Key Takeaways
- Generative AI can produce essays quickly, making the final text a poor indicator of learning.
- The author shifts assessment from the polished product to the student’s visible writing process.
- Students must label AI‑generated passages (bold for verbatim, italics for revised work) and explain their revisions.
- Longer assignments include an AI appendix detailing prompts, raw output, and verification steps.
- Rubric now rewards argument quality, organization, lucidity, referencing, and the judgement exercised when working with AI‑generated material.
- Small‑group seminars allow instructors to gauge ownership of work through dialogue about drafts and AI use.
- Dishonesty remains possible, but removing penalties for AI use raises the cost of cheating relative to the reward.
- The approach can be scaled down to a single student‑led conversation per term or a short appendix for larger courses.
- Emphasizing process over product restores insight into how students develop arguments, even when AI contributes early drafts.
- The model demonstrates a practical, feedback‑rich response to AI that preserves academic integrity without relying on detection or bans.
Challenge of AI in Academic Writing
“Chatbots can now produce a passable essay in seconds.” This capability undermines the traditional assumption that a polished final essay reflects the learner’s effort, understanding, and judgment. As the author notes, “A polished final essay now tells us less than it used to about how the work was done.” Consequently, instructors face a dilemma: banning tools, reviving closed‑room exams, deploying AI detectors, or demanding proof of authorship each address symptoms but not the core issue—opacity of the writing process.
Vulnerability of the Author’s Course
The academic writing course the author teaches is especially exposed because it requires students to produce “structured, grammatical, formal prose.” From the submitted text alone, the instructor can no longer discern how the work was produced, whether claims have been verified, or how much judgement shaped the argument. This uncertainty existed prior to AI—students could hire tutors or ghostwriters—but generative models amplify the problem by delivering ready‑made content at scale.
Broadening Assessment to the Writing Process
In response, the author broadens what is assessed, moving away from treating the finished text as sufficient evidence of learning. “The student’s account of how they worked becomes the starting point for feedback.” By asking students to make more of the writing process visible, the instructor can evaluate judgement within that process rather than merely policing the final product. Drafting remains part of the work; a chatbot may supply an early version, which the student then transforms into an argument they understand and can defend.
Making AI Contributions Visible
For shorter tasks, students submit a marked‑up text. “Text copied verbatim from AI is in bold. AI output that has been meaningfully revised is in italics. Text written entirely by the student is left unmarked.” They also add brief comments explaining what they changed and why. This colour‑coding lets the instructor see, at the sentence level, how students describe their writing process: Did they blindly copy output? Did they spot weaknesses, refine unclear statements, edit organization, and ultimately take responsibility for the argument?
Handling Longer Assignments with an AI Appendix
When assignments become unwieldy for inline annotation, students submit an AI appendix containing the prompts used, the raw text generated by the LLM, and a note on how they verified and edited the material. The rule is simple: “Students may use AI or avoid it without penalty, but concealing its use is penalised.” This transparency preserves academic integrity while allowing legitimate AI assistance.
Adjusted Rubric Emphasising Judgement
The rubric now assesses the quality of argument, organisation, lucidity, and referencing—elements manageable because tasks are sequential. Students begin by summarising a selected paper, using it as a framing device for a literature review, then discuss the introduction and bibliography in a seminar, enabling the instructor to challenge or recommend further reading. Finally, the instructor verifies and cross‑checks new references and spots inappropriate citation of common texts in the final assignment. Part of the grade reflects the judgement students exercise when working with generated material: Can they flag fake references, misplaced confidence, or incoherent framing? Can they fix those problems and explain their decisions?
Seminars as a Diagnostic Tool
On assignments requiring AI use, students produce outputs, edit them, report on their process, and reflect on the experience in seminars. “Here, students bring their drafts and LLM outputs and work through a paragraph from inception to final form, explaining where AI entered the process and what has changed.” The discussion starts with argument and sources, then flows into editing style and wording. Through these dialogues, the instructor gauges whether students truly “own” the work, observing variations in how they engage with AI.
Student Variations and Responsibility
The approach reveals differences: some students copy and paste verbatim with minimal edits; others start with AI‑generated content but edit, shape, restructure, and rewrite to convey different ideas; a few barely engage with the technology. Any of these strategies can support learning, provided students take responsibility for the final text and show how they checked and developed it. While a student could lie and present AI‑generated text as their own, no educational design can eliminate dishonesty entirely. However, the author notes that in a small course with weekly seminars of five to seven students, asking them to explain their reasoning makes sustaining a false story difficult. Moreover, because there is no penalty for AI use, “the cost of dishonesty is higher than the reward, so it is simpler to be truthful.”
Scaling the Model for Larger Courses
The author acknowledges that this level of engagement isn’t feasible in all courses, especially large lectures in finance, public health, engineering, political theory, or marketing. Nevertheless, a scaled‑down version could work: a single student‑led conversation per term where learners choose a paragraph and explain its development and AI involvement, or, for larger assignments, a short narrative appendix. Such touchpoints preserve the core benefit—making the writing process visible—without demanding excessive instructor time.
Conclusion: Restoring Insight into Learning
These changes reframe the meaning of the work. A draft is now merely one step in the writing process; AI may contribute an early draft, but students continue to shape the argument as it develops. By the end, they should be able to justify what they’ve written and feel comfortable defending it. “GenAI has accelerated certain types of writing significantly and introduced new challenges to assessors. Silence and prohibition don’t reveal anything about the underlying process. A more productive response is to create opportunities to see and assess how students develop their writing when AI is involved.” The author’s model demonstrates that, by focusing on process, transparency, and reflective dialogue, educators can uphold academic integrity while harnessing AI’s potential as a learning aid.
https://www.timeshighereducation.com/campus/assessing-students-academic-writing-when-ai-can-produce-first-draft
