• Overview
• Materials and methods
• Image captions
• Appendix
• Repository
• Additions to AI disclosure for local models

An important aspect of scientific integrity is the practice of documenting your methodology. This documentation allows other scientists to both reproduce your research and to judge its contents within the context of your applied practices (e.g. did it follow the academic standards?). For AI this practice also applies, though the manner in which the usage of AI is documented may need to be more extensive than you first think.

In this article, we distinguish four manners of using AI, for which the documentation requirements are later explained. Moreover, we give more information on the documentation of both local and online models.

In the Materials and Methods (M&M) section of your document it is convenient to add a paragraph per relevant tool at the end of your M&M to explain their usage.

In this example, Whisper was used to transcribe interviews. These interviews were manually corrected and anonymized, whereafter, NotebookLM was used to code the transcribed interviews.

When using AI-generated images, the source of the image should be included in the caption of the image, just like you would do for real images. However, for AI-generated images this information may require more extensive disclosure than expected.

• The tool used and the web location.
• The version of the model or the date of image generation.
• The (negative) prompt provided to the model.
• (If applicable) the prompt used by the model.

For the image below we have used Ideogram with Magic Prompt enabled to provide an extensive example of a complete caption.

Whether you use AI or not, you should add a statement at the end of your document (typically above the References) officially declaring your use of AI and stating you take full responsibility for the contents of the document written. When doing so, it is important to be concise about the name, developer and access location of the tool. Moreover, you should tell what specifically you used the tool for.

To allow reproducibility and give better insight into your data handling, you must allow your readers to access your research data, and give them insight into your data handling. Part of this may be the information and prompts you provide to AI tools, plus the generated outputs. Provide these outputs in a designated appendix, or a separate repository.

The more complete this repository is, the better. The minimum required information per AI model is the name, version, developer, access location and model configurations/settings of the tool. Still, we advise you to be reasonable in this documentation: Only copy significant changes and revisions, rather than all minor adjustments.

Conversations can often be exported to a Word or PDF-document and saved, though some models also do so via JSON-files (is allowed, but less easily readable) or via web-links. The usage of web-links to save your conversations is not recommended, as the link becomes inactive once the conversation is deleted on the original account.

When using AI models on your own device, you are often given a larger array of settings to play with. All model parameters relevant to allow reproducibility of your research settings should also be documented in the appendix. In these cases, be as complete as reasonably possible.