Many of us find ourselves at the crossroads: do we ban or allow the use of Generative Artificial Intelligence tools? There are many AI tools now appearing for text, images, design, audio, music and video, even AI that help with literature reviews. Some are free, others require payment, so it can be hard to know the right thing to do.
How can we define AI use in terms of our assessment tasks? If we allow AI use, what do we suggest that students do? As educators, we need to provide guidance to students about when AI tools are permitted and how these tools can be used in the context of doing assessment with academic integrity. The approach you use should be guided by the context of the discipline and the requirements of each assessment task.
But first, a caveat: any opinions expressed within this post are that of the author and not necessarily representative of any official position by Macquarie University administration on these matters. [Update – see the advice on citing ChatGPT from MQ library here]. Now, let’s get on with exploring what we could do!
Spelling out the rules
It is a good idea to clarify the permitted and non-permitted use of generative AI (and other such tools) for each assessment task. This will help students understand the rules for each task, rather than leave them guessing, which can lead to trouble.
MQ Academic Literacies Unit have prepared a ‘Traffic Light’ guide to generative AI tool use along with an associated assessment checklist that has three levels:
- Not permitted. Any use would constitute a breach of academic integrity. Given that AI will become increasingly integrated into many software systems and tools, prohibiting AI use is likely a short-term strategy for all except selected assessment tasks.
- Some use permitted. The scope of limited use is outlined for students while the majority of the work is still composed, written or made by the student. Any use of AI tools or content must be acknowledged or cited as appropriate. Any AI tool use beyond that specified would be considered a breach of academic integrity. Examples of limited use could include: Brainstorming (e.g. an AI tool can be used to brainstorm or rough-out ideas), Summarising material such as journal articles to aid understanding (but the generated summary itself can’t be used as part of the submission) or Editing (e.g. the student’s original work is edited for clarity by the AI tool). Careful thought will be required on the part of the assessment designer as to the impact of AI tool use on assurance of learning and how to make the boundaries of use clear to students.
- Full use permitted (with attribution). Adaptive use of generative AI tool output where content generated from an AI tool is edited, mixed, adapted and integrated into the student’s final submission – with attribution of the source.
Students should be reminded that they are responsible for what they submit. If they have doubts about what is permitted then it is in their best interests to check with their educators.
Be sure to clarify rules that apply to other supporting tools such as the use of grammar and spell checking tools (where relevant for language assessment) or the use of calculators, code repositories or other software as may be relevant to effective assessment of student performance on the learning outcomes of the task.
Be aware of AI limitations
Another important piece to developing AI literacy is in understanding the positives as well as the negatives of using generative AI tools. Monash University flags some concerns to share with students:
- Accuracy: the output of generative AI tools may be inaccurate or contain false or invented information. For example, ChatGPT is essentially a ‘next word guessing engine’ not a ‘truth machine’. It is designed to produce human-like text given the context of the user’s prompt, but the tool has no ability to check for accuracy. This has implications for the quality of what it produces and this can include the generation of fake academic references. Other AI tools may have similar limitations.
- Privacy: In general the input and output of the tool is recorded by the organisation offering the tool. The information may become public or shared with other organisations. For students this means that the university may gain access to their interactions at some future time. At present it is not advisable for MQ academics to submit student work to AI tools without permission from the student (unless the tool is covered by a licensing agreement with MQ).
- Intellectual property: Be sure to read the terms of service for each AI tool to check the licence terms that cover input and output of each tool you use. The legal status of AI generated content is being tested in the courts in the US and UK but the law is far from settled. The source material for the models used to drive many of these tools is in part based on copyrighted material and the output may yet be deemed as inadvertent plagiarism or a breach of copyright.
While there are reasons to be cautious, there are certainly permissible uses for AI tools. If we follow good academic practices we should be on stable ground.
Once it has been made clear to students as to the permitted use of generative AI tools for each assessment task, then we can say:
- Any *unauthorised* use of AI content generation tools in the preparation of assessment responses may constitute academic misconduct.
- Where use of AI tools for an assessment task is permitted then such use should be acknowledged. The contribution from the AI source should be cited using a relevant format or technique. This includes where direct quotes are used or where AI generated images are used as well as where ideas or concepts are sourced, remixed, or edited based on AI generated content. Where such is not cited then this may constitute academic misconduct (similar to plagiarism).
How could we acknowledge AI generated content?
There are no current established guidelines as to how to acknowledge generative AI tools or their outputs in academic works. The main standards bodies that define APA, Harvard etc are no doubt working on it. However we can consider the affordances of each tool, the context of the work, and what is of value to the reader as starting points.
- Is the output of the tool original (never seen before) or is it a copy or mashup? (The makers of ChatGPT, DALLE2 and Stable Diffusion claim to produce original output each time).
- Are the results unique each time a query is run? (Tools such as ChatGPT, DALLE2 and Stable Diffusion produce unique results each time).
- Is the output retrievable by a 3rd party reader? (Scribble Diffusion generates a unique URL for each output and this can be provided to others).
- Does the user have a record of the interaction? (or can it be kept or saved? – ChatGPT saves a log of each conversation to the user’s account).
