AI text generators like ChatGPT are still in their early days. Still, large technology companies like Google, Microsoft, and Amazon are already using large language models to design new writing platforms and writing technologies. In most cases, human users will use AI to augment their writing processes rather than as a replacement for human writers. The challenge for users will be to make decisions about what AI can do effectively (summary and general description) and what may still depend on human agents (attribution, evaluation, and judgment).
Policy Options for Incorporating ChatGPT in Undergraduate Courses
The Provost’s office offers two recommendations for allowing the use of Chat GPT and other Large Language Models in their three recommended syllabus policy statements.
For instructors who wish to embrace ChatGPT
Artificial intelligence (AI) language models, such as ChatGPT, may be used for any assignment with appropriate citation. Examples of citing AI language models are available at: libguides.umn.edu/chatgpt [or provide an alternative reference appropriate for your class]. Students are responsible for fact-checking statements composed by AI language models.
For instructors who wish to allow limited usage of ChatGPT
Artificial intelligence (AI) language models, such as ChatGPT, may be used for [assignment types A, B & C] with appropriate citation, but not for [assignment types D, E & F]. If you are in doubt as to whether you are using AI language models appropriately in this course, I encourage you to discuss your situation with me. Examples of citing AI language models are available at: libguides.umn.edu/chatgpt [or provide an alternative reference appropriate for your class]. Students are responsible for fact-checking statements composed by AI language models.
Documentation guidelines for ChatGPT
If you allow ChatGPT in your courses, students should document their use. Different stylesheets have different expectations for how ChatGPT-generated language should be documented.
American Psychological Association Style
APA treats text generated by AI tools as personal communication, much like you might treat an email to an author. The in-text citation follows the author-date convention typical of APA.
ChatGPT offers the following description of the APA style: “The APA style refers to the formatting and citation guidelines established by the American Psychological Association (APA). It is commonly used in academic writing, particularly in the social sciences such as psychology, sociology, education, and other related fields” (OpenAI, 2023).
In the References section of the document, the entry would list ChatGPT as an author.
OpenAi. (2023). ChatGPT (July 20 version). http://chat.openai.com
Modern Language Association Style
MLA treats the user-generated prompt as the title of a document and does not credit ChatGPT (or other AI) as an author. The in-text citation and reference below follow the preferred conventions of MLA, which treats ChatGPT as a source without an author rather than an agent that generates text.
“MLA style, short for Modern Language Association style, is a commonly used citation and formatting style in academic writing, particularly in the fields of literature, arts, humanities, and some social sciences (“What is MLA Style?”).
On the Works Cited page, the entry would have no author listed.
“What is MLA Style” prompt. ChatGPT. 20 July version. Open AI. 26 July, 2023. chat.openai.com/chat.
In each example above, the ChatGPT-generated text is offered as a direct quotation set off by quotation marks. Preliminary agreements regarding AI regulation include efforts to build watermarks and other identifiers to help distinguish AI-generated text from human-generated writing. Indirect quotations, paraphrases, and summaries of AI outputs would be documented similarly.
Recommendations for Instructors
Enter your writing assignments into ChatGPT to see what can be generated from a Large Language Model
ChatGPT can accept queries up to 2000 characters and will attempt to generate a response to most prompts. The text generated will be based on statistical probabilities of matching responses and will follow the language parameters upon which it is trained. Pay careful attention to the elements where generated text is successful and where it is unsuccessful. Users may notice that the form of the in-text citation is correct (author, date) but that the content doesn’t refer to an actual source or author. Similarly, a ChatGPT-generated reference page may organize alphabetically but not follow APA conventions correctly (or consistently) for capitalization and typography.
Helping students recognize the limits of generative AI is vital. While digital detectors have difficulty distinguishing between AI- and Human-written prose, experienced human readers often identify the errors and hallucinations produced by generative AI. Novice readers are less likely to catch these errors and may, unfortunately, be more likely to be persuaded by the authoritative style produced by GPTs.
Teach students strategies for prompt design and effective narrowing
Initial outputs from ChatGPT are often exceedingly conventional and lack relevant details. However, by using a process of prompt refinement, users can generate text that comes closer to serving the purposes and audiences intended with its output. ChatGPT offers six basic strategies for improving results from initial queries. In addition, many technology companies are developing Application Program Interfaces (APIs) to assist ChatGPT with complex tasks like statistical analysis and complex mathematics. Such ‘bolt-on’ technologies can improve the accuracy of responses and limit the generation of “hallucinated” results.
Remind students of alternative or better resources
Students sometimes use ChatGPT as a virtual tutor or librarian to generate ideas or learn strategies for beginning an unfamiliar assignment. Remind students that their instructors and TAs are subject matter experts who may have tested and proven strategies for explaining materials and concepts in many ways. The information, analogies, summaries, and comparisons generated by actual experts rarely contain the kinds of errors and omissions common to LLMs like ChatGPT. Similarly, while ChatGPT can generate keywords to guide literature research, librarians and subject matter experts can help students to use library tools and interfaces, controlled vocabulary searches, Boolean operations, and limiters to generate more effective results with greater accuracy.