AI and qualitative analysis, f4analysis tips and tutorials

Citing ChatGPT and other AI content scientifically: A practical guide for researchers

23. July 2025 13 minutes reading time

Specify AI correctly - here's how

  • No direct quotation:
    LLM responses are not quotable like conventional text sources (book, academic essay, newspaper article) or personal communications.
  • Describe the LLM application in the method section and document the dialog in the appendix:
    Name the tool (e.g. ChatGPT-4o, OpenAI), the purpose (e.g. for brainstorming) and how you used it (e.g. central prompts, procedure). In some LLM-supported research methods (such as hybrid text interpretation; cf. Krähnke et al. 2025), the entire chat dialog (prompts and outputs) is listed in the appendix as part of the method documentation, but not directly cited as a source.
  • List in the bibliography:
    Reference the LLM itself correctly, e.g:
    OpenAI. (2023). ChatGPT (Version 4o) [Large language model]. https://chat.openai.com

Important: Always check specific requirements (university, publisher)! Details & reasons can be found below:

Introduction

Large language models such as ChatGPT, Claude or Gemini support literature research, text drafts or qualitative data analysis and AI supports automatic transcription. These tools raise a key question: how to document and cite generated content correctly in scientific papers, as the transparent documentation of AI use is now standard good scientific practice.

Since LLM responses are not permanently and identically retrievable like books or articles, they are not considered bibliographically verifiable sources. Nor should they be treated as “personal communication”. This is because the content of generative AI comes from machines and is neither stable nor replicable. For reasons of transparency, however, it must be documented where and how exactly the LLM was used in the research process: in the methods section, through complete prompts and outputs in the appendix and through a formal inclusion of the tool used in the bibliography.

This article summarizes the current recommendations of the common styles (APA, Harvard) and shows how you can make the use of LLM transparent in your research. You will also receive formulations and templates for your method section based on practical examples. We present this detailed guideline because – in view of the rapid development of AI tools and the scientific discussion about them – various approaches to citation exist, but some common recommendations do not, in our opinion, take sufficient account of aspects such as reproducibility or methodological embedding. This article therefore shows a way forward that takes particular account of the principles of good scientific practice.

Supplemented 26.05.25: Graduated transparency: Differentiated approaches to documenting AI usage

It is important to emphasize that the scientific debate on how to deal with AI use is still in flux. In addition to the specific question of how to cite AI content, there are also approaches that suggest a differentiation according to the type and scope of AI use. For example, the editors of the Zeitschrift für Evaluation (ZfEv) suggest in their current notes and an AI handout (as of December 2024) that different requirements should be placed on the documentation depending on whether AI was only used as a writing assistant, to create content, to support content or in empirical work. This can range from no necessary labeling (in the case of pure writing assistance) to classic source references to detailed descriptions in the methods section or a separate disclosure section. Regardless of the approach chosen, the primary objective remains to ensure traceability and scientific integrity.

APA citation style: focus on methodological transparency

The American Psychological Association (APA) recommends in its guideline (APA, 2023):

“If you’ve used ChatGPT or other AI tools in your research, describe how you used the tool in your Method section or in a comparable section of your paper. In your text, provide the prompt you used and then any portion of the relevant text that was generated in response.”
Source: https://apastyle.apa.org/blog/how-to-cite-chatgpt

In addition, Columbia College provides these implementation tips:

“APA recommends that you include an explanation of how you used ChatGPT, and an Appendix (or Appendices) containing the full text of your prompt(s) and output(s).”
Source: https://columbiacollege-ca.libguides.com/c.php?g=713274&p=5355771

In concrete terms, this means

Information in the methods section or in a comparable chapter:

    • Which AI tool was used (e.g. ChatGPT-4o)?

    • What for (e.g. brainstorming, text drafts, summaries)?

    • How (central prompt, example of a generated response)?

    • No direct citation of individual AI answers, as they are not available in a stable manner (i.e. they are NOT always answered in the same way).

Appendix:

  • Complete AI dialogs (prompt + response), (e.g. for hybrid interpretation).

Example of the reference entry according to APA

OpenAI. (2023). ChatGPT (Version 4.0) [Large language model]. https://chat.openai.com

Harvard citation style: focus on formal bibliographic coverage

The University College Dublin (UCD) recommends the Harvard citation style (UCD Library, 2025):

“When referencing AI-generated text in Harvard UCD, you should credit the creator of the AI tool as the author, using both an in-text citation and a reference list entry. If a shareable URL is available, please include it in the reference list entry. In the absence of a shareable URL, include the chat session with the AI tool as an appendix and refer readers to this appendix when citing the AI-generated text.” Source: https://libguides.ucd.ie/harvardstyle/harvardgenAI?utm_source=chatgpt.com

Harvard style requires that the developer of the AI system is cited as the author. This is done by in-text citation and entry in the bibliography. If the URL is missing, the complete documentation of the chat session is recommended as an attachment.