- Does the tool version or the date have any baring on the information produced by the tool? Many tools are improving with each version. When a newly trained model is produced, the source material within it may also be updated to include more recent information. As of writing, the model for ChatGPT (GPT3.5) was created in 2021, any information after that is unavailable until the next iteration of the model is created. Other tools such as Perplexity.ai (and pending Bing and Google bots) perform live web search as part of constructing a response.
- What information is helpful to the reader? Is the query or ‘prompt’ or the ‘conversation’ log of value? For example, it may be useful for the reader to be able to run the query themselves as point of comparison or as part of research transparency.
- Are there requirements on the part of the person using the tool to maintain records of their use of tools or data sources? (as may be the case for a research project where keeping records of source material is part of ethical research practice).
Given the variation in the features of each tool, I would argue we need a nuanced approach that considers what will be of value to the reader. The following guidance is loosely based on Monash, JCU and UQ library resources. MQ library has also produced guidance for citing and critiquing generative AI tool output and the MQ library referencing guides have been updated too. The advice below is centred on the principle of being transparent about the use of AI tools and their outputs.
As is the practice of acknowledging other types of sources, it is suggested that where AI content is used without changes (this includes AI generated images and media) then direct quotes or ‘source’ attribution is used. Where ideas or concepts are adapted or remixed, with the author using their own words, then it is advisable to cite according to practices for ‘paraphrased’ of ‘adapted’ sources.
Where content is not retrievable the minimal option is to treat interactions with a generative AI tool as a form of personal communication where the AI tool name, link and date is used within the flow of the text. In this case there would be no reference list entry. For example, University of Queensland suggests to follow advice as per personal written communication.
Where generated content is retrievable, an option for the reference list entry is to treat the output as one would for information available on the web. Asking students to include live link (where possible) to each cited source in their reference list, such as to AI generated images will also help assessors check the source. For example:
Replicate (2023) Scribble Diffusion, “Tree house on top of a mountain…” Accessed 28 Feb. Generated output https://scribblediffusion.com/scribbles/txtcn3wtp5acdl6jzy5qus4a6y
It may be appropriate to do more to inform the reader, such as to include the model version number, date and the query used. Including the query will allow a reader to try the query themselves. In this case, a reference list entry could include the tool developer/author, the tool name, the version number, a link to the tool, the date and time the AI tool was used, and the text or content of the query used. For example:
Open AI (2023) ChatGPT (v3), Version 13 Feb 2023, Accessed 28 Feb. https://chat.openai.com Query “Which city in Australia is most often confused as being the national capital? Why is this so?”
Where more than the equivalent of say 30 words was used in the query or where the query comprised something other than text (e.g. a sketch, photo or other media) then attaching the query content as an appendix may be more appropriate.
No doubt in time the style guides produced by various bodies such has APA, Harvard etc will be updated to include various types AI generated sources.
Having students attach their AI conversations to the work as an appendix as one would do with a data set is reflective of the increasing practice of including links to data sets in journal articles. A conversation for this purpose includes the input and the output the user received. This can apply to text outputs such as from ChatGPT or visual outputs such as digital art from DALL-E or Scribble Diffusion. In the case of ChatGPT it enables a discussion over multiple connected queries and a transcript of the ‘discussion’ with the AI bot is recorded in the user’s account. While the chat log cannot be directly externally linked, a free browser add-on is available that enables a conversation ‘export’ feature to be added to the ChatGPT account menu. This exported conversation log could then be used for an attachment, appendix or a series of such logs could form a data set shared via an institutional repository or service such as FigShare. While other generative AI tools may not keep a record within an account, some may provide a URL to share output (such as Scribble Diffusion). In other cases, the onus will be on user to keep adequate records as is the custom for other forms of research. Providing advance notice to students of the expectations means they can take the necessary steps to preserve the record (e.g. by using screen imaging tools or saving files as they go).
We could also ask students to explain their working methods – e.g. “Show your working”. In this way the student is asked to surface their thinking and the process they used to arrive at the final product (See Lodge 2023 for an exploration). The student may be asked to provide a separate learning journal or learning log about the strategies used, search terms, sources, drafts, queries, outputs and how they adapted, remixed and integrated the information they obtained into their own assessment response. In this case the work involved in “showing the working” may need to become part of the assessment criteria given it would add work and skill sets to the task. See further ideas from MQ educators on how they are planning to have students surface their working including using paper log books, regular check-ins and in-class written reflections.
1) Be clear to students about if, and how generative AI tools (and other tools) can be used for each assessment task.
2) Where use is permitted, provide guidance to students on how you want them to acknowledge and evidence the contribution of AI tools in their submission.
As a last word, it looks like I was not the only one to be thinking along the above lines. I have since been pointed to this proposal by Hossain (2023) on citing AI sources who is a teacher, librarian and researcher.
Share your experience
We welcome your thoughts in the comments below with respect to permitting and acknowledging generative AI sources in student assessment and other academic works. You can also contribute your ideas by emailing firstname.lastname@example.org.
Join the conversation: Thursday, 30th March 2023 from 13:00-14:00. Register free for a MQ Community Roundtable: Gen AI tools – implications for learning, teaching and assessment at Macquarie.
Found an ‘AI generated’ academic integrity breach? See this advice on how to gather evidence and report it to MQ for investigation.
Acknowledgements: Banner image: Stable Diffusion “Two robots meeting over a book” (28 Feb 2023). M. Hillier. CC0 1.0 Universal Public Domain.