03 · What You Need to Know
Why AI Tool and Model Identification Matters in Research Reporting
What Is the Difference Between an AI Tool and an AI Model?
An AI tool is the application, platform, or service through which a researcher interacts with artificial intelligence.
An AI model is the computational system underlying some or all of that functionality.
For example, ChatGPT is an application provided by OpenAI. The application may provide access to different models and capabilities, depending on the service configuration and time of use.
Similarly, an AI-powered research platform may incorporate a language model supplied by another organization or combine multiple systems within one workflow.
The distinction matters because naming an application does not always identify the exact model that produced a particular output.
AI tool or application
The software interface or service used by the researcher, such as ChatGPT or an AI-assisted research platform.
AI model or version
The underlying model identifier, release, or documented configuration responsible for generating or processing the relevant output.
Researchers should not treat these terms as interchangeable when the distinction is important to the methodology.
Why Might the Exact Model Version Matter?
Different AI models can produce different outputs from similar inputs.
They may vary in their ability to follow instructions, process documents, generate code, classify text, or interpret research materials.
For example, a researcher using a language model to classify open-ended survey responses may obtain different classifications after switching to another model.
Such differences could affect category frequencies and subsequent findings.
Identifying the model used helps readers understand the computational context in which the outputs were produced.
However, model identification does not establish that the results are accurate or that the procedure can be reproduced exactly.
What Do Major Publishing Authorities Require?
The International Committee of Medical Journal Editors (ICMJE) recommends disclosure of AI-assisted technologies used in submitted work and an explanation of how they were used.
The Committee on Publication Ethics (COPE) similarly emphasizes identifying AI tools and explaining their role in manuscript preparation or research activities.
Springer Nature and Elsevier maintain publisher-specific generative AI policies that address the reporting of AI assistance.
These authorities emphasize transparency, but their general policies should not be interpreted as imposing one identical model-version requirement across all journals and activities.
A journal may request the application name and purpose for manuscript-language assistance while requiring more detailed model information for an AI-based analytical method.
Researchers should consult the current instructions of the intended journal.
Does Naming ChatGPT Satisfy AI Disclosure Requirements?
It may satisfy the tool-identification component of a disclosure, but it does not necessarily satisfy the complete reporting requirement.
A statement identifying ChatGPT without explaining its role leaves readers uncertain whether it was used for drafting, translation, statistical code, or research interpretation.
Where the journal requires the model version, naming the application alone may also be insufficient.
For example, a methodological declaration might identify ChatGPT, the underlying model where known, the classification task, and how the outputs were verified.
For routine language assistance, the journal may require less detail.
The appropriate disclosure should reflect both the tool and the actual activity performed.
Does Every AI-Assisted Writing Activity Need a Model Version?
Not necessarily. The requirements depend on the publisher and the nature of the assistance.
Some publishers exempt defined routine AI-assisted copyediting, while others require declarations for broader forms of writing assistance.
Where disclosure is required, the application name and purpose may be sufficient under a particular journal's instructions.
For example, a statement that ChatGPT was used to improve language clarity may communicate the assistance without identifying the underlying model.
However, researchers should not assume that all publishers accept the same level of detail.
The question of disclosing AI-assisted grammar correction must be resolved separately from whether the model version needs to be reported.
When Is Model Identification Particularly Important?
Model information becomes more consequential when AI contributes directly to research procedures or produces outputs that are themselves analyzed.
| AI-Assisted Activity |
Relevance of Model Identification |
Reporting Consideration |
| Routine manuscript-language refinement |
May have limited methodological significance. |
Follow the journal's writing-assistance policy. |
| Statistical code generation |
May help document how analytical software was developed. |
Report the model where required, alongside the actual tested code. |
| Automated qualitative coding |
Different models may produce different classifications. |
Model identification may support evaluation of the coding procedure. |
| AI-assisted document screening |
Model behavior may influence study inclusion. |
Document the system and relevant screening configuration. |
| AI model performance evaluation |
The model is central to the research object. |
Precise identification is particularly important. |
| AI-generated research stimuli |
Model differences may affect experimental materials. |
Report relevant model and generation details. |
These distinctions concern methodological relevance rather than universal publisher rules.
What If the AI Provider Changes the Model?
Commercial AI services may update their models, interfaces, system instructions, or supporting infrastructure over time.
Consequently, the same application name may refer to different technical configurations at different dates.
For example, a researcher may use an AI service during data collection and return months later to find that the available models have changed.
Even when the model identifier remains the same, provider-side updates may influence outputs.
Researchers should therefore consider recording the date of use and relevant configuration information for consequential AI-assisted procedures.
The current version of an application should not be assumed to be the version used during an earlier research interaction.
Does Reporting a Model Version Make AI Research Reproducible?
Not by itself. Reproducibility may depend on several additional factors.
These can include the prompts, input materials, model settings, software environment, random variation, and provider-side changes.
Some generative AI services may not guarantee identical outputs even when researchers submit the same prompt to the same named model.
For example, a qualitative classification procedure might produce slightly different labels across repeated interactions.
Researchers should therefore distinguish identifying the system from reproducing its outputs.
Model identification supports methodological documentation, but it cannot substitute for preserving consequential outputs and verification records.
Should Researchers Record Model Settings?
Settings may be relevant when they influence how a model generates outputs.
Depending on the system, these may include temperature, output limits, sampling parameters, or other configurable options.
However, consumer applications may not expose every setting, and providers may change the available controls.