Example of the Harvard reference entry

OpenAI (2023) ChatGPT (Version 4.0) [Large language model]. Available at: https://chat.openai.com (Accessed: [date]).

Recommended practice in academic papers

Why you shouldn’t treat AI answers like quotes

Many researchers ask themselves the question: Can or should I simply insert the answer from ChatGPT into my work like a quote from a book? The clear scientific answer is: No, AI answers are not citable sources in the traditional sense. Here is a summary of the main reasons:

  1. No stability and reproducibility: The most important practical reason is that AI responses are not stable. Unlike a printed text, an AI such as ChatGPT does not provide identical answers to the same question (prompt). Even if you save an answer, other readers cannot reliably replicate or verify that specific answer. However, scientific evidence must be verifiable.
  2. Lack of scientific authorship: AI systems are tools, not authors. They lack human judgment, responsibility and the ability to stand up for their statements in scientific discourse. As the German Research Foundation (DFG, 2023) emphasizes, authorship requires agency and responsibility, which can only be found in humans (see also Buck, 2025, p. 87). AI does not have its own “opinion” or scientific position (Buck, 2025, p. 91).
  3. Unsuitable as a “personal message”: Sometimes it is suggested to treat AI responses as “personal communication”. That is not correct either. This category is intended for communication between people that is not publicly accessible but can be traced (e.g. emails, interviews). An AI answer does not fit here (see notes from the Harvard Library).
  4. No scientific contribution: Simply copying an AI answer does not constitute a research achievement in itself. Scientific work requires a critical examination of information, its classification and the development of one’s own conclusions. The power lies in the way you use AI as a tool and evaluate and integrate the results.

The right approach is:

  • Transparency in the methods section: Describe exactly which AI tool (with version, developer) you have used, for what purpose and in what way. Explain your approach, any key prompts and how you checked and used the results of the AI.
  • Referencing the tool: List the AI tool used (not the individual answer!) correctly in the bibliography. Depending on the citation style (e.g. according to Harvard/UCD), the developer (e.g. OpenAI) is then formally named in the author position. This serves to identify the tool, not to attribute authorship to the AI for the output.
  • Documentation in the appendix (optional/depending on method): For certain approaches, such as hybrid interpretation (Krähnke et al., 2025), in which the dialog process itself is part of the method, it can be useful to document the prompts and responses in the appendix in order to ensure the traceability of the process. However, this serves the purpose of method transparency, not the proof of a substantive statement by the AI.

To summarize:

Treat AI as a tool whose use you must document for transparency reasons, but not as a source that you can cite directly to support substantive arguments. As the citation standards for AI tools are still in a state of flux, you should also observe the current requirements of your university, department or publisher. It is important that your own scientific achievement remains clearly recognizable and that readers can follow your steps.

Sample formulations for the method section:

Example 1 (general template):

“In this work, [name of AI system] was used for [specific purpose, e.g. ‘identification of relevant literature’, ‘text summarization’, ‘data analysis’]. The complete dialog with the system, including all prompts used and responses generated, is documented in Appendix [X]. The use of AI was limited to [precise narrowing], while [aspects without AI support] were performed without AI support.”

Example 2 (specific application for hybrid interpretation with several LLMs):

“For the qualitative analysis of the interview data, the method of hybrid interpretation with several dialogically integrated Large Language Models (LLMs) was used, as described by Krähnke et al (2025). As part of this approach, three different generative language model versions (Anthropics Claude 3.7 Sonnet, openAI’s ChatGPT-4o and Google’s Gemini 2.5 pro) were integrated into an iterative interpretation process in order to generate multiple perspectives on the data material. The text interpretation took place in three phases: initial, elaborating and synthesizing interpretation. All prompts and the generated LLM outputs are fully documented in Appendix B. The selection of the text passages to be interpreted, the critical evaluation of the AI-generated interpretations and the final analytical summarization were carried out by me as a researcher without AI support in order to ensure a methodologically controlled and object-appropriate analysis.”

Special challenge: mediating AI systems

MAXQDA AI Assist, MAXQDA Tailwind, DocumetAI, Queludra and similar tools in the field of qualitative research sometimes integrate their own system prompts, which remain invisible to users. In addition, it is usually not possible to identify which language model (ChatGPT, Claude etc.) was used in which version (4o, 01 preview, sonnet 3.5 etc.) in the background – because: the manufacturers have not developed their own AI models, but use existing tools, but usually do not make this transparent down to the last detail.

