer to a new task. They should
examine differences associated with prior corpus experience and language
proficiency. Boulton and Vyatkina (2024)
call for broader methodological approaches in DDL research. Han (2024)
likewise argues for systematic inquiry into ChatGPT and second language
acquisition. These concerns support studies that combine outcome measures with
observations of evidence use. A favourable
questionnaire can indicate acceptability. It cannot establish that learners
have developed language knowledge or the ability to evaluate a new claim.
Claims
about replacing teachers require particular restraint.
Handley (2024)
questions whether AI has made language teaching obsolete. For vibe-coded corpus
tools, teacher involvement remains relevant to selecting tasks, interpreting
learner difficulties and adjusting support. Automation may change how that work
is organised. Its educational consequences will
depend on which decisions are transferred and what expertise remains available.
Studies should report the teacher input required to achieve their results.
Otherwise, a tool may receive credit for learning that depended on sustained
human guidance. Reporting this input also makes replication and realistic
adoption more feasible.
Future
research should extend the populations and purposes represented in corpus–AI
studies. O’Keeffe (2021) calls for a broader research gaze in
DDL. Curry and McEnery (2025) identify research directions connecting
corpus pedagogy to curricula, AI and wider language education. Vibe coding
could help develop applications for less commonly taught languages or
under-resourced settings. That possibility needs locally grounded evaluation.
Researchers should involve relevant teachers and language communities in
defining useful outputs. They should also investigate whether the supporting
models and corpus resources handle local variation adequately. Rapid
development alone cannot resolve the absence of suitable language data.
Critical
literacy should become observable in this research. Pérez-Paredes et al. (2025)
connect corpus work to critical engagement with AI. Building on that
connection, a study could ask whether learners identify a misleading
explanation, select an appropriate comparison or justify a revision using
source evidence. These actions would provide stronger grounds for claims about
critical learning than general confidence alone. They would also help explain
differences between designs that provide similar answers. The educational
contribution of a corpus–AI interface may lie in how users investigate a claim,
rather than the amount of information the interface supplies.
Crosthwaite
(2026)
challenges the claim that generative AI makes DDL obsolete. The present
argument extends that debate to the tools through which DDL
and corpus research are conducted. AI-supported development could make specialised instruments more accessible. Its lasting
contribution would depend on reliable components, documented decisions and
support that survives the initial prototype. The strongest future trend would
therefore be a growth in accountable instrument development and evidence-based
learning practices. A rapid increase in custom applications would be
insufficient evidence of that change. Research must establish what those
applications allow people to know and learn.
Limitations
This
article has several limitations. It is conceptual and reports no new data.
Hence, the usefulness of the framework has yet to be shown in practice. Its
literature base is purposive and a systematic search
might locate studies that qualify the argument. The illustrative case rests on
a single published tutorial and was not checked against the tool itself, whose
features may have changed since publication. The three forms of validation are
separated for clarity and real projects will meet
cases in which they are hard to tell apart, as when a tokenisation
choice is at once a computational and a linguistic decision. Finally, the field
is moving quickly. Statements here about what current tools can and cannot do
should be read as dated to 2026.
Conclusion
Vibe
coding could widen participation in corpus-based language research and
learning. It allows more researchers and teachers to influence the tools they
use. However, it also places methodological decisions within prompts, generated
code and interface defaults. Those decisions need explicit justification. The
proposed framework links computational checking to linguistic interpretation
and pedagogical evaluation. It explains why correct outputs, credible language
claims and learning gains require different evidence.
The
contribution of current corpus–AI research is to make this evaluation possible.
Corpus pedagogy clarifies the role of teacher knowledge and context. Corpus
literacy supports the inspection of language evidence. Methodological research
connects instrument development to documentation and replication. The present
framework adds the step that follows documentation: stating which claim a tool
is meant to support and gathering the evidence that claim requires. These
foundations should guide future studies of vibe coding. Comparative workflow
studies and delayed learning measures are especially needed.
The
future of this approach will depend on how institutions support it. Independent checking, shared technical expertise and
continued maintenance can make local innovation more sustainable. Without that
support, easier creation may expand the number of tools faster than the
capacity to evaluate them. Vibe coding becomes a credible research and educational
method when users can explain how their tools work, why the evidence supports
their claims and what learners can do independently after using them.
Declaration
of Interest Statement
There is no potential conflict of interest.
Generative
AI statement
Generative AI was employed solely to enhance linguistic
clarity during the final proofreading stage. All conceptualisation,
analysis and interpretive judgements remain entirely the author’s own.
Funding
There is no funding received.
References
Alqheedan, A., & Alzughaibi,
S. (2026). The VIBE-HI framework: A conceptual model for evaluating vibe coding
appropriateness, quality, and safety in health informatics. Frontiers in Artificial Intelligence, 9,
Article 1882814. https://doi.org/10.3389/frai.2026.1882814
Anthony, L. (2025). Integrating AI
technology into corpus-based language learning through ChatAI.
Computer Assisted Language Learning.
Advance online publication. https://doi.org/10.1080/09588221.2025.2589747
Baker, P., Gabrielatos, C., KhosraviNik, M., Krzyżanowski, M., McEnery, T., & Wodak, R. (2008). A useful methodological synergy?
Combining critical discourse analysis and corpus linguistics to examine
discourses of refugees and asylum seekers in the UK press. Discourse & Society, 19(3),
273–306. https://doi.org/10.1177/0957926508088962
Barrot, J. S. (2023). Using ChatGPT for
second language writing: Pitfalls and potentials. Assessing Writing, 57,
Article 100745. https://doi.org/10.1016/j.asw.2023.100745
Biber, D. (1993). Representativeness in
corpus design. Literary and Linguistic
Computing, 8(4), 243–257. https://doi.org/10.1093/llc/8.4.243
Biber, D., Conrad, S., & Cortes, V.
(2004). If you look at …: Lexical bundles in university teaching and textbooks.
Applied Linguistics, 25(3), 371–405. https://doi.org/10.1093/applin/25.3.371
Boulton, A. (2010). Data-driven learning:
Taking the computer out of the equation. Language
Learning, 60(3), 534–572. https://doi.org/10.1111/j.1467-9922.2010.00566.x
Boulton, A., & Cobb, T. (2017).
Corpus use in language learning: A meta-analysis. Language Learning, 67(2),
348–393. https://doi.org/10.1111/lang.12224
Boulton, A., & Vyatkina,
N. (2021). Thirty years of data-driven learning: Taking stock and charting new
directions over time. Language Learning
& Technology, 25(3), 66–89. https://doi.org/10.64152/10125/73450
Boulton, A., & Vyatkina,
N. (2024). Expanding methodological approaches in DDL research. TESOL Quarterly, 58(3), 1193–1204. https://doi.org/10.1002/tesq.3269
Celik, I. (2023). Towards
Intelligent-TPACK: An empirical study on teachers’ professional knowledge to
ethically integrate artificial intelligence (AI)-based tools into education. Computers in Human Behavior, 138, Article 107468. https://doi.org/10.1016/j.chb.2022.107468
Chambers, A. (2019). Towards the corpus
revolution? Bridging the research–practice gap. Language Teaching, 52(4),
460–475. https://doi.org/10.1017/S0261444819000089
Charles, M. (2024). Benefits and
challenges of using do-it-yourself corpora for academic writing development. TESOL Quarterly, 58(3), 1205–1214. https://doi.org/10.1002/tesq.3276
Cheung, L., & Crosthwaite, P. (2025).
CorpusChat: Integrating corpus linguistics and
generative AI for academic writing development. Computer Assisted Language Learning. Advance online publication. https://doi.org/10.1080/09588221.2025.2506480
Crosthwaite, P. (2026). Is DDL dead yet?
Not so fast! Revisiting earlier claims on the state of corpus linguistics, DDL,
generative AI, and Language for Specific Purposes. Ibérica, 51, 5–19. https://doi.org/10.17398/2340-2784.51.5
Crosthwaite, P., & Baisa, V. (2023).
Generative AI and the end of corpus-assisted data-driven learning? Not so fast!
Applied Corpus Linguistics, 3(3), Article 100066. https://doi.org/10.1016/j.acorp.2023.100066
Crosthwaite, P., & Baisa, V. (2024).
A user-friendly corpus tool for disciplinary data-driven learning: Introducing CorpusMate. International
Journal of Corpus Linguistics, 29(4),
595–610. https://doi.org/10.1075/ijcl.23056.cro
Crosthwaite, P.,
& Özer, M. (2026). Vibecoding as method: A
structured workflow for AI-assisted research instrument development in applied
linguistics. Research Methods in Applied
Linguistics, 5(3), Article
100356. https://doi.org/10.1016/j.rmal.2026.100356
Crosthwaite, P., Sun, S. A., & Jiang,
F. K. (2026). Prompt sensitivity in AI writing: A methodological caution for
corpus-based analysis of AI-generated language. The Modern Language Journal. Advance online publication. https://doi.org/10.1111/modl.70089
Curry, N., & McEnery, T. (2025).
Corpus linguistics for language teaching and learning: A research agenda. Language Teaching, 58(2), 232–251. https://doi.org/10.1017/S0261444824000430
Dong, J., & Wang, H. (2025).
