by Thomas Mejtoft, PhD

This guide is an ongoing attempt to discuss how to make references to generative AI in publication. Changes will continuously be made to this guide. Please send suggestion of changes of pages that could be used to develop this page.

Any sufficiently advanced technology is indistinguishable from magic.

Arthur C. Clarke^*

This guide should not be regarded as a policy on whether generative AI is allowed to be used as a tool within a certain course or program. Please refer to your university teacher for more information on that issue.

Generative AI technologies (e.g., ChatGPT for text and DALL·E for images) is a fairly new technology to the public and the use has many purposes. However, using these tools should be done with caution. It is important to understand that it is you as the author of an original piece that is responsible for what the generative AI tool produces if it is used. Even though the contribution of a generative AI tool can be substantial, many of the big publishers, e.g., Nature and ACM has clearly declared that a generative AI cannot be a co-author on a paper. Nevertheless, the use of these tools are in some cases permitted but should be transparent and fully disclosed.

Never mislead your readers about the use and involvement of AI in any part of the research and writing process!

It is important to understand that the copyright issues and the legal consequences of the use of AI generated content is still blurry and unclear. As an example, in the US, content created by generative AI cannot be considered for copyright. However, this is still a disputed question and might depend on the degree of human input (2023-08-18).

It is always the responsibility of the author (e.g., the student or researcher) of a paper/report/thesis to ensure that all information is correct, and no copyright violations has been included by an generative AI before the paper/report/thesis is published or submitted for examination. The AI cannot take responsibility. 

These guidelines are loosely based on the following sources:

ACM Policy on Authorship by ACM
How to cite ChatGPT by Timothy AcAdoo (APA Style Blog)
Educator considerations for ChatGPT by OpenAI
Citation, Documentation of Sources by The Chicago Manual of Style
How should I credit DALL·E in my work? by OpenAI
Who Ultimately Owns Content Generated By ChatGPT And Other AI Platforms? by Joe McKendrick (Forbes)

Hope you find this material useful!
If you are looking for other resources around writing to use, here is a page with resources and material. If you are looking for how to use and cite figures, screenshots, code etc. please refer to the following documents: How to use and cite figures from other sources, How to cite screenshots, References to secondary sources and review articles, Writing references to personal communication, Writing references to programming code, and Citing content created by generative AI. Regarding quotes and visualizing data, please read the following documents: Master quotes in writing and How to visualize your data in an understandable way.

How to cite AI generated text and images

Using generative AI tools as references. Since this the area of generative AI is fastmoving and changing, there are two philosophies around how to make a reference to generative AI – as an algorithm (most common and recommended) and as personal communication. The general rule for having something as a post in the reference list is that it is possible to retrieve or reproduce. Hence, e.g., published papers and literature are acceptable sources in the reference list, while personal communication is not. Personal communication is cited in-text but not included in the reference list.

Most common among publishers (and recommended here) is that output of generative AI should be regarded as the use of an algorithm to produce output that is used in the writing and not as personal communication. The generative AI should therefore be included in the reference list much like the use of programming code or any other algorithm. Consequently, the credit as author should be to the creator (individual or company) of the algorithm in the post in the reference list and an in-text citation should be done.

References examples (name, year) (APA Style):
OpenAI. (2024). ChatGPT (November 2024 release) [large language model]. Available at https://chatgpt.com/

Google. (2024). Gemini (1.5 Flash release) [Large Language Model]. Available at https://deepmind.google/technologies/gemini/

References, suggested examples [numerical]:
OpenAI. 2024. ChatGPT (November 2024 release) [Large Language Model]. https://chatgpt.com/

Google. 2024. Gemini (1.5 Flash release) [Large Language Model]. https://deepmind.google/technologies/gemini/

Since these systems are updated on a regular basis, it is of great importance to keep track of when certain information is generated and what version of the tool that is used. This information should be included in the reference. The best way to know which version and release that is currently in use is to ask the generative AI with e.g., the following prompt

Prompt:
What version and release is used in this chat

ChatGPT is rather specific on the current release, while e.g., Gemini is more vague. Include the information that is available to you in the reference.