It should therefore be made clear in the methods section that the complete system prompts are not accessible and that it may remain unclear which LLM generated the answers. This lack of transparency can have methodological consequences: The non-reproducibility of the AI analysis makes it difficult to trace, hidden system prompts could influence the analysis unnoticed and some models only have limited knowledge (in terms of time/content) from their respective training data. In order to make these limitations and biases transparent, it is necessary to document the specific work steps carried out with AI support. This includes a precise description of the software functions used (e.g. “text summarization”, “topic extraction”), the data entered and user instructions as well as the methods for critically reviewing the AI-generated results.

Example formulation

“MAXQDA AI Assist was used for the initial categorization as part of the qualitative content analysis. As the underlying system prompts and the language model used are not transparently disclosed, the analysis cannot be fully reproduced. To mitigate this limitation, we tested the AI proposals in a multi-stage discursive validation process (manual coding by two researchers, external discussion). Only those categories that passed were accepted. The AI support was limited to the exploratory phase; the final analysis and interpretation were carried out without AI.”

Citation of mediating systems

VERBI Software. (2024). MAXQDA AI Assist [Computer software]. https://www.maxqda.com/products/ai-assist

Also cite the use of automatic transcription correctly!

So far in this article, we have mainly dealt with the citation of Large Language Models (LLMs) such as ChatGPT. However, the principle of transparent documentation also applies to other AI-based tools used in the research process. A key example of this is the automatic transcription of audio or video data. Here, too, it is essential to clearly describe their use in the methods section to ensure scientific traceability. For example, if you use the automatic transcription from audiotranskription.de – either via browser via f4x.audiotranskription.de or within the f4 2024 software – this should be documented.

What belongs in the methods section?

Naming the tool: Specify the service or software used precisely (e.g. “AI-based transcription service from audiotranskription.de via f4x” or “automatic transcription function of the software f4 2024”), ideally with the date of use.

Purpose: Briefly describe the reason for use (e.g. “for the initial, time-saving transcription of expert interviews”).

Presentation of quality assurance: Since AI-based transcription can contain errors, the presentation of quality assurance is central. Have the automatically generated raw transcripts been fully checked against the recordings and manually corrected? According to which transcription rule system(e.g. Dresing & Pehl, 2023, content-semantic) was the final transcript revised?

Example formulation for the method section

“The AI-based transcription service of audiotranskription.de was used for the initial transcription of the audio recordings (accessed via [f4x / f4 2024] on [date]). The raw transcripts generated in this way using automatic speech recognition were then checked and corrected manually by [person/team] using the original recordings (e.g. in f4 2024 or Microsoft Word). The revision was carried out according to the rule system for content-semantic transcription by Dresing & Pehl (2023) to ensure methodological fit and consistency.”

Example of the transcription service/software entry in the bibliography

Variant f4 Use via browser only:
audiotranskription.de – Dr. Dresing & Pehl GmbH. (year of use). Automatic transcription service [Web service]. Retrieved on [date of retrieval] from https://f4x.audiotranskription.de

Variant Use via software f4:
audiotranskription.de – Dr. Dresing & Pehl GmbH. (2024). f4 2024 (Version X.X) [Computer software]. https://www.audiotranskription.de

Example of the rule system entry:
Dresing, T. & Pehl, T. (2023). Content-semantic transcription. In T. Dresing & T. Pehl, Praxisbuch Interview, Transkription & Analyse: Anleitungen und Regelsysteme für qualitativ Forschende (9th ed., p. [insert page number]). Dr. Dresing & Pehl GmbH.

Literature

You can quote this blog article as follows:

Dresing, T., Krähnke, U., & Pehl, T. (2025, April 4). Citing ChatGPT and other AI content scientifically: A practical guide for researchers. audiotranskription.de Blog. https://www.audiotranskription.de/ki-richtig-zitieren

The latest posts

10 minutes reading time
Prompts for your qualitative research project

AI can support you in various phases of your qualitative research project. The basis for good results are good prompts …

DSGVO-konforme, hybride Interpretation
15. May 2025 21 minutes reading time
GDPR-compliant, hybrid interpretation A workshop report on local interpretation with four LLMs on a MacBook Pro with Gemma 3, Qwen 3, Mistral 3.1 …
01. August 2024 4 minutes reading time
Documentary method – Proposal for the use of QDA software.

A simplified example, implemented with f4analyse. Introduction Here we show an example of the technical implementation of the evaluation steps …

Back to overview