The AI Paradox: Why Classroom Automation May Be Undermining Student Engagement and Performance

For years, the promise of generative artificial intelligence in education has been presented as a pedagogical panacea. Proponents envisioned a future where teachers, freed from the drudgery of lesson planning, grading, and administrative overhead, could spend more time fostering deep, meaningful connections with their students. AI, we were told, would be the ultimate teaching assistant: a tireless generator of creative materials, syllabi, and personalized feedback.

However, a groundbreaking, albeit sobering, new study suggests that the reality of AI in the classroom may be far more complex—and potentially detrimental—than the optimistic marketing suggests. The first randomized trial to test AI’s impact on real-world instruction indicates that when teachers gain access to AI-driven teaching assistants, it can actually lead to a decline in student motivation and, in certain cases, academic performance.

The Study: A Real-World Stress Test

The study, titled "Generative AI Can Harm Teaching," was released as a working paper in June 2025. Led by Alp Sungu, an assistant professor at the Wharton School at the University of Pennsylvania, the research team—which included renowned educational psychologist Angela Duckworth—sought to move beyond theoretical speculation and examine how AI functions in the high-stakes environment of a classroom.

The experiment was conducted during the spring of 2025 within a large private school chain in Turkey. The researchers followed a cohort of 193 teachers and over 2,800 middle and high school students. The methodology was straightforward: teachers were randomly assigned either to receive access to a customized, ChatGPT-based teaching assistant aligned with the Turkish national curriculum or to continue their instructional practices without this specific tool. Over a ten-week period, the participating teachers utilized the AI primarily for the creation of lecture notes, exam questions, and homework assignments.

The findings were striking. Students whose teachers were provided with the AI assistant reported that their classes were less interesting, less enjoyable, and ultimately less important than their peers in the control group. While the decline in "intrinsic motivation" was described as modest, it was statistically significant. More concerning was the impact on academic achievement: while overall averages remained steady, students under the tutelage of "weaker" instructors—those who had demonstrated lower performance metrics prior to the study—saw a marked decline in both standardized test scores and student confidence.

Chronology of a Failed Optimization

The trajectory of the study highlights a growing tension between technological "efficiency" and the human art of teaching.

• Pre-Experiment (Early 2025): Researchers established a baseline for teacher efficacy and student achievement. The study identified a subset of teachers who, by historical data, were already struggling to maintain high engagement or academic benchmarks.
• Implementation Phase (Spring 2025): The 10-week intervention began. Teachers in the experimental group were given full access to the AI tool. The researchers did not intervene in the daily classroom flow, allowing for an "organic" usage pattern.
• Mid-Experiment Observations: As teachers began to lean on the technology, the "personal voice" of the instruction appeared to dilute. The AI provided grammatically correct and curriculum-aligned material, but it lacked the idiosyncratic touch that characterizes effective human-led instruction.
• Post-Experiment Assessment (Late Spring 2025): Standardized final exams were administered. The data revealed the "crutch effect," where the tool served to replace the labor of teaching rather than augment the quality of the instructor’s output.

Analyzing the "Crutch" Phenomenon

Why would a tool designed to enhance efficiency result in lower student achievement? According to Professor Sungu, the answer lies in the distinction between "tool-assisted work" and "automated replacement."

"Teachers, just like students or coders, might be using AI as a crutch," Sungu explained in an interview regarding the findings. "Instead of doing the actual work, they’re using AI to delegate the task, and that lowers the quality of their teaching."

The hypothesis is that the most effective teachers treat AI output as a rough draft. They use the AI to generate a framework, but then invest significant time and energy into customizing the content to fit their unique teaching style, the specific needs of their students, and the nuanced context of their classroom. Conversely, less effective teachers appear to use AI as an "answer machine," accepting the generated output with minimal interaction or revision.

This creates a "uniformity trap." When lesson plans and materials are generated entirely by an algorithm, they lack the narrative thread and human emotion that make lessons memorable. "When you start using AI-generated material, you’re losing your personal voice," says Sungu. "It might be technically good enough, but it doesn’t really carry your own style. If everything is very uniform, it just becomes a bit more boring."

The Data: Where the Damage Occurs

The study utilized externally administered standardized exams to measure achievement, which effectively ruled out the possibility of grading bias. The most significant finding—that students of lower-performing teachers suffered the most—suggests that AI may widen the equity gap in education.

In environments where a teacher is already struggling, the temptation to "outsource" the cognitive load of lesson design to AI is higher. When that teacher uses an AI tool to replace their planning process, the resulting instruction becomes even more detached from student needs. The data indicated that these students not only saw their test scores dip but also experienced a sharper decline in their own sense of academic confidence.

It is important to note that the study did not observe a universal failure of AI. It was not a comparison between "AI vs. No AI," as teachers in the control group were free to use other digital tools or even different, non-customized AI platforms. The study effectively compared supported, curriculum-specific AI usage against independent pedagogical choice. If anything, the study suggests that the risks of unchecked AI reliance might be even higher than the data currently shows.

Implications for the Future of EdTech

The findings of the UPenn study serve as a necessary, if uncomfortable, correction to the industry’s hype cycle. As schools across the globe rush to integrate generative AI into their digital infrastructure, this research provides a roadmap of the pitfalls that await.

1. The Death of the "Time-Saver" Myth

Sungu himself is an active user of AI in his own university-level instruction, but he rejects the notion that it saves time. He notes that when he generates a poll or a game using AI, he spends an equal amount of time "calibrating" the output to ensure it is accurate and relevant. "It’s not a time saver," he asserts. The lesson for educators is that AI should be viewed as a catalyst for deeper thought, not a mechanism for clearing a to-do list.

2. The Need for Guardrails and Training

The study underscores that access alone is insufficient. If schools provide AI tools without accompanying teacher training, they are likely to see the exact outcomes recorded in the Turkish study. Professional development programs must shift their focus from how to use the tool to how to maintain human agency when using the tool.

3. Preserving the Human Element

Education is, at its core, a human-to-human relationship. If AI-generated materials are used as a replacement for a teacher’s unique pedagogical voice, the student-teacher connection is severed. Future interfaces must be designed to encourage teachers to insert their own examples, stories, and assessments, rather than simply accepting the "first draft" offered by a chatbot.

Conclusion: A Call for Cautious Innovation

Professor Sungu is careful to emphasize that the study should not be interpreted as a blanket condemnation of AI. "It would be a mistake to conclude that AI is terrible and will ruin education," he cautions. Instead, the research highlights a systemic challenge: the technology is currently being used as a replacement for the "difficult work" of teaching, rather than as a scaffold for better instruction.

As education systems continue to grapple with the transformative power of generative AI, the focus must shift from the convenience of automation to the preservation of human judgment. The "crutch" of AI is a tempting one, but the research suggests that by leaning on it, the education system risks losing the very qualities—creativity, nuance, and connection—that make teachers the most essential component of the learning process.

The era of AI in the classroom is no longer a future prospect; it is here. The question now is whether educators will use it to amplify their craft or allow it to hollow out the profession from within. For now, the evidence suggests that the answer depends entirely on how much of themselves teachers choose to keep in the loop.