OpenAI recommend the following use of BibTeX entry when refereeing to e.g., ChatGPT:

Suggested BibTeX entry for ChatGPT
@Misc{ChatGPT:2024,
  author = {OpenAI},
  title = {ChatGPT},
  version = {November 2024 release} [Large Language Model],
  year = {2024},
  month = {nov},
  day = {19},
  url = {https://chatgpt.com/},
}

@Misc{ChatGPT:2024,
  author = {OpenAI},
  title = {ChatGPT (November 2024 release) [Large Language Model]},
  year = {2024},
  month = {nov},
  day = {19},
  url = {https://chatgpt.com/},
}

Depending if the version id showin in the current reference format or not, it can be included in the title instead of using version. Look at the output to see of the version is showing or not.

Using generative AI tool as a general co-author. Many of the big publishers have stated that generative AI tools do not meet the criteria for being a co-author of a paper, e.g., Nature and ACM. Mostly based on its inability to take responsibility for the work. Nevertheless, the use of these tools might in some cases be permitted if it is fully disclosed in the paper. In a ACM publication the use should be stated in the Acknowledge section as follows:

Acknowledgement
ChatGPT was utilized to generate sections of this work, including text, tables, graphs, code, data, citations, etc.

If there are uncertainties about the use and whether an acknowledge should be included or not, go with caution and include a disclosure.

Using text, images of other original content created by generative AI tools. When using work created by generative AI it is important to disclose where this work come from. If a text is directly taken from a generative tool, it is important for the reader to understand where this text comes from and an in-text reference and a post in the reference list should be used. Likewise, if an image is generated by a tool an in-text reference and a post in the reference list should be used. In the case of images, e.g., OpenAI recommend that it is stated which tool that is used. For example: “This image was created with the assistance of DALL·E 2”. However, the most important thing is not to mislead your readers about the use and involvement of AI. Furthermore, OpenAI states that “You may remove the DALL·E signature/watermark in the bottom right corner if you wish, but you may not mislead others about the nature of the work. For example, you may not tell people that the work was entirely human generated or that the work is an unaltered photograph of a real event”.

Example with text:
Asking the question "What is the best structure of a design project?", text generated by ChatGPT indicate that a successful process is to "define goals, research user needs, brainstorm concepts, create designs, iterate based on feedback, collaborate with developers, test, finalize, document, evaluate" (OpenAI, 2023).

Reference:
OpenAI. (2023, August 16). ChatGPT (August 3 version) [large language model]. Available at https://chat.openai.com/

Examples of figure text to image created using generative AI:

Above are a couple of different examples of good ways of writing figure text to an AI generated image.

Reference:
OpenAI. (2023, August 16). DALL·E (version 2) [large language model]. Available at https://labs.openai.com/

Even though there are some cases where it is stated that no credit is needed for the use of a generative AI tool, a note acknowledging the use of the tool should always be made to avoid misleading the audience and to give credit to the creators of the tool. In other words, always use one of the examples above.

Using generative AI from a methodological point-of-view

From the standpoint of describing the method used in a paper, it is important to give a complete picture of the approach. Hence, if generative AI has been used in a paper for e.g., creative work, brainstorming etc., it is important that questions asked to the generative AI is clearly stated in the method. Furthermore, it might also be necessary to provide printouts of “conversations” with the generative AI.

^*Clarke’s Third Law, quoted from the essay Hazards of Prophecy: The Failure of Imagination in Profiles of the Future (Clarke, 1973, p. 21).
Clarke, A. C. (1973). Profiles of the Future: An Inquiry into the Limits of the Possible, Revised Edition. Harper & Row Publishers.

Cite this page as (APA style):

Mejtoft, T. (2024). Citing content created by generative AI. Notes on (scientific) writing, no 6. Retrieved from https://www.mejtoft.se/thomas/education/academic-writing/citing-content-created-by-generative-ai/

If you have any questions, praise, critique, or feed-forward, please do not hesitate to contact me.

(First published by Thomas Mejtoft: 2023-08-16; Last updated: 2024-11-25)