Handout for Citing AI Tools
According to the Guideline for Dealing with Artificial Intelligence by Berlin Universities Publishing (BerlinUP), the use of artificial intelligence (AI) or AI tools for BerlinUP publications requires a transparent description.
The following recommendations are based on current educational and academic policy discourses and known citation styles.
Not every use of digital tools or software applications in relation to scholarly writing and publishing falls under the term artificial intelligence as defined in the BerlinUP guideline. For example, the use of tools as aids does not necessarily require explicit mention.
Tools that use AI to check formal criteria (e.g. checking spelling, grammar and style, reference management programs and translation tools) are referred to as aids. However, these tools are not permitted to generate or create content themselves, but only to make formal changes. This does not affect the authors' responsibility for the formal and content-related correctness of the content published under their name.
According to the BerlinUP guideline, AI tools are therefore generative models for creating new content in the form of text, images, audio, video, software code or data sets. Currently known examples include text-generating chatbots such as ChatGPT, Copilot, Gemini or LLaMA as well as image-generating AI tools such as DALL-E, Stable Diffusion or Midjourney.
When adopting content generated by AI tools, the following information is required for a citation:
WHAT?
It is described which parts of a publication were generated with AI tools: e.g. text sections, figures, tables, data.
WITH WHAT?
It is described which AI tools were used in each case: e.g. provider, model, URL.
WHEN?
It is described at which time AI tools were used: e.g. date, version.
HOW?
It is described in which way AI tools were used: e.g. prompt, configuration, chat history.
However, if AI tools are used as an aid, i.e. without generating content intended for publication or as a rudimentary source of inspiration, a general reference is recommended, for example as a note or possibly in the methods section of the publication.
Generating content intended for publication
In principle, AI-generated content should be cited in the same way as other references in scholarly publications, i.e. in the form of a citation. This can be done in the text or in a footnote or endnote as well as within the reference list. However, it should be noted that AI-generated content is not a source in the scientific sense.
In addition, a description in the introduction or in the methods section may be appropriate. In special cases, more extensive chat histories with AI tools can also be documented as an appendix or supplement.
The publisher reserves the right to include a corresponding reference to the use of AI tools in the metadata of a publication.
Generating content not intended for publication
The use of AI tools should also be described transparently for content not intended for publication, such as for internal communication between the submitter and the publisher (e.g. abstract, synopsis). It is recommended that the information is also provided here in the form of citations.
Furthermore, the use of AI tools in the production process by editors should be described transparently. For journals, it is recommended that the information is provided on the journal's website. For books, the publisher reserves the right to provide the information on the publisher's website.
However, the use of AI tools in the review process is not permitted, as the confidentiality of submissions cannot be guaranteed. Furthermore, the reviewers are fully responsible for the content of their reviews. In particular, it should be noted that a text protected by copyright or licensing law (even in part) may not be used as an input in a commercial or freely accessible AI tool without the consent of the respective rights holder.
In principle, AI tools cannot legally function as authors.
Examples of relevant citation styles that have established rules for citing AI tools are presented below. If another standardized citation style not listed here is used, we recommend consulting the relevant rules and regulations.
In accordance with the recommendation of the American Psychological Association (APA), documentation of how and for what purpose a generative AI tool was used should be included in the methods section of the publication. In shorter texts (e.g. literature reports, essays), it is recommended that the documentation of the use of the tool is addressed in the introduction.
When reproduced in the text, both the prompt and the relevant text passages from the model's response should be reproduced.
Verbatim citation – example:
Regarding the question of whether the EU health statistics allow conclusions to be drawn about a possible link between sugar and fat intake and body weight, ChatGPT answers: „Yes, there is substantial evidence that both sugar and fat intakes are related to body weight. Numerous studies have examined the association between dietary behaviors, particularly sugar and fat consumption, and body weight.” (OpenAI, 2024)Reference – example:
OpenAI. (2024). ChatGPT (Version 3.5) [Large language model]. https://chat.openai.com/share/5ea85ee7-06a9-4760-ab5c-65ee3e8f54aaFor longer responses or prompt histories, the APA requires these to be attached as supplemental material to the paper, as the texts generated by AI tools are not always reproducible. Unless the chat history can be made accessible via an individual link (see ChatGPT example above).
According to the Chicago Manual of Style (CMoS), inclusion in the list of references is only recommended if the prompt and response of the AI tool can be made accessible via an individual link. If this is not the case, the AI tool's response should be included either in a footnote or endnote or in the text.
A. Footnote citation
Footnote citation – example:
When asked whether the EU health statistics allow conclusions to be drawn about a possible link between sugar and fat intake and body weight, ChatGPT responds in the affirmative, stating that there is a substantial link that has been proven by numerous studies. In addition, the language model points out that general dietary patterns and lifestyle factors must also be taken into account in order to understand and combat the obesity epidemic.³ --------------------- ³ Answer by ChatGPT to the prompt „Regarding EU health statistics and related research on dietary behavior, is there any evidence that sugar and fat intakes are related to body weight?”, OpenAI, March 7th, 2024.Reference – example:
ChatGPT, answer to the prompt „Regarding EU health statistics and related research on dietary behavior, is there any evidence that sugar and fat intakes are related to body weight?” OpenAI, March 7th, 2024. https://chat.openai.com/share/5ea85ee7-06a9-4760-ab5c-65ee3e8f54aaB. In-text citation
In-text citation – example:
When asked whether the EU health statistics allow conclusions to be drawn about a possible link between sugar and fat intake and body weight, ChatGPT responds in the affirmative, stating that there is a substantial link that has been proven by numerous studies. In addition, the language model points out that general dietary patterns and lifestyle factors must also be taken into account in order to understand and combat the obesity epidemic (ChatGPT, March 7th, 2024).Reference – example:
ChatGPT. 2024. Answer to the prompt „Regarding EU health statistics and related research on dietary behavior, is there any evidence that sugar and fat intakes are related to body weight?” OpenAI, March 7th https://chat.openai.com/share/5ea85ee7-06a9-4760-ab5c-65ee3e8f54aaModifications to the AI-generated text must be indicated in the same way as modifications to verbatim quotations.
The recommendation of the Modern Language Association (MLA) also refers to the paraphrased and verbatim use of AI-generated content. In both cases, the source is cited by means of a short title in the text.
Paraphrased citation – example:
When asked whether the EU health statistics allow conclusions to be drawn about a possible link between sugar and fat intake and body weight, ChatGPT answers in the affirmative and notes a substantial link that has been proven by numerous studies ("Regarding EU health statistics"). In addition, the language model points out that general dietary patterns and lifestyle factors must also be taken into account in order to understand and combat the obesity epidemic.Verbatim citation – example:
We asked various language models whether there were any indications of a connection between sugar and fat intake and body weight with regard to EU health statistics. The GPT 3.5 language model answered our question in the affirmative: “Yes, there is substantial evidence that both sugar and fat intakes are related to body weight. Numerous studies have examined the association between dietary behaviors, particularly sugar and fat consumption, and body weight.” (“Regarding EU health statistics”).Reference – example:
“Regarding EU health statistics and related research on dietary behavior, is there any evidence that sugar and fat intakes are related to body weight?” Prompt. ChatGPT, Version 3.5, OpenAI, 7. Mar. 2024, https://chat.openai.com/share/5ea85ee7-06a9-4760-ab5c-65ee3e8f54aaDue to dynamic developments, this handout will be regularly reviewed and updated as necessary.