The research behind 11 teaching techniques, ready for your lesson plan.

It can go either way, and how the AI is set up seems to matter more than which tool it is. In a randomized field experiment with nearly 1,000 high school math students, unrestricted GPT-4 access raised practice scores but left students about 17% below a no-AI control group on a later exam, while a version built to give hints instead of answers largely avoided that harm (Bastani et al. 2025). In a randomized trial in one Harvard physics course, students learned more in less time from a purpose-built AI tutor than in an active learning class (Kestin et al. 2025). Both studies are single-site and short-term. See a guardrailed AI tutor for practice, Explain It Yourself, or every technique built around AI.

Not reliably. A test of 12 publicly available detection tools and two commercial systems (Weber-Wulff et al. 2023) concluded that none was accurate or reliable, and the errors leaned toward labeling human writing as AI-generated. Liang et al. 2023 found that detectors frequently misclassify writing by non-native English speakers as AI-generated. A sturdier approach builds evidence of the student's own thinking into the assignment itself. See process-based assessment.

Use three parts: the outcome students will reach, the research behind the method (what was found, in what setting, and how certain it is), and the evidence of learning you will collect. The justification guide has the template and a worked example, and every technique page ends with a one-sentence and a full version you can copy, with AMA references.

Four techniques here carry the strong label: retrieval practice, spaced practice, worked examples, and cooperative groups with individual accountability. For example, a meta-analysis of 222 classroom studies found a medium-sized achievement benefit from quizzing (Yang et al. 2021), and social interdependence theory, behind cooperative learning, draws on more than 1,200 studies (Johnson & Johnson 2009). The method page explains the labels.

Have them pull it out of memory, and come back to it later. In experiments by Roediger & Karpicke 2006, students who took recall tests remembered much more after two days and a week than students who reread, even though rereading made them feel more confident. Cepeda et al. 2008 studied more than 1,350 people and found that the best gap between study sessions grows with how long you need them to remember. See retrieval practice and spaced practice.

Peer instruction works with no special room: pose a conceptual question, have every student vote alone, let neighbors discuss, vote again, then explain. A ten-year report from introductory physics courses found gains in both conceptual understanding and problem solving (Crouch & Mazur 2001). More broadly, active learning raised exam and concept inventory scores by 0.47 standard deviations in a meta-analysis of 225 undergraduate STEM studies (Freeman et al. 2014). See peer instruction, or techniques for large in-person classes.

Link the outcomes and check each person. Structured cooperative learning rests on group members' outcomes depending on each other and on each member staying individually accountable (Johnson & Johnson 2009). In practice that means small groups, rotating roles, and an individual check whose results feed the group's result. In a meta-analysis of undergraduate STEM courses, small-group learning raised achievement and persistence (Springer et al. 1999). See cooperative groups.

Show complete solutions first, then fade the support. In five algebra experiments, worked examples cut processing time and errors (Sweller & Cooper 1985), and gradually fading examples into problems beat the usual example-then-problem pairing on near transfer (Renkl et al. 2002). The benefit shrinks as expertise grows (Kalyuga et al. 2003), which is why the fading matters. See worked examples.

Add video to your teaching instead of replacing it, keep it short, and put a question after each segment. A systematic review of randomized trials found larger gains when video was added to existing teaching (g = 0.80) than when it replaced teaching (g = 0.28) (Noetel et al. 2021). In an observational study of 6.9 million viewing sessions, median engagement was at most about 6 minutes (Guo et al. 2014). See instructional video design.

Keep it about the task, return it while it can still be used, and end with a next step. In a meta-analysis of 607 effect sizes, over one-third of feedback interventions decreased performance, especially when attention shifted from the task to the self (Kluger & DeNisi 1996). The often-quoted effect size of about .70 for formative assessment is not supported by a careful K-12 meta-analysis, which found about .20 (Kingston & Nash 2011). See feedback with a next step.

State the purpose, the task, and the criteria for success on the assignment itself (the TILT framework). A multi-institution pilot reported statistically significant gains of medium to large effect, especially for first-generation, low-income, and underrepresented students (Winkelmes et al. 2016). The evidence is early and comes from one project, so it is labeled emerging. See transparent assignments.

Two low-prep options: a 3-to-5-minute retrieval quiz at the start of class, or one assignment rewritten with a purpose, task, and criteria (transparent assignments). See everything that takes under an hour to set up.

Rarely directly, and the pages say so. Much of the evidence comes from four-year universities, K-12 schools, and laboratories. One comparison of a two-year college with a research university found stronger conceptual learning and lower attrition in introductory physics with peer instruction (Lasry et al. 2008). Each technique page lists its limits. Treat any technique as something to try in your own class and compare with a previous term.