Teaching academic writing in the age of AI

Nearly all undergraduate students (92%) now report using artificial intelligence for assignments, most often for writing and editing, according to a survey from the Higher Education Policy Institute (Freeman, J., Student Generative AI Survey 2025, Feb. 2025opens in new window). That has posed a major challenge for instructors, especially those who teach academic writing or courses built around reading and discussing scholarly research.

How instructors handle AI—specifically, generative AI, or tools that create text, images, or other new content—runs the gamut from total bans to head-on adoption. Those decisions are complicated and personal—and must be made as technology and institutional policies continue to evolve. But a growing number of instructors believe it’s important to engage with AI and update assignments and classroom rules accordingly.

“Engaging doesn’t have to mean adopting. But it can mean having conversations with students, having a clear AI policy, and being intentional around this topic,” said Liz Delf, an AI teaching and learning fellow with the Center for Teaching and Learningopens in new window and a senior instructor in the School of Writing, Literature, and Film at Oregon State University.

To that end, Delf and others are exploring how to create assignments that reduce students’ reliance on AI, including by breaking up academic writing into smaller chunks. They’re also finding creative ways to explore AI’s benefits and limitations in the classroom. The goal is to prepare students for a professional world where they will need to decide when using AI is appropriate or helpful and navigate AI-use policies from employers and publishers alike.

“In the long run, we want to be thinking about what students are going to do outside of school, whether that’s psychology or something else,” said David Williamson Shaffer, PhD, a professor of learning analytics and learning sciences at the University of Wisconsin–Madison. “What tools will they need to use, and how can we make sure they are prepared to use them?”

AI has so rapidly and drastically changed the learning environment that some teachers have taken extreme measures, such as reverting to Scantrons (standardized multiple-choice forms) or blue books, to ensure that student work is original. Others have turned to oral discussions and exams as primary ways to evaluate students, but these approaches are not always possible for high-enrollment or online courses.

Even as some instructors move to sidestep or limit writing assignments in response to AI, “decades of research support writing as an important mode of learning,” Delf said. “My goal is to help students develop critical thinking, inquiry, and rhetorical judgment—and writing remains central to that, even in a world with generative AI.”

That truth has propelled Delf and others to find creative ways to reduce students’ reliance on AI, leaning into established best practices in writing pedagogy. First and foremost: Emphasize process over product by dividing large writing tasks into a series of smaller, scaffolded assignments. Instead of grading students on a polished, 20-page paper, grade their topic proposals, partial drafts, drafts, annotated bibliographies, and peer review.

“Breaking academic writing into chunks can help students feel less overwhelmed,” said Chris Hakala, PhD, director of the Center for Excellence in Teaching, Learning, and Scholarshipopens in new window and a professor of psychology at Springfield College in Massachusetts. “Part of the reason they turn to AI is that academic writing feels like such a daunting process.”

Reducing page length and extending timelines can also help, he said. Easing students’ stress and workload can help make them less likely to turn to AI. Decreasing the length and number of assignments can also help instructors manage the additional workload of multiple rounds of grading.

Providing feedback to students offers another opportunity. For example, giving students feedback face-to-face can help demonstrate in a tangible way that human input matters, Delf said. Hakala gives students both written comments and verbal feedback via video, and then asks them to respond with a written or video message explaining how they intend to address it. Even in larger courses, this approach can be scaled—such as giving a one-sentence response per student, he said.

Group assignments and multimedia assignments are also harder to offload to AI, Delf said, because the final product makes students’ authorship more visible. When students create a podcast or comic strip, they may use AI for brainstorming, but the finished work requires student-produced elements—such as original audio or hand-drawn illustrations—that cannot be outsourced to AI. She also recommends tapping into personal or local knowledge (having students tackle campus problems, for example, rather than national ones), as well as real-audience or real-world purposes.

“Is there a way to give the assignment real value through a natural audience, an opportunity for publication, or a document that students can use later in their careers?” Delf said. “Students are often more invested when the work feels purposeful” and may therefore be less likely to outsource work to AI.

Redesigning assignments around AI may require more up-front work from instructors, but “it also offers an opportunity to clarify the purpose of writing in higher education—and to adapt in ways that ultimately benefit students,” said Chelsea Lee, instructional lead for the APA Style team.

In a growing number of classrooms, teaching students about effective AI use is seen as an essential part of preparing students for the workforce.

During a December 2025 APA webinar on AI in academic writing, 442 participants reported allowing students to use AI in various ways—mostly around the edges of the writing process. A majority allowed students to use AI to generate ideas (72%) or get feedback on writing (59%), while about half allowed its use for research or outlining. Just 9% said students could use AI to generate text, for an introduction, for example.

Delf starts the conversation about AI when the term begins, giving students opportunities both during and outside class to test AI’s capabilities. For example, she asks students to search for scholarly sources using both Google Scholar and an AI tool of their choice, then compare the results and assess whether the AI-generated citations are authentic.

Shaffer teaches a course on AI where students are required to use AI to help complete assignments, such as writing summaries of scientific papers. He does not restrict how students use AI, but he does require them to submit conversation transcripts, as well as a reflection on how the tool helped—or hindered—their work.

