Recent advances in artificial intelligence (AI) for writing (including ChatGPT) can quickly create coherent, cohesive prose and paragraphs on a seemingly limitless set of topics. The potential for abuse in academic integrity is clear, and our students could be using these tools already. There are similar AI tools for creating images, computer code, and many other domains. As AI represents a permanent addition to society and students’ tools, we need to permanently re-envision how we assign college writing and other projects. As such, FCTL has assembled this set of ideas to consider, falling into three categories.
Category 1: Neutralize the Software
- Hyper-customize your writing assignments. As is the case for contract (“for hire”) writing by professionals, academic misconduct can be curtailed or detected when the writing prompts are so specific to the course and the discussions within the class that an outsider, or an AI, would have little chance of producing an output that would earn a good grade. It can also help to specify heavy citations and a specific length, both of which are difficult for the AI to deliver convincingly.
- Break major assignments into smaller graded chunks. By scaffolding assignments into smaller bits, students are not only less likely to cheat, they are more likely to create stronger final products. An annotated bibliography might be an especially good idea to blunt the advantages of AI-generated writing.
- Prioritize writing in an authentic environment. While some students may lack experience with in-class writing (on paper), this high-touch method of collecting and grading writing offers the best chance to eliminate academic misconduct. In larger classes, grading may be kept more manageable by assigning shorter, but more frequent, in-class writing assignments.
- Collect at least one diagnostic of in-person writing to compare to submitted essays. A student whose formal essay writing style deviates significantly from their spontaneous, hand-written writing might warrant additional scrutiny.
- Assign writing with heavy citations. The AI software is more likely than a student to use citations that you (let alone a student) might never think to use, making them appear suspicious. Moreover, the chosen citations might poorly reflect what you had in mind with your hyper-customized writing prompt.
- Double check AI-generated citations. AI software is known to “hallucinate” citations (essentially: invent plausible, but fake examples)
- Assign longer essays. The AI software has trouble creating essays longer than five paragraphs. If prompted to continue, the result will look like bolted-on content to a high-school level essay.
- Be suspicious of the AI-specific pattern of writing. Many AI-written essays are exactly five paragraphs long. Most paragraphs are exactly three sentences long. Most paragraph transitions (and often most sentences in general) begin with a one or two word transition, followed by a comma.
- If feasible, assign a writing prompt that requires information after 2021. ChatGPT only includes information up to 2021, so anything from 2022 and beyond will stymie the software.
- Preview your writing prompt on the AI platform yourself. The type of prose produced by ChatGPT is remarkably cohesive, but the style can be recognized over time. Certain markers, like the flat topic sentences that begin most paragraphs, can help identify the prose as machine-generated. The rhetorical level of the prose can also be a marker—for many topics, the produced essay is superficial and can be characterized more as summary than analysis. It can also be helpful to know what your students might be seeing as output if they ask the AI a similar question, which can aid in detecting misconduct on student-submitted essays. However, do not rely on plagiarism detection software (e.g. Turnitin), since an identical prompt given to ChatGPT on two occasions will yield two unique essays.
- Specify your policies about AI writing on the syllabus. If you ban the use of AI writing, say so directly on the syllabus. Alternately, if you allow its use but want it acknowledged (cited, referenced, etc.), be explicit in the syllabus about your expectations. This level of detail is necessary because students might have other faculty that expect students to use AI writing, even if you do not.
- Use detection software that flags AI-generated writing. OpenAI created its own “AI Text Classifier“, and there are other such as ZeroGPT. Note that asking ChatGPT itself if it wrote a snippet of text will return a false negative: we have documented that it will claim it did not write text that it generated for someone else (note: it WILL recognize text it generated for your own account, however). Students may also use a paraphrasing website such as QuillBot to further scramble the words and decrease the utility of the detector. Student may also ask the AI to maximize the burstiness and complexity of the output, which may decrease its chances of being detected.
- Double-check citations to catch AI “hallucinations”. Since some AI generative text invents citations, it’s wise to spot-check student citation submissions via Google to verify they really exist and these are the correct authors.
- Require citation screenshots. If you’re assigning a research essay with a bibliography, require students to append a screenshot of the citation as listed on the university library’s webpage for each citation.
- Explore formats beyond traditional essays. In some cases, there may be other ways to communicate thinking, analysis, or evaluation without using a written essay. Examples might include interviews, asynchronous “face the camera” extemporaneous (rather than read) videos, mind maps, podcasts, vlogs, debates, or applications (both long and short) of interactive techniques.
Category 2: Teach Ethics, Integrity, and Career-Related Skills
- Discuss the ethical and career implications of AI-writing with your students. Early in the semester (or at least when assigning a writing prompt), have a frank discussion with your students about the existence of AI writing. Point out to them the surface-level ethical problem with mis-representing their work if they choose to attempt it, as well as the deeper problem of “cheating themselves” by entering the workforce without adequate preparation for writing skills, a quality that employers highly prize.
- Create and prioritize an honor code in your class. Submitting AI-created work as one’s own is, fundamentally, dishonest. As professionals, we consider it among our top priorities to graduate individuals of character who can perform admirably in their chosen discipline, all of which requires a set of core beliefs rooted in honor. Make this chain of logic explicit to students (repeatedly if necessary) in an effort to convince them to adopt a similar alignment toward personal honesty. A class-specific honor code can aid this effort, particularly if invoked or attested to when submitting major assignments and tests.
- Reduce course-related workload to disincentivize cheating. Many instances of student cheating, including the use of AI-writing, is borne out of desperation and a lack of time. Consider how realistic the workload you expect of students is.