Researchers should record settings that were actually available and relevant to the procedure.
They should not invent values for parameters that were hidden or inaccessible.
Where AI is central to the research method, the limitations created by unavailable settings may need to be acknowledged.
Should the Date of AI Use Be Included?
The date may be particularly useful when the underlying service can change over time.
For example, an AI model evaluated in January may not behave identically to a service accessed under the same application name several months later.
Recording the date helps establish when the interaction occurred and which publicly documented configurations may have been available.
However, the date alone does not establish the exact model version.
Not every journal requires exact dates for routine manuscript assistance.
The requirements concerning recording AI-use dates should therefore be considered in relation to the methodological significance of the interaction.
What If the Exact Model Version Is Unknown?
Researchers should not guess.
An application may not display its underlying model identifier, or the researcher may not have preserved the information at the time of use.
Where possible, the researcher may consult saved interaction records, service documentation, account information, or research logs.
However, these sources may not establish the precise model responsible for an earlier output.
If the journal requires model identification and the information is unavailable, the researcher should report the known details accurately and seek clarification where necessary.
A statement acknowledging that the exact underlying model version was not available is preferable to an invented technical identifier.
Can Researchers Infer the Model Version From the Date?
Not reliably in every case.
A provider may offer several models simultaneously or use different configurations for different users and services.
The publicly announced release date of a model does not establish that every interaction occurring afterward used that model.
Likewise, an application may change its routing or supporting infrastructure without providing a complete public record of every configuration.
Researchers should therefore distinguish documented model information from assumptions based on release timelines.
Where a precise identifier cannot be substantiated, the limitation should be acknowledged.
Should Researchers Name the Provider as Well as the Tool?
Identifying the provider can improve clarity, particularly when tool names are ambiguous or when an application integrates models supplied by another organization.
For example, "ChatGPT (OpenAI)" identifies both the application and its provider.
However, a third-party research platform may use an external model provider, making the relationship more complex.
Where this distinction is methodologically relevant, researchers should describe the information available about the application and underlying system.
The disclosure should avoid implying that the application developer necessarily created the underlying model.
Should Researchers Cite the AI Model as a Reference?
Identifying AI assistance in a disclosure is different from citing an AI system as a source of scholarly evidence.
A journal may provide instructions for citing software, model documentation, or other research tools.
However, an AI-generated answer should not be treated as equivalent to a peer-reviewed source merely because the model is named.
Where the AI model itself is the subject of investigation, citing relevant technical documentation may be appropriate.
Researchers should follow the journal's software citation and AI disclosure requirements without confusing the two purposes.
What About Locally Hosted or Open-Weight Models?
Locally hosted models may provide researchers with greater control over certain aspects of the execution environment.
For example, researchers may be able to preserve the model checkpoint, inference software, settings, and code used during analysis.
However, model names alone may still be insufficient if different checkpoints, quantization formats, or configurations are available.
Where a local model materially affects the research, the relevant technical details should be documented according to methodological and journal requirements.
Researchers should also consider licensing and data-protection obligations.
The goal is to describe the actual system used rather than assume that local execution automatically guarantees reproducibility.
Should Researchers Preserve the Original Model Information?
Yes, where the information is relevant to consequential research activities.
Useful records may include the tool name, provider, model identifier, access date, settings, and task performed.
Researchers may also preserve prompts, outputs, and verification notes where appropriate.
These records can help reconstruct the workflow if the provider later changes the application or removes access to the model.
However, the appropriate documentation should remain proportionate to the research activity and applicable retention requirements.
The broader issue of recording AI model versions in research concerns preservation of methodological information as well as disclosure.
What Should a Disclosure With Model Information Look Like?
The statement should distinguish verified model information from unavailable details.
Illustrative Disclosure
AI Assistance With a Known Model Identifier
"The authors used ChatGPT (OpenAI; model identifier recorded in the research log) to generate preliminary Python code for data processing. The generated code was reviewed, tested, and revised before use in the final analysis. The executed scripts and relevant software environment were retained as part of the research records."
In an actual submission, the researcher should replace the descriptive model reference with the exact recorded identifier where required.
The statement should not imply that the model information was documented if no such record exists.
Illustrative Disclosure
AI Assistance When the Model Version Is Unavailable
"The authors used ChatGPT (OpenAI) to assist with language refinement during manuscript preparation. The exact underlying model version was not recorded. The authors reviewed the revisions and retained responsibility for the final manuscript."
Whether this level of detail is necessary or sufficient depends on the intended journal's requirements.
Does Model Disclosure Establish Human Accountability?
No. Naming the model does not demonstrate that its outputs were verified.
A researcher may accurately identify the AI system while incorporating unsupported interpretations or incorrect analytical code.
Conversely, careful verification does not automatically eliminate a journal's requirement to identify the tool.
Disclosure and verification serve different purposes.
The broader requirements concerning human accountability in AI-assisted authorship remain applicable regardless of how precisely the technology is identified.
What If the Journal Does Not Specify Whether Versions Are Required?
Researchers should review the current author instructions, publisher-level AI guidance, and any submission declarations.
If AI played a substantial methodological role, recording the available model information is prudent even when the journal's requirements are ambiguous.
Where uncertainty remains, the researcher may ask the editorial office whether the model identifier, access date, or additional technical information is expected.
The inquiry should describe the actual AI activity.
For example, a question about reporting a model used for automated qualitative coding is more informative than asking whether every ChatGPT interaction requires a version number.
