课堂中的生成式AI学习成果：实证证据述评

Executive summary This report presents a review of recent empirical evidence of generative AI (GenAI) impact on learningoutcomes in formal education. Its purpose is to provide educators with an overview of top concerns forensuring students’ learning gains when using LLM-based learning tools and concludes with research-derivedguidance for deciding when and how to use these tools in the classroom. The reportunfoldsas follows: Section 1 distinguishes between the needs of education and industry,where the benefits of LLMs were firstexplored, primarily for productivity gains.Educators’ priorities are different.Pedagogical concerns includeconsideration ofinequities in education, developing students’ critical thinking skills, and the potential for GenAIto inhibit social development. These concerns extend beyond technologists’ focus on mitigating technicalharms such as toxic content, bias, or accuracy in system outputs. Section 2 presents several key variablesthat affect learningwith GenAI:(1)AI literacy—understanding thecapabilities and limitations of an AI system—is a critical new variable for student success when using GenAI.(2)Educational equity is a variable where GenAI renders mixed experiences for marginalized groups. Studies showhow GenAI can be an effective resource for students with disabilities. In other contexts, it entrenches existingpatterns in academic performance of the weakest students and can exacerbate inequities for economicallymarginalized students.(3)GenAI can impact psychological and social conditions long recognized to facilitatelearning: self-efficacy, individual pace, and human connection. On self-efficacy, studies show that students canbe overconfident about their skill mastery when using GenAIand need help calibrating their mental model oflearning gains. For self-paced learning, GenAI introduces both efficiencies and pitfalls depending onlearningdomain andcontext, including whether AI tools are general purpose chatbots or scaffolded tutors.Studies alsohighlight GenAI impact on human connection, the foundation for developing higher-order skills of criticalthinking and creativity.GenAI’s on-demand availability but lack of social presencecan present opportunitiesand disadvantages, from providing a nonjudgmental environment for exploring topics to reducingcollaboration with peers in group projects. Yet, studies show that human tutors remain students’ preferredsource for trusted information. Section 3 examines how GenAI usage aligns with learning objectives in Bloom’s taxonomy.Basic cognitiveskills—Bloom’srememberingandunderstanding—are fundamental to success across academic domains.Studies show that there can be an overdependence and lack of engagement that result in impaired memory formation when using LLM chatbots. Development of higher-order thinking—analysis, reasoning,andcreativity—can be compromised if GenAI is used in ways that bypass the necessary struggle that is integral toacquiring skills. Studies illustrate how use of general-purpose GenAI tools such as ChatGPT, without scaffoldingor other pedagogical guardrails, can be detrimental to critical thinking. GenAI can also impact creativity.Students using GenAI for creative problem-solving can benefit from fast prototype iteration and greater projectcompleteness or detail but can also tend toward idea fixation and less originality and complexity in their work. Section 4 highlights how GenAI learning tools need greater pedagogical complexity.Up to now, state-of-the-art tools have been ChatGPT or similar, with prompt engineering for the model to assume an instructorrole or restrain itsoutputs. However, modified general-purpose chatbots cannot address the broad range ofpedagogical considerations involved in learning success. New types of experimental AI tutors with embeddedproven pedagogical strategies—for example, capable of detecting and effectively responding to a range ofstudent cognitive states—show promise. Consulting educators in the design is key for success of systems likethese that are on the horizon. A concluding synthesis of the empirical evidence offers four guidelines for integrating GenAI in learningenvironments: (1)Ensure student readiness—avoid introducing GenAI too early, before students masterdomain basics.(2)Teach AI literacy—build an awareness of GenAI capabilities and limitations so students canassess system outputs and learn domain-specific techniques for optimal results.(3)Use GenAI as a supplementto traditional learning methods—GenAI explanations and examples are capabilities thatstudents value, butteacher guidance with these explanations remains necessary.(4)Promote design interventions that fosterstudent engagement—limiting copy-paste functionality, supporting students’ metacognitive calibration toreduce overestimation of their learning progress, nudging learnerstowardscritical thinking, and evaluatingGenAI tools for provenengagement strategies. Authors Mihaela Vorvoreanu, PhDPrincipal Applied Scientist