Using AI
Using AI
This section highlights the importance of leveraging AI as a supportive tool rather than a substitute for human intellect. If the intention of using AI is to augment human thinking, then it is likely an appropriate use of AI. If the intention of using AI is to replace human thinking, then it is likely an inappropriate use of AI.
See various examples of both appropriate and inappropriate uses of AI below. These examples illustrate how AI can be a powerful tool in tasks such as proofreading, translation, and data generation, while also cautioning against its limitations in areas requiring deep conceptual understanding and personal judgment.
Appropriate Usage Examples:
AI can support education in many ways, including proofreading for English as a Foreign Language (EFL) and English Language Learner (ELL) students, generating course materials like problems and rubrics (while avoiding uploading student work for privacy reasons), and creating example datasets for demonstrations. It is also effective for translation with better contextual accuracy than standard tools, for converting formats such as YAML to Markdown, and for tutoring when guided by carefully designed prompts. By tailoring prompts, AI can help learners practice, solve problems step by step, and receive feedback, making it a versatile tool for teaching and learning.
AI is especially useful for English Language Learners (ELL) for proofreading. While AI can sometimes replace your words and alter your style, it excels at correcting structure and grammar. It’s a good idea for students to use AI as a proofreader first.
AI can be used for problem creation (e.g., homework), lesson planning, or grading rubrics. Do not upload copyrighted or confidential course material to the AI though, since we have no guarantee that the AI will not incorporate the text into its training.
AI can create example data sets for various applications. For example, in a lecture about reading comma separated value data in Python, the instructor used a prompt “Give me a list of 50 names in comma separated value format with three columns: name, age, and zip code. The names should be traditional African American first names, the ages between 20 and 60 with a normal distribution, and the zip codes taken from 61801, 61820, 61822, 61853, and 61874.”
AI can assist in translating text between languages, and often it does a better job than Google Translate in terms of understanding context, idioms, and nuances. It can be particularly useful for translating documents,such as letters of recommendation written in other languages.
AI can convert formats, such as from YAML to Markdown tables. To do this it is helpful to give the LLM and example of the before and after. Here is an example used by one of the authors: “I have this list of dates and lectures:
Date,Lecture
2024-08-22,Course Introduction
2024-08-24,Recursion
….
I need a file of task entries. Each row should have three entries. The first row would look like this:
Date,Task
2024-08-20,Post Course Introduction Pre-lecture
2024-08,22,Post after-lecture activity for Course Introduction
2024-08,24,Post solution set for Course Introduction
Also, change the dates for Spring 2025 from January 21 to May 7th. The lectures occur on Tuesday and Thursday. Spring break is from March 15 to 23, no lectures should occur there.”
Given an appropriate prompt, an LLM can function well as a tutor. One method is to give it a prompt such as “Generate two paragraphs in Korean appropriate for an A2 level learner. We do not know participles yet, so do not use them. The story should be about a woman who is going to work and discovers that it snowed the night before, and now has to shovel the snow. Make the story end with a surprise. After the story, output a glossary of all the nouns and verbs used.” Similarly, it is usually very effective to copy in the text of a homework and give a prompt like “Here is a recent homework set. Generate a new one of the same difficulty level and similar content so we can get more practice.”
For tutoring, you can give it a prompt such as “Generate a problem involving solving an integral over a polynomial inside a transcendental function such as cosine. Help me to solve it on my own using the socratic method. Never give me the answer, but give me feedback on how I did and what is the next step to take.”
Inappropriate Usage Examples:
There are three main categories of inappropriate usage. First, the AI might give incorrect information that sounds right until inspected more closely. Second, anything you upload to the AI could be used to train it, and we don’t have data-safety guarantees that we can verify, which has implications for FERPA-protected information. Third, the strength of AI is it saves us from having to spend time thinking about things, but sometimes that is precisely what we are supposed to be doing. Finally, consider that the usefulness and appropriateness of the LLM output depends heavily on the quality of the prompt. “Garbage in, garbage out.” Here are some examples:
Writing Conclusions for Papers
Using an LLM to summarize a paper to get a high-level idea about its content can be helpful, for example, in deciding whether to read it., but copy-pasting that summary as a concluding paragraph for a paper you have written could be problematic. AI might produce text in perfect English, but it often produces over-generic text less useful for a reader.
AI struggles with tasks that require deep understanding beyond surface-level information. This will undoubtedly improve as LLMs become more advanced, but the human operator is ultimately responsible for checking the correctness of the output.
AI is not reliable for programming tasks that aren’t easily found on platforms like Stack Overflow.
Writing Recommendation Letters Without Context
If you just ask the AI to write a letter, it will be very generic and the reader may well realize that the AI wrote it. If you give the LLM examples of your own writing and ask it to mimic your style you will get more useful output. You also need to be careful about FERPA protections.
AI should not be used to grade student assignments, mainly from the data-safety and privacy concerns mentioned above. That said, there have been some teams that have trained a smaller LLM locally to grade specific assignments and have had good results.
AI is not suitable for writing peer reviews of academic manuscripts, for similar reasons as point 1 above. Successful peer review requires significantly more contextual knowledge than the AI is going to have from just the text of the paper.
While AI might provide correct answers, they are often not useful for learning because “being shown the answer” is not as useful as having to generate the answer yourself.
Students generally prefer their professors to respond personally rather than using AI-generated replies. Consider how you feel about “I hope this email finds you well”.
Applying to Jobs or Graduate School
When a recruiter has to look at hundreds of applications they will quickly develop the skill of recognizing the style that LLMs use, particularly if the applicant doesn’t rework the text to make it their own. This phenomenon is already being discussed in the press.