AI adoption in education is outpacing the evidence. A 2026 natural experiment found that ChatGPT availability had effectively zero impact on high school student test scores (Huntington-Klein, 2026). When researchers have reviewed adaptive learning platforms, they have found few large-scale studies proving that these tools actually improve student learning. LLM-based automated scoring degrades on partial-credit responses, the exact responses where teachers need the most help (Education and Information Technologies, 2026). And yet district procurement conversations proceed as if the evidence base were settled.
I don’t say this to dismiss AI. I say it because the gap between what edtech marketing claims and what the research supports is widening, and educators are caught in the middle.
My work sits at the intersection of three pillars: cognitive science, evidence-based instruction, and AI/EdTech. Not as separate domains, but as one integrated practice. The cognitive science and evidence-based practices tell us how learning works and how to teach. The AI tools are the new layer that either supports or undermines both.
The Foundation: Three Pillars
Cognitive science has established a set of principles that are not controversial among researchers, even as they remain underused in classrooms. Retrieval practice, pulling information out rather than cramming it in, consistently improves long-term retention across ages, subjects, and formats, with effect sizes of d ≈ 0.30–0.50 at delayed test (Agarwal, Nunes, & Blunt, 2021). Spaced practice produces durable retention gains (Dunlosky, Rawson, Marsh, Nathan, & Willingham, 2013). Together, desirable difficulties, spacing, interleaving, and varied conditions accelerate long-term learning when implemented under the right conditions (Bjork & Bjork, 2011). Memory is the residue of thought; what students think about is what they remember (Willingham, 2009). These are not hypotheses. They are established findings with decades of converging evidence.
Evidence-based instruction translates those principles into classroom practice. Kirschner, Sweller, and Clark (2006) demonstrated that minimal guidance during instruction overloads working memory, particularly for novices. Mayer’s (2009) cognitive theory of multimedia learning specifies how to design visual and verbal instruction so it builds knowledge rather than decorates slides. Laski, Jor’dan, Daoust, and Murray (2015) extracted four principles from cognitive science for making math manipulatives effective: consistency, transparency, simplicity, and explicit explanation. Ophuis-Cox, Catrysse, and Camp (2023) showed that retrieval practice via flashcards improved multiplication fact fluency in authentic elementary classrooms. The research exists. The question is whether we decide to use it.
The third pillar, AI and EdTech, is where the trouble starts.
The Core Framework
Three claims anchor my framework.
First, pedagogy precedes tooling. Early on, Kasneci et al. (2023) laid out the opportunities and challenges of LLMs in education comprehensively, and the challenge side has grown since publication. An AI tool that generates an essay for a student is not to amplify or accelerate learning; it is outsourcing the cognitive work that produces learning. The distinction matters. A diagnostic tool that identifies a student’s specific misconception and routes them to a worked example is cognitive offloading. An AI that writes the answer is cognitive outsourcing. The first supports the teacher. The second replaces the student’s thinking.
Second, mid-range grading degrades. LLM-based automated scoring performs reliably on binary correct/incorrect items but degrades systematically on partial-credit responses (Education and Information Technologies, 2026). This is the range where formative feedback matters most, where a student has partially grasped a concept and needs targeted correction. AI scoring is least reliable exactly where it would be most useful. This finding should pause any district considering AI auto-grading as a workload solution for teacher grading.
Third, I am not anti-AI. I am pro-evidence. Hendrick and Kirschner (2020) argued that education has a long history of adopting interventions before testing them. AI is the latest instance. The response is not rejection. It is a procurement standard that asks for evidence before full adoption. We should be going slower before we go fast.
Implementation: A Practical Sequence
If a school or district wants to build AI literacy, the sequence matters.
- Pre-assessment. Before selecting tools, assess what teachers and students actually need. What instructional problems are you trying to solve? What does the cognitive science and evidence-based practice say about how to solve them? Start with the learning problem, not the technology.
- Tool selection. Apply a two-axis evaluation: does the tool offload cognition or outsource it? Does it preserve the teacher’s role in feedback and judgment or replace it? Overlay Willingham’s (2009) filter: what is the student thinking about? Does the tool respect working memory limits? Does it build long-term memory or just performance?
- Professional development. PD that does not change classroom practice targets the wrong system layer. Teachers need to understand the cognitive science first, then the tool. Reversing that order produces tool-worship without pedagogical grounding. Ground PD in formative assessment, retrieval practice, spacing, and feedback, then show how AI can support those strategies.
- Measurement. Demand evidence of impact on student learning over time; not engagement metrics, not teacher satisfaction surveys, not usage data. Standardized or independently designed assessments, not researcher-made measures, which can roughly double reported effect sizes (Cheung & Slavin, 2016). If a vendor cannot provide this, that is the answer.
Closing: Implementation Conditions