“Nobody was actually relying solely on AI to write things and turn them in, because they very quickly realized that AI was not doing as good a job as they could,” Shaffer said.

His students did find other creative ways to use AI to support their work. Several non-native English speakers uploaded an article and discussed its main points with AI in their native language before writing their own English summary to submit. Other students asked AI to create a reading guide to help direct them toward key concepts in an article.

Overreliance on AI has been shown to weaken critical-thinking skills, so teaching students when not to use these tools is an important way to mitigate that risk (Gerlich, M., Societies, Vol. 15, No. 1, 2025opens in new window). In that vein, some instructors are experimenting with using AI as a questioning partner. Rather than asking AI to generate text, students prompt it to ask higher-order questions—probing assumptions, requesting counterarguments, or testing the strength of an interpretation. Learning science research suggests that generating and responding to such questions supports deeper comprehension and transfer of knowledge. When AI is positioned to challenge students’ thinking rather than replace it, it can function more like a Socratic tutor than a shortcut.

AI also offers opportunities to improve accessibility for students with disabilities. When used as a planning aid, language support, or organizational scaffold, AI can increase independence without replacing learning, much like other assistive technologies.

Hakala has found it helpful to have students complete an assignment with AI, then critique the result to see what works and what falls short. Hakala’s experience has shown that using AI for brainstorming can limit students’ creativity, but he allows students to ask AI for feedback after generating their own ideas.

“I’m a cognitive psychologist by training, so I think about AI in terms of learning outcomes,” Hakala said. “If it helps scaffold learning, I’m open to it. If it impedes or replaces learning, it has no role in the classroom.”

In 2025, 45% of institutions had adopted AI policies covering academic integrity, teaching, research, or some combination of these—up from just 20% the year before, according to Inside Higher Ed (2025 Survey of College and University Chief Academic Officersopens in new window; 2024 Survey of College and University Chief Academic Officers, Inside Higher Edopens in new window).

Such policies generally give broad guidelines and then defer to instructors to choose how to handle AI in their courses. Some institutions mandate the use of custom generative pretrained transformers (GPTs) for privacy reasons, while a growing number require instructors to include an AI policy in their syllabi.

“Classroom policies really run the gamut—a lot of instructors don’t allow for any use of AI, while others actively encourage their students to use it in different ways,” said Samantha Denneny, development manager for the APA Style team.

Given the range of approaches, Delf said it’s important to set clear expectations for students. On her syllabus, she includes a disclosure about her own AI use (for example: “I sometimes use AI to develop examples or course materials, but I do not use it for grading or feedback”), as well as initial guidelines for students.

Delf then works collaboratively with students during class to draft and commit to a classroom AI policy, which usually permits AI use for brainstorming, research, and feedback but not for generating sentences or paragraphs. She then posts the policy or adds it to the syllabus to give students a sense of ownership over their AI use.

“When we’re transparent and have those conversations up front, my hope is that students are more likely to buy in,” Delf said. “I can’t control how they use AI once they leave my classroom, but I want them thinking critically about their choices going forward.”

Some instructors even allow students to create their own personal AI policies. This may involve declining to use AI because of ethical concerns around privacy, bias, environmental burden, and more. In such cases, Delf offers alternative exercises that don’t require the use of AI.

For additional guidance on creating a classroom AI policy, see the AI Pedagogy Projectopens in new window and the Harvard Office of Undergraduate Education’s generative AI guidanceopens in new window, which includes sample syllabus language.

How to detect unauthorized AI use is another ongoing question. Nearly 8 in 10 U.S. faculty members said cheating has increased on their campus since AI tools became widely available, according to a November 2025 survey (Watson, C. E., & Rainie, L., The AI Challenge: Faculty Concerns About Generative AI in Higher Education, American Association of Colleges and Universities, 2025opens in new window). About half of those who attended APA’s academic writing webinar reported that their institution uses AI detection software, such as Turnitin or GPTZero, but just 1% said they completely trust these tools.

One reason for concern is that AI checkers have been shown to disproportionately flag original work from neurodivergent students and students writing in a second language (Pindell, N., “The Challenge of AI Checkers,” University of Nebraska–Lincoln Center for Transformative Teaching, 2024opens in new window). Research also suggests that instructors may overestimate their abilities to detect AI-generated student essays (Fleckenstein, J., et al., Computers and Education: Artificial Intelligence, Vol. 6, 2024opens in new window).

“Use of these checkers can lead to false accusations and can create an adversarial relationship with students,” Delf said. “It makes me feel like I’m policing instead of teaching.”

As AI detectors get better at flagging AI-generated content, “humanizer” tools have emerged to help users evade detection. Shaffer describes this as “an arms race that we’re eventually going to lose,” pointing to the long history of failed attempts to ban new technologies in education, including the so-called calculator wars in mathematics.

“What I tell my students is this: Whatever you turn in—whether it came from AI or not—you are responsible for it,” he said. “If the AI plagiarized, you plagiarized. If the AI made a mistake, you made a mistake. Using AI doesn’t change who gets blamed if there is a problem with the work.”