Professor Emily DeJeu argues that general-purpose AI tools like ChatGPT can hinder student learning by encouraging shortcuts, but custom-built AI systems designed for pedagogy are proving effective at amplifying skill-building and deep engagement.
When it comes to artificial intelligence (AI) and education, we are in a moment of dueling truths. Proponents of generative AI in education say that these large language models (LLMs) belong in classrooms because they give students personalized, just-in-time support. In a recent 60 Minutes feature, high school chemistry teacher Melissa Higgason explained that “Khanmigo,” the new AI-powered tutor from Khan Academy, supports student learning by being a helpful resource available 24/7, helping them work through problems step-by-step and answering every question they have right when they feel stuck.
Critics claim that these tools hinder learning by encouraging students to offload cognitive effort instead of wrestling with challenges that build real understanding. This is William A’s story: when he graduated from high school in Tennessee, his 3.4 GPA suggested success, but he couldn’t read. Diagnosed with dyslexia years earlier, he often failed in-class assessments like tests, but he leaned heavily on speech-to-text tools and later generative AI like ChatGPT to complete essays and projects outside of class, which allowed him to pass courses without developing literacy. His case, which ended in a court ruling that the district had denied him a meaningful education, highlights how detrimental new AI systems can be to students’ development when they’re allowed to do students’ thinking for them.
Who is right in this debate? The answer depends on the tools themselves. Publicly available tools like ChatGPT, which are built for ease and efficiency, have a different impact on learning than custom-built AI-powered tools specifically designed to support learning.
ChatGPT and Gemini are optimized for speed, fluency, and user satisfaction, making them excellent productivity assistants in workplace settings. However, students who are building the foundational knowledge they will need to succeed at work won’t benefit from speed and efficiency. Learning is an inherently slow, difficult process. Learning requires a student to hold new information in their working memory, focus sustained attention on it, and gradually integrate it into what they already know – a laborious, difficult process sometimes fraught with failure. But decades of educational research confirm that productive struggle and failure, as unpleasant as they may be, are crucial to learning.
New research shows that when students use publicly available LLMs for classroom tasks, they can bypass that productive struggle and seriously impair their learning. A team from MIT found that students who used LLMs to write essays worked efficiently but engaged in shallower mental processing and felt less ownership over their essays than students who worked without help from digital tools. Repeated reliance on LLMs created a “cognitive debt” for participants, as it allowed students to produce fluent writing without having to engage deeply and critically with source material. Similarly, a group of German researchers found that when students in their study used ChatGPT to research and write a recommendation paper, it alleviated some of the cognitive burden associated with long-form writing, but those students ultimately produced lower-quality arguments than students who used more traditional methods for conducting research and developing their ideas.
These trends are not exclusive to writing: a team of computer science researchers found that when LLMs generated full code solutions in response to students’ prompts, those students engaged less with the code, often choosing to simply run it and see the results instead of digging into its components. Students who received LLM-generated code also had the lowest average performance on manual coding tasks that required them to apply their knowledge, compared to students who did not outsource coding work to an LLM. These trends are also true for math education. Researchers from the University of Pennsylvania found that students who used ChatGPT for math help performed better on homework assignments than students without access to an AI tool; later, however, those students performed worse during an exam that did not allow LLM access, while students who had done their math homework independently performed better. Why the discrepancy? Researchers found that frequently, students asked ChatGPT for answers to homework problems instead of first trying to solve problems themselves, so they did not build durable math-related knowledge and problem-solving skills.
These studies show that, across different academic domains, LLMs can produce smooth, high-quality work that masks shallow engagement. Even worse, we are not always good at gauging our own learning, and we often avoid challenges in favor of easier tasks. It’s natural for students to conflate the fluency and ease that attends LLM use with a productive learning experience. A Harvard-based research team confirmed this when they compared how well users performed on basic tasks when an LLM simply gave them suggestions versus when the LLM forced them to think through a problem or set of options themselves and reason their way to an answer before seeing LLM-generated guidance. They found that users performed better when they had to grapple with a task first, using their own critical thinking skills; however, users liked that condition least and reported low levels of trust in their final answers. By contrast, they preferred simply receiving LLM-generated guidance and assumed their answers were more accurate in that condition, even though they weren’t.
If general-purpose LLMs encourage us to offload the difficult cognitive work of learning, and if we’re likely to mistake the fluency and ease they offer for real learning, do LLMs have any role to play in high-quality education? The answer is yes: AI-powered tools custom-built to support effective pedagogy can produce substantial learning gains. For example, a team of Harvard physicists and engineers built an AI-powered science tutor that doubled students’ learning gains compared to students who learned in an active-learning classroom. They designed their system according to research-based pedagogical principles, such as careful scaffolding (offering enough assistance to help, but not shortcut critical thinking) and iterative feedback, which helped students feel more engaged and motivated. A team of European researchers developed an AI-powered dashboard that helped students set goals, monitor progress, and reflect on their performance; they found that the dashboard improved students’ exam scores significantly, and students reported higher levels of motivation, engagement, and self-efficacy.
Custom-built AI tools can also make learning more fun. For instance, three Chinese researchers developed an LLM–driven science game where students interacted with an adaptive virtual guide that responded to their questions, gave real-time feedback, and adjusted challenges based on their progress; the result was higher achievement, greater immersion, and lower cognitive load compared to a conventional lesson plan. Similarly, three European researchers built an AI system that allowed students to learn programming concepts through lessons delivered by pop-culture characters such as Batman or Wednesday Addams, and they found that using the system doubled students’ study time and made their study experience more engaging.
Based on this emerging evidence, it’s clear that the future of business education lies not in enabling students to outsource their thinking to general-purpose AI systems but rather in developing custom tools that expand and enrich learning opportunities. Purpose-built systems that align with established learning science can help students engage more deeply with complex problems and practice decision-making in realistic contexts while at the same time giving them opportunities to build AI literacy, which is essential for navigating and shaping the data-driven workplaces of the future. By embedding custom AI-powered tools within well-designed pedagogical frameworks, educators can harness their power while keeping learning human-centered, ensuring that technology amplifies intellectual work without reducing human skill-building.
Tepper School Professor Zoey Jiang argues that generative AI can make less-skilled workers more productive, but experts or highly skilled workers may see productivity decline. However, strategies such as upskilling, tailored tool design, and continuous feedback can help companies better integrate AI.
Yan Huang reveals how Zillow's Zestimate, despite occasional algorithmic bias and varying accuracy across neighborhoods, benefits both buyers and sellers in underserved areas by reducing market uncertainty and improving transaction outcomes, even potentially lessening existing inequality compared to human-only systems.
Emily Diana explores algorithmic bias in machine learning and outlines three intervention stages: pre-processing, in-processing, and post-processing to mitigate algorithmic discrimination.