Category 3: Lean into the Software’s Abilities
- Re-envision writing as editing/revising. Assign students to create an AI essay with a given prompt, and then heavily edit the AI output using Track Changes and margin comments. Such an assignment refocuses the work of writing away from composition and toward revision, which may be more common in an AI-rich future workplace. ChatGPT is spectacular at providing summaries, but they lack details and specifics, which could be what the students are tasked to do. Other examples include better connecting examples to claims, and revising overall paragraph structure in service of a larger argument.
- Re-envision writing as first and third stage human work, with AI performing the middle. Instead of asking students to generate the initials drafts (i.e., “writing as composition”), imagine the student work instead focusing on creating effective prompts for the AI, as well editing the AI output.
- Focus student learning on creative thesis writing by editing AI-created theses. The controlling statement for most AI essays can be characterized as summary in nature, rather than analytical. Students can be challenged to transform AI output into more creative, analytical theses.
- Refine editing skills via grading. Assign students to create an AI essay and grade it, providing specific feedback justifying each of the scores on the rubric. This assignment might be paired with asking students to create their own essay responding to the same prompt.
- Write rebuttals. Ask the AI to produce a custom output you’ve intentionally designed, then assign students to write a rebuttal of the AI output.
- Evaluate AI writing for bias. Because the software is only as good as information it finds and ingests (remember the principle of GIGO: garbage in, garbage out), it may well create prose that mimics structural bias and racism that is present in its source material.
- Teach information literacy through AI. Many students over-trust information they find on websites; use AI software to fuel a conversation about when to trust, when to verify, and when to use information found online.
- Give only open-book exams (especially online). Assume that students can and will use the Internet and any available AI to assist them.
- Assign essays, projects, and tests that aim for “application” and above in Bloom’s taxonomy. Since students can look up knowledge/information answers and facts, it’s better to avoid testing them on such domains, especially online.
- Teach debate and critical thinking skills. Ask the AI to produce a stance, then using the tools of your discipline evaluate and find flaws/holes in its position or statements.
- Overcome writer’s block. The AI output could provide a starting point for an essay outline, a thesis statement, or even ideas for paragraphs.
- Treat it like a Spellchecker. Ask your students to visit ChatGPT, type “suggest grammar and syntax fixes:” and then paste their pre-written essay to gain ideas before submission. (Note: for classes where writing ability is a main learning outcome, it might be advisable to require that students disclose any such assistance).
- Make the AI your teaching assistant. When preparing a course, ask the AI to explain why commonly-wrong answers are incorrect. Then, use the Canvas feedback options on quiz/homework questions to paste the AI output for each question.
- Teach sentence diagramming and parts of speech. Since AI can quickly generate text with variety in sentence structures, use the AI output to teach grammar and help students how better to construct sophisticated sentences.
- Engage creativity and multiple modes of representation to foster better recall. Studies show that student recall increases when they use words to describe a picture, or draw a picture to capture information in words. Using AI output as the base, ask students to create artwork (or performances) that capture the same essence.
- Teach AI input strategies as a discreet subject related to your field. AI-created content is sure to be a constant in the workplace of the future. Our alumni will need to be versed in crafting specific and sophisticated inputs to obtain best AI outputs.
- Create sample test questions to study for your test. Given appropriate prompts, AI can generate college-level multiple choice test questions on virtually any subject, and provide the right answer. Students can use such questions as modern-day flash cards and test practice.
Category 4: Use the software to make your teaching/faculty life easier
- Create grading rubics for major assignments. Give specifics about the assignment when asking the software to create a rubric in table format. Optionally, give it the desired sub-grades of the rubric.
- Write simple or mechanical correspondence for you. The AI is fairly good at writing recommendation letters and formulaic emails. The more specific the inputs are, the better the output is.
- Adjust, simplify, shorten, or enhance your formal writing. The software could be asked to shorten your grant proposal, or to suggest grammar and syntax fixes (particularly useful for non-native speakers of English!) In short, treat it like Spellchecker before you submit it.
- Summarize one-minute papers. If you ask students for feedback, or to prove they understand a concept via one-minute papers, you can submit these en masse and ask the AI to provide a summary.
Clearly, the strategies in these categories are not mutually exclusive and many can be used in combination with each other.
Be aware that there are multiple “large language model” software solutions similar to (and competing with) ChatGPT. Most of the large technology companies have their own. There are related AIs for drawing pictures, and many other domains.
- Bard – the text-generating AI created by Google using LaMDA technology
- BlenderBot – the text-generating AI created by Meta
- ChatGPT – the text-generating AI created by OpenAI
- Claude – the text-generating AI created by Anthropic (ex-employees of OpenAI)
- DALL-E – the image-generating AI created by OpenAI
- ErnieBot – the text-generating AI (in Mandarin) from Chinese search engine Baidu
- LaMDA (Language Model for Dialogue Applications) – a LLM trained specifically on dialogue, such as Google’s Bard
- LLM (Large Language Model) – a type of software / AI that accesses large databases it’s been trained on to predict the next logical word in a sentence, given the task/question it’s been given. Advanced models have excellent “perplexity” (plausibility in the word choice) and “burstiness” (variation of the sentences).
- Jasper – a for-pay text-generating AI aimed at businesses and blog posts
- OpenAI – the company that created ChatGPT and DALL-E
- Sydney – the name of the AI that supports Microsoft’s Bing search engine
Feel free to contact email@example.com with any further questions or ideas–we are actively seeking additional strategies to add to this list.