Should Researchers Report Which AI Model Was Used for Each Language in a Multilingual Study?

If Different Languages Were Processed With AI, Should the Model Used for Each Language Be Reported?

You are conducting a multilingual study involving English, Filipino, and Korean interview responses. Generative AI assists with translation, preliminary coding, or summarization.

Perhaps you use the same AI model for every language. Alternatively, you select different models because one performs better on Korean text while another appears more suitable for Filipino-English code-switching.

When writing the methodology, you consider stating that "generative AI was used to assist with multilingual data analysis."

Would that description provide enough information for readers to understand how the analysis was conducted? If different models processed different languages, could those choices have influenced the results?

Language-specific model reporting matters because AI systems may differ in their ability to interpret terminology, follow instructions, preserve cultural meaning, and classify research material. Without an identifiable account of which system processed which data, readers may struggle to evaluate those differences.

Report AI Models by Language When Their Use Could Affect the Research

Why Language-Specific AI Model Reporting Matters

AI Model Identity Is Part of the Research Method When AI Influences Analysis

Research methods should describe procedures sufficiently clearly for readers to understand how evidence was generated and analyzed.

When an AI system translates participant responses, assigns analytical categories, or produces information incorporated into findings, its contribution becomes relevant to the methodological record.

A general statement such as "AI was used" does not identify the system, its function, or the material it processed.

Researchers should distinguish between an AI product or service and its underlying model. A commercial interface may change models over time or provide different models through different configurations.

Where available, the research record should identify the model, relevant version information, service or deployment, date of use, and task performed.

If the exact model identifier is not disclosed, researchers should report the information actually available rather than infer a version.

Different Models May Perform Differently Across Languages

Generative AI systems do not necessarily demonstrate equivalent performance across languages.

A model that produces accurate English summaries may perform less reliably when interpreting another language, a regional variety, or mixed-language material.

Differences may involve terminology, factual accuracy, cultural interpretation, instruction following, or omission of relevant details.

Consequently, researchers should not assume that using the same model across languages guarantees equivalent processing.

Similarly, using different models does not automatically establish that each language receives equally reliable treatment.

Language-specific reporting helps readers examine these possibilities in light of differences in generative AI performance across languages.

Model Choice Can Become a Methodological Confound

A confounding problem can arise when the language group and AI model are systematically linked.

Suppose all English responses are analyzed using Model A, while all Filipino responses are analyzed using Model B.

If the resulting categories differ, the researcher cannot immediately determine whether those differences reflect participants' experiences, model behavior, translation procedures, or other features of the study.

The model and language are confounded in the analytical workflow because each language is associated with a different processing system.

This does not mean that researchers must always use one model for every language. A different model may be more suitable for a particular task.

However, the selection should be justified, and researchers should consider whether the resulting differences could influence cross-language comparisons.

Report the Task Performed, Not Merely the Model Name

The same AI model can be used for substantially different activities.

For example, a researcher may use one model to translate interviews, another to summarize translated material, and a third to suggest preliminary codes.

Reporting only the model names does not explain how they contributed to the analysis.

Researchers should identify which model performed each task and which languages or materials were processed.

The distinction is particularly important when translation precedes analysis. An error introduced during translation may influence later coding even when a different model performs the coding.

A clear methodological account should allow readers to reconstruct the sequence of AI-assisted procedures.

Prompts and Configuration Can Matter as Much as Model Identity

Two researchers using the same model may obtain different results because they provide different instructions, contextual information, or output constraints.

For example, one prompt might request literal translation, while another asks for culturally adapted academic English.

These instructions can produce different representations of the same source material.

Researchers should preserve significant prompts, language-specific instructions, terminology glossaries, and relevant configuration settings where available.

For API-based workflows, this may include parameters affecting output generation and any system-level instructions controlled by the researchers.

However, recording identical settings does not guarantee identical results, particularly when the service or underlying model changes.

Model Versions and Access Dates Help Explain Changes Over Time

Commercial AI systems may be updated without researchers controlling every aspect of the underlying service.

A model used during initial analysis may behave differently when the same prompts are submitted months later.

Researchers should record the date or period of use and the available model version or identifier.

Where the provider offers stable model snapshots, these identifiers can improve traceability.

Where only a product name is available, researchers should acknowledge that limitation.

Model reporting supports reproducibility, but it does not guarantee that an external service can later reproduce the same output.

Language-Specific Prompts Should Be Documented

Researchers may adapt prompts to account for differences in language structure, terminology, or cultural context.

For example, a prompt used for Filipino interviews may instruct the model to preserve Taglish expressions, while a prompt for Korean interviews may specify how honorifics should be represented.

Such adaptations can be methodologically reasonable.

However, they also introduce differences in processing procedures.

