03 · What You Need to Know
How Generative AI Can Support Multilingual Qualitative Research
Multilingual Qualitative Analysis Is an Interpretive Process
Qualitative analysis seeks to understand meanings, experiences, practices, and processes represented in research material.
When participants use different languages, researchers must also consider how those meanings are expressed and transferred across linguistic contexts.
For example, a participant describing academic pressure in Filipino may use expressions involving family responsibility or social expectations that do not have exact English equivalents.
A literal translation may communicate the general statement while losing distinctions relevant to the research question.
Multilingual analysis therefore requires decisions about language, interpretation, coding, and representation. These decisions are methodological rather than merely technical.
Generative AI can contribute to this work, but the quality of its output depends on whether it preserves the meanings that the research seeks to investigate.
AI Can Suggest Preliminary Codes From Multilingual Responses
One possible application is using AI to identify candidate codes in interview transcripts, open-ended questionnaire responses, or other textual data.
A code is an analytical label assigned to a segment of material to represent something relevant to the research question.
For example, a participant might explain that unreliable internet connectivity makes it difficult to attend synchronous classes.
An AI system may suggest codes such as "connectivity barriers," "disrupted participation," or "technological constraints."
These suggestions can help researchers consider possible analytical directions. However, a plausible code is not automatically appropriate.
Researchers should examine whether the code accurately represents the participant's statement, whether it fits the study's analytical framework, and whether it introduces assumptions not supported by the source material.
In multilingual studies, the same code may also conceal different meanings across language groups.
Inductive and Deductive Coding Require Different AI Instructions
Qualitative studies use different approaches to coding.
In deductive coding, researchers apply categories informed by an existing framework or theoretical model. AI may assist by identifying passages that appear relevant to predefined categories.
In inductive coding, categories are developed through engagement with the data. AI may suggest candidate labels, but these should not determine the analytical structure before researchers have examined participants' accounts.
The distinction matters because AI systems may favor familiar categories from widely represented academic literature.
For instance, a model analyzing students' experiences may repeatedly propose conventional categories such as motivation, satisfaction, and engagement, even when participants describe experiences that do not fit those concepts neatly.
Researchers should decide whether AI-generated suggestions are consistent with the study's epistemological position and analytical approach.
AI Can Summarize Interviews but May Remove Analytical Detail
Interview summaries can help researchers become familiar with large collections of qualitative material.
Generative AI may identify the principal topics discussed, describe recurring concerns, or prepare concise summaries of individual interviews.
However, summarization involves selection and interpretation.
A model may omit hesitation, contradictory statements, culturally specific expressions, or seemingly minor details that later become important to the analysis.
For example, a participant may describe an online learning platform as generally useful while expressing discomfort about asking questions publicly. A summary that describes the experience as simply positive would conceal the tension between these observations.
Researchers should treat summaries as navigational aids rather than substitutes for the original transcripts.
When AI-generated summaries are used to support coding, consequential interpretations should be checked against the relevant source passages.
Translation Can Change the Data Before Coding Begins
One of the most consequential decisions in multilingual qualitative research is whether to translate participant responses before coding or analyze them initially in their original languages.
Translating everything into English may make it easier for a multilingual research team to work with a common dataset.
However, translation can also remove culturally meaningful expressions, alter emotional intensity, or substitute familiar concepts for locally specific experiences.
Alternatively, source-language coding may preserve more contextual meaning but require suitable linguistic expertise and procedures for comparing codes across languages.
There is no universal sequence appropriate for every study.
Researchers should consider whether multilingual data should be translated before or after coding in relation to their analytical framework, team expertise, and research objectives.
AI May Interpret Equivalent Concepts Differently Across Languages
A model may assign different codes to conceptually similar statements depending on the language in which they are expressed.
For example, expressions describing difficulty participating in class might be coded as "low confidence" in one language and "social pressure" in another.
These differences may reflect genuine contextual distinctions, but they may also arise from uneven model performance or language-specific associations.
Researchers should therefore evaluate whether AI-generated categories are applied consistently where consistency is analytically appropriate.
The problem of comparing qualitative themes across languages requires more than matching translated labels.
Researchers must establish whether apparently similar categories represent comparable experiences and whether apparent differences are supported by the source data.
Code-Switching Can Carry Meaning That AI Normalization Removes
Participants may switch languages within sentences or conversations.
For example, a Filipino university student might say, "I understand the lecture, pero nahihiya akong magtanong."
An English translation may communicate that the student understands the lecture but feels hesitant or embarrassed about asking questions.
However, the expression "nahihiya" may involve context-dependent interpersonal meanings that are not fully captured by a single English word.
Language switching itself may also be relevant when the study examines identity, classroom communication, or social relationships.
Researchers should evaluate AI interpretations of code-switched responses against the original material rather than assuming that fluent monolingual translations preserve every analytically relevant feature.
AI Can Help Compare Candidate Themes but Cannot Establish Their Meaning Independently
After researchers develop preliminary codes, AI may help organize them, identify overlapping labels, or suggest relationships among candidate themes.
For example, codes concerning unstable connectivity, limited device access, and difficulty attending synchronous sessions might be considered in relation to a broader pattern of technological barriers.
However, a theme is not simply a frequently occurring topic or a cluster of similar words.
In approaches such as reflexive thematic analysis, themes are developed through interpretive engagement with the data and the research question. They represent patterns of shared meaning rather than automatically discovered categories.
AI-generated groupings may therefore be useful as analytical prompts, but researchers must evaluate whether the proposed themes are coherent, distinctive, and adequately supported.
Different qualitative traditions may require different forms of analytical justification.
Human-AI Agreement Is Not Automatically Evidence of Qualitative Validity
Researchers may compare AI-generated codes with human coding to evaluate consistency.
Such comparisons can be useful when the study employs a structured coding framework with clearly defined categories.
However, agreement statistics must be interpreted according to the analytical approach.
In reflexive thematic analysis, coding is understood as an interpretive practice rather than a procedure requiring coders to produce identical labels. Treating agreement as the primary measure of quality may conflict with the method's assumptions.
Researchers should avoid applying inter-rater reliability metrics mechanically across all qualitative traditions.
Where agreement is relevant, independently established coding criteria, suitable evaluation samples, and examination of disagreements are important.
AI May Reproduce Dominant Theoretical Assumptions
Generative AI systems learn from patterns in available material, including widely represented scholarly language and conceptual frameworks.
When analyzing multilingual responses, a model may interpret unfamiliar experiences through categories that are prominent in English-language academic literature.
This can make locally specific knowledge less visible.
For example, an expression concerning family obligations might be categorized as individual motivation without sufficient attention to the relational context.
Researchers should examine whether AI-generated interpretations impose theoretical assumptions that are not supported by participants' accounts.
The possibility of English-language conceptual bias in non-English research deserves particular attention when the study aims to understand locally situated experiences.
Language-Specific Performance Should Be Evaluated Separately
A model that performs adequately on English-language interviews may be less reliable for another language or regional variety.
Researchers should not assume equivalent performance merely because the model can produce fluent responses in each language.
Evaluation may involve comparing AI-generated summaries or codes with independently developed interpretations by qualified reviewers.
Where possible, researchers should examine errors separately for each language and relevant variety.
They should also consider the uncertainty associated with small evaluation samples.
These procedures contribute to validating AI-assisted multilingual qualitative analysis.
Data Confidentiality Must Be Considered Before Uploading Transcripts
Qualitative transcripts may contain sensitive personal information, identifiable experiences, or confidential institutional details.
Removing names does not necessarily eliminate re-identification risks, particularly when contextual information makes participants recognizable.
Before uploading transcripts to an AI service, researchers should examine participant consent, ethics approval, institutional policies, contractual restrictions, and the service's data-handling arrangements.
Where external processing is not permitted, researchers may need an approved local environment, restricted-access system, or a different analytical procedure.
AI convenience does not override obligations to protect research participants.
Document the Analytical Role of AI
Researchers should identify which stages involved AI assistance and how its outputs influenced the analysis.
Useful records may include the model information available at the time, significant prompts, language-specific instructions, generated codes or summaries, human revisions, and decisions about accepting or rejecting interpretations.
Documentation should distinguish AI suggestions from final researcher-developed findings.
When translation is involved, researchers should also document consequential translation decisions.
These records help readers understand how multilingual material was interpreted and where AI may have influenced the analytical process.
