Case Study 2: When AI Assistance Backfired

This case study is composite — it draws from patterns observed across multiple students and reported by educators, rather than a single individual. The specific details are illustrative, but the underlying dynamic is well-documented in discussions among educators since the wide availability of AI writing assistants.

Maya was a third-year undergraduate studying international relations. She was a strong student: intellectually curious, conscientious, well-organized. She'd used AI writing assistants since her first year, initially with some guilt and increasingly with comfort.

Her process, by her third year, had become fairly fixed: she would read the assigned material, develop a general sense of what she wanted to argue, and then ask an AI assistant to help her develop the essay. She didn't ask it to write the essay for her — she'd say that clearly, and she believed it. What she actually did was describe her argument to the AI, ask it to help her develop the supporting structure, refine each paragraph's wording "with clearer academic language," and address counterarguments that the AI suggested she consider.

The result was essays that consistently received B+ to A- grades. Clear, well-organized, correctly referenced. Her professors found them competent if somewhat generic.

At no point in her first two years of using this approach did Maya see a problem. Her grades were fine. She was learning the subject matter.

The Diagnostic Moments

Two events, in her third year, revealed the problem she hadn't noticed.

First event: A final exam, no outside resources permitted.

The exam had three questions. The first two were factual and analytical, and Maya handled them adequately. The third question asked her to construct an argument about a current geopolitical situation, integrating perspectives from multiple theoretical frameworks she'd studied across the year.

She stared at the question for a long time. She had the theoretical frameworks in her memory. She knew the facts about the geopolitical situation. She understood what the question was asking.

But she had no idea how to build the argument.

She produced something — words appeared on paper — but when she read it afterward she recognized it as thin: the claims were asserted rather than argued, the evidence was assembled without genuine analytical scaffolding, the theoretical frameworks were named without being genuinely applied. She could tell it wasn't good, but she couldn't quite figure out how to make it good.

She received a C+ on that question, the lowest grade she'd received in her university career.

Second event: A seminar discussion.

Her seminar on international security required each student to defend a position on a current policy question, then respond in real time to challenges from classmates and the professor.

When Maya was challenged on a specific aspect of her position — asked to explain the logical connection between a premise and her conclusion — she found herself fumbling. She could sense that there was a connection, but she couldn't articulate what it was. She had argued the position in her essay (with AI assistance); she could not defend it in conversation.

The seminar professor took her aside afterward. Not unkindly: "Your essays are well-organized. I'm not sure they're actually your thinking. Can we talk?"

The Diagnosis

In conversation with her professor, Maya articulated something she'd been sensing but not fully acknowledging: she didn't know exactly what her thinking was anymore. The essays were good, but she wasn't sure they represented her analysis — or whether the analysis was the AI's, organized by her.

The professor offered a diagnosis that Maya found clarifying and difficult simultaneously: "Writing is thinking. When you wrote essays with significant AI assistance, you weren't doing the work that produces analytical thinking. You were reading and organizing and adjusting, but not generating the analysis from scratch. And that means you didn't develop the analytical capacity you thought you were developing."

The grades had been adequate. The learning had not been.

What Had and Hadn't Developed

Looking carefully at Maya's situation, a pattern emerges:

What did develop: Content knowledge (she genuinely knew the subject matter), reading comprehension (she processed the assigned material carefully), organizational understanding (she knew how academic arguments were structured), and — ironically — skill at using AI effectively as a writing collaborator.

What didn't develop: The ability to construct an original argument from a blank page, the analytical capacity to identify logical relationships between claims and evidence without assistance, and the fluency of moving from thought to written argument in real time.

The exam and seminar situation were designed specifically to test the second category — which they did. The AI had masked a skills deficit that only became visible when the AI wasn't available.

This is not uniquely an AI problem. Students have always found ways to produce adequate work without adequate learning: having someone else write essays, copying from other sources, producing work that looks like learning without being learning. What AI changes is that this shortcut is more accessible, more seamless, and harder to detect from the outside than previous shortcuts.

It's also more dangerous, because it feels like learning in a way that copying from a source or hiring a ghostwriter didn't. When Maya worked with the AI, she was reading, making decisions, revising, engaging with the material. It didn't feel like cheating. It didn't feel like shortcutting. It felt like assisted learning.

But it was shortcutting the specific cognitive work — constructing an argument from nothing, identifying logical connections, building analytical structures — that the assignments existed to develop.

The Recovery

Maya's professor designed a structured recovery for the second semester.

The core: essay-writing in stages, with each stage evaluated before the next was permitted.

Stage 1: An argument outline, written from memory without looking at notes. Just: what is your thesis? What are your three main supporting points? What is the best counterargument? No notes, no AI, no research materials.

Stage 2: A first draft, written from the outline. Notes permitted, AI not permitted.

Stage 3: Evidence integration and revision. Research permitted, AI may be used only for checking grammar and sentence-level clarity — not for developing arguments.

Stage 4: Final draft.

The first-semester experience with Stage 1 was humbling. Maya's outlines were thin. She had opinions but not arguments. Her "supporting points" often weren't logically related to her thesis.

That thinness was valuable information. It showed her where her analytical capacity actually was — and gave her a baseline to improve from.

By the end of the second semester, working consistently in this constrained format, her outlines had become dense and well-structured. Her arguments, when she started writing from them, were genuinely hers. Her exam performance recovered.

She also found — and this surprised her — that the essays she wrote in the constrained format felt more satisfying than the ones she'd produced with heavy AI assistance. "When I wrote with the AI, the essay was fine but I didn't feel much connection to it. Now the essay is mine. I can defend every claim in it, because I thought of every claim."

The Broader Lesson

Maya's story doesn't suggest that AI writing assistance is always or uniformly harmful. For professionals who already have strong analytical and writing skills, AI assistance can accelerate work without degrading skills. For learners who are developing those skills, AI that does the analytical work can prevent the development from happening.

The key question is: what is the purpose of this activity? If an assignment exists to develop a specific skill, using AI to perform that skill is defeating the purpose — regardless of the quality of the output.

The calibration is: use AI to support your thinking, not to replace it. Use AI after you've done your thinking, not before. Use AI to check, refine, and develop — not to generate the foundation you're supposed to be building yourself.

This isn't a rule against using AI. It's a rule for protecting the learning that AI can inadvertently displace.