Researchers should record consequential language-specific instructions and explain their purpose.

Without this information, readers may incorrectly attribute differences in AI output to the language or participants when they partly reflect different instructions.

Model Reporting Should Include Relevant Performance Evidence

Identifying a model does not establish that it is suitable for the research task.

If AI contributes substantially to multilingual analysis, researchers should consider reporting how its outputs were evaluated.

For structured classification, this may involve language-specific performance metrics and examination of errors.

For translation, bilingual reviewers may assess semantic fidelity, technical terminology, and culturally consequential expressions.

For interpretive qualitative analysis, researchers may examine whether AI-generated suggestions are supported by original-language evidence and consistent with the study's methodology.

The appropriate procedures are discussed in greater depth in validating AI-assisted multilingual analysis.

Researchers should distinguish demonstrated performance from assumptions based on a model's general reputation.

What Information Should a Language-Specific AI Record Contain?

A structured record can help researchers maintain consistent documentation across languages and analytical stages.

The record need not appear in full within the manuscript. Depending on the journal and study, detailed information may be retained internally or included in supplementary materials.

Should the Same Model Be Used for Every Language?

Using one model can simplify documentation and reduce variation in the processing system.

However, the model may perform unevenly across languages.

Using different models may improve task-specific performance when supported by appropriate evidence, but it complicates direct comparison because model effects and language effects may become intertwined.

Neither strategy is universally superior.

Researchers should evaluate the available alternatives, document the selection criteria, and consider whether cross-language comparisons remain defensible.

Where feasible, a common evaluation set or carefully designed sensitivity analysis may help investigate whether model choice materially affects the findings.

Reporting Is Particularly Important for Qualitative Coding

In multilingual qualitative research, AI-generated codes and candidate themes may reflect both the source-language material and the model's learned conceptual patterns.

If different models are used across languages, they may suggest different analytical categories even when participants describe related experiences.

Researchers should preserve the original-language evidence and distinguish AI-generated suggestions from researcher-developed interpretations.

The risk of misleading cross-language thematic comparisons makes model provenance particularly relevant when AI materially influences coding or theme development.

Code-Switching May Require Additional Reporting Detail

Some research material contains more than one language within a response.

For example, Filipino participants may combine Filipino and English when discussing academic experiences.

Researchers should not necessarily classify these responses as belonging exclusively to one language without considering the analytical purpose.

Where code-switching influences processing, the record should identify how the material was handled.

This may include whether the original mixed-language text was preserved, translated, normalized, or analyzed directly.

Model reporting should reflect the actual workflow rather than an oversimplified language label.

Published Disclosure and Internal Documentation Are Different

The full internal record may contain detailed prompts, model identifiers, translation versions, reviewer comments, and performance assessments.

The published methodology should communicate the information necessary to understand and evaluate the study.

Researchers should follow current journal, publisher, institutional, and funder requirements regarding AI use and disclosure.

There is no single reporting rule applicable to every discipline or type of AI assistance.

However, when AI materially affects the generation or interpretation of research evidence, describing its methodological role is important even when a publisher's general AI disclosure statement is brief.

Do Not Invent Missing Model Information

Some AI services do not disclose the exact model or snapshot used for a particular interaction.

Researchers should record the service name, date, interface information, and any version details that are actually available.

They should not infer an exact model identifier from the product name alone.

Where the model cannot be identified precisely, this limitation should be acknowledged if it affects reproducibility or methodological interpretation.

Accurate reporting of incomplete information is preferable to an apparently precise but unsupported description.

Reporting AI Models Used for English, Filipino, and Korean Interviews

The team might include a supplementary table identifying each language, model, task, and validation procedure.

Such reporting does not prove that the resulting analyses are equivalent. It makes the methodological differences visible and allows readers to evaluate the evidence supporting the comparisons.

Common Misconceptions About Reporting AI Models in Multilingual Studies

How Much Language-Specific AI Reporting Does Your Study Need?

The appropriate reporting detail depends on whether AI affects the research evidence or serves only a minor supporting function.

Researchers should prioritize transparency about procedures that could influence measurements, classifications, interpretations, or comparisons.

A simple language-by-model record can prevent important methodological details from being lost during a long research project.

Researchers should establish this documentation process before AI-assisted analysis begins rather than reconstructing it from memory during manuscript preparation.

Before Reporting AI Use in a Multilingual Research Study

Frequently Asked Questions About Reporting AI Models by Language

Language-Specific Model Reporting Makes Multilingual AI Research More Transparent

How to Cite This Guide

This guide is intended to be read, shared, and used in research, teaching, and academic work. If you draw on its ideas, explanations, or other content, please acknowledge the source by citing the guide. Doing so gives appropriate credit and helps your readers locate the original resource.