Exploring the effectiveness of a DDL-GenAI teaching mode in students’ EFL
writing classroom. Computer Assisted
Language Learning. Advance online publication. https://doi.org/10.1080/09588221.2025.2577369
Elgendy, I. A., Dwivedi, Y. K.,
Al-Sharafi, M. A., Hosny, M., Helal, M. Y., Crick, T., & Al-Busaidi, A. S. (2026). Responsible vibe coding:
Architecture, opportunities, and research agenda. Journal of Computer Information Systems. Advance online
publication. https://doi.org/10.1080/08874417.2026.2621186
Frankenberg-Garcia,
A. (2014). The use of corpus examples for language comprehension and
production. ReCALL,
26(2), 128–146. https://doi.org/10.1017/S0958344014000093
Gries, S. T. (2008). Dispersions and
adjusted frequencies in corpora. International
Journal of Corpus Linguistics, 13(4),
403–437. https://doi.org/10.1075/ijcl.13.4.02gri
Han, Z. (2024). ChatGPT in and for second
language acquisition: A call for systematic research. Studies in Second Language Acquisition, 46(2), 301–306. https://doi.org/10.1017/S0272263124000111
Handley, Z. (2024). Has artificial
intelligence rendered language teaching obsolete? The Modern Language Journal, 108(2),
548–555. https://doi.org/10.1111/modl.12929
Kane, M. T. (2013). Validating the
interpretations and uses of test scores. Journal
of Educational Measurement, 50(1),
1–73. https://doi.org/10.1111/jedm.12000
Karpathy, A. [@karpathy]. (2025, February
2). There’s a new kind of coding I call
“vibe coding” [Post]. X. https://x.com/karpathy/status/1886192184808149383
Kern, R. (2021). Twenty-five years of
digital literacies in CALL. Language
Learning & Technology, 25(3),
132–150. https://doi.org/10.64152/10125/73453
Kern, R. (2024). Twenty-first century
technologies and language education: Charting a path forward. The Modern Language Journal, 108(2), 515–533. https://doi.org/10.1111/modl.12924
Kilgarriff, A. (2005). Language is never, ever,
ever, random. Corpus Linguistics and
Linguistic Theory, 1(2), 263–276.
https://doi.org/10.1515/cllt.2005.1.2.263
Kohnke, L., Moorhouse, B. L., & Zou,
D. (2023). ChatGPT for language teaching and learning. RELC Journal, 54(2),
537–550. https://doi.org/10.1177/00336882231162868
Lu, X. (2010). Automatic analysis of
syntactic complexity in second language writing. International Journal of Corpus Linguistics, 15(4), 474–496. https://doi.org/10.1075/ijcl.15.4.02lu
Lu, X. (2012). The relationship of
lexical richness to the quality of ESL learners’ oral narratives. The Modern Language Journal, 96(2), 190–208. https://doi.org/10.1111/j.1540-4781.2011.01232.x
Ma, Q., Yan, J., Crosthwaite, P., &
Lee, C.-K. J. (2026). TPACK for corpus technology: TESOL student teachers’
growth and self-efficacy in independent language learning and teaching. ReCALL, 38(1), 94–110. https://doi.org/10.1017/S0958344025100347
Ma, Q., Yuan, R. E., Cheung, L. M. E.,
& Yang, J. (2024). Teacher paths for developing corpus-based language
pedagogy: A case study. Computer Assisted
Language Learning, 37(3),
461–492. https://doi.org/10.1080/09588221.2022.2040537
Marsden, E., Morgan-Short, K., Thompson,
S., & Abugaber, D. (2018). Replication in second
language research: Narrative and systematic reviews and recommendations for the
field. Language Learning, 68(2), 321–391. https://doi.org/10.1111/lang.12286
McEnery, T., & Hardie, A. (2012). Corpus linguistics: Method, theory and
practice. Cambridge University Press.
Messeri, L., & Crockett, M. J. (2024).
Artificial intelligence and illusions of understanding in scientific research. Nature, 627, 49–58. https://doi.org/10.1038/s41586-024-07146-0
Mizumoto, A. (2023). Data-driven learning
meets generative AI: Introducing the framework of metacognitive resource use. Applied Corpus Linguistics, 3(3), Article 100074. https://doi.org/10.1016/j.acorp.2023.100074
Mizumoto, A., & Eguchi, M. (2023).
Exploring the potential of using an AI language model for automated essay
scoring. Research Methods in Applied
Linguistics, 2(2), Article
100050. https://doi.org/10.1016/j.rmal.2023.100050
Ng, D. T. K., Leung, J. K. L., Chu, S. K.
W., & Qiao, M. S. (2021). Conceptualizing AI literacy: An exploratory
review. Computers and Education:
Artificial Intelligence, 2,
Article 100041. https://doi.org/10.1016/j.caeai.2021.100041
O’Keeffe, A. (2021). Data-driven learning
– a call for a broader research gaze. Language
Teaching, 54(2), 259–272. https://doi.org/10.1017/S0261444820000245
O’Sullivan, Í. (2007). Enhancing a
process-oriented approach to literacy and language learning: The role of corpus
consultation literacy. ReCALL,
19(3), 269–286. https://doi.org/10.1017/S095834400700033X
Ouyang, S., Zhang, J. M., Harman, M.,
& Wang, M. (2025). An empirical study of the non-determinism of ChatGPT in
code generation. ACM Transactions on
Software Engineering and Methodology, 34(2), 1–28. https://doi.org/10.1145/3697010
Pérez-Paredes, P. (2022). A systematic
review of the uses and spread of corpora and data-driven learning in CALL
research during 2011–2015. Computer
Assisted Language Learning, 35(1–2),
36–61. https://doi.org/10.1080/09588221.2019.1667832
Pérez-Paredes, P. (2026). Corpus
linguistics and AI in the reconfiguration of language learning ecologies. Language Teaching. Advance online
publication. https://doi.org/10.1017/S0261444825101109
Pérez-Paredes, P., Curry, N., &
Aguado Jiménez, P. (2025). Integrating critical corpus and AI literacies in
applied linguistics: A mixed-methods study. Computer
Assisted Language Learning. Advance online publication. https://doi.org/10.1080/09588221.2025.2569351
Pérez-Paredes, P., Sánchez-Tornel, M.,
& Alcaraz Calero, J. M. (2012). Learners’ search patterns during
corpus-based focus-on-form activities: A study on hands-on concordancing.
International Journal of Corpus
Linguistics, 17(4), 482–515. https://doi.org/10.1075/ijcl.17.4.02par
Pérez-Paredes, P., Sánchez-Tornel, M.,
Alcaraz Calero, J. M., & Aguado Jiménez, P. (2011). Tracking learners’
actual uses of corpora: Guided vs non-guided corpus consultation. Computer Assisted Language Learning, 24(3), 233–253. https://doi.org/10.1080/09588221.2010.539978
Ren, Y., Wong,
W. L., Barrot, J., & Zhang, L. J. (2026). Learning styles, engagement and
anxiety in AI-mediated writing: A multimodal feedback study. International Journal of Applied
Linguistics. Advance online publication. https://doi.org/10.1111/ijal.70120
Risko, E. F., & Gilbert, S. J. (2016). Cognitive
offloading. Trends in Cognitive Sciences,
20(9), 676–688. https://doi.org/10.1016/j.tics.2016.07.002
Shumailov, I., Shumaylov, Z., Zhao, Y., Papernot,
N., Anderson, R., & Gal, Y. (2024). AI models collapse when trained on
recursively generated data. Nature, 631, 755–759. https://doi.org/10.1038/s41586-024-07566-y
Sinclair, J. (2005). Corpus and text: Basic principles. In
M. Wynne (Ed.), Developing linguistic
corpora: A guide to good practice (pp. 1–16). Oxbow Books.
Vaccaro, M., Almaatouq, A., & Malone, T. W. (2024). When
combinations of humans and AI are useful: A systematic review and
meta-analysis. Nature Human Behaviour, 8,
2293–2303. https://doi.org/10.1038/s41562-024-02024-1
Wang, Q., Deng, Y., & Liu, D. (2025). Corpus-driven and
AI-assisted learning in teaching business academic writing: Examining their
efficacies and the predicting/moderating effects of learner variables. Computer Assisted Language Learning.
Advance online publication. https://doi.org/10.1080/09588221.2025.2589739
Wong, W. L., & Fong, S. (2025). A new era with
artificial intelligence powered learner corpora for teaching and learning of
English as a second language in primary school education. Applied Language Sciences, 1, Article 250105. https://doi.org/10.65553/ALS.250105
Wong, W. L., Abuhassna, H.,
Barrot, J., & Zhang, L. J. (2026). Trends in corpus linguistics and
artificial intelligence language technology research: A bibliometric review. Asian-Pacific Journal of Second and Foreign
Language Education, 11, Article
35. https://doi.org/10.1186/s40862-026-00412-w
Zhang, N.,
Wong, W. L., & Barrot, J. (2025). Primary teachers’ acceptance and
sustained adoption of AI powered learner corpora for writing instruction
through TAM and ECM perspectives. Scientific
Reports, 15, Article 37399. https://doi.org/10.1038/s41598-025-21205-0
Author Bios
Wei Lun Wong
is a senior lecturer at Faculty of Education, Universiti Kebangsaan
Malaysia. His research interests include GenAI learner corpora, teacher
education and applied linguistics. His works appear at International Journal of Applied Linguistics, The Asia-Pacific
Education Research, Educational Review, SAGE Open and Scientific Reports.