Schools are adopting artificial intelligence (AI) tools faster than researchers can evaluate them. Survey evidence suggests that student use of AI for school is already widespread. In a nationally representative 2025 RAND survey, 54% of middle and high school students reported using AI for school, and 53% of English language arts (ELA), math, and science teachers reported using AI in instructional tasks.1 New systems promise to provide tutoring, generate instructional materials, deliver automated feedback, assist teachers with planning and assessment, and help schools coordinate learning opportunities.
Despite rapid growth in AI-related education research, the causal evidence base remains small, especially for U.S. K–12 settings. A recent review identified more than 800 relevant papers, but only 20 high-quality causal studies on impacts for students or educators.2 The review found no high-quality causal studies conducted in U.S. K–12 schools on student-facing AI tools. Most of the studies examine short-term task performance rather than durable impacts on students' knowledge and skills.
This evidence gap creates a practical problem. Because new tools are emerging continuously and their capabilities are changing quickly, educators cannot rely only on product-specific studies to guide decisions. A tool evaluated rigorously today may look different by the time findings are published. Schools, therefore, need a framework for evaluating AI that can keep pace with technological change while remaining grounded in what research already shows about learning.
Research consistently demonstrates that cognitive, social, and emotional competencies influence both academic performance and long-term success.3 Drawing from five established social and emotional learning (SEL) frameworks informed by research across education, psychology, and labor economics, we identify the following competencies for thriving: academic knowledge and skills, higher order thinking skills, social skills, metacognition, self-regulation, adaptability, autonomy, motivation, interest, curiosity, belonging, interpersonal connection, self efficacy, a growth-oriented mindset and self-concept, and management of content specific anxiety, boredom, and frustration.
The empirical evidence on which experiences develop these competencies is well established and independent of the rapid cycles in which technological solutions develop. Key among these experiences are personalized instruction, complex problem-solving, meaningful collaborative projects, substantive discussion, and sustained relationships with educators.5 This chapter argues that the most useful way to evaluate AI in education is to ask whether AI tools expand students' access to those experiences. This question applies across technologies and connects adoption decisions to the educational goals schools are trying to achieve.
The chapter first reviews the evidence base and its limitations. It then examines student-facing AI tools, for which the latest causal evidence is clearest. Subsequent sections discuss design choices, educator-facing tools, system-level applications, and assessment. The chapter concludes with implications for practice and priorities for future research. The extensive list of priorities in the final section reflects the scarcity of empirical evidence on AI applications in K–12 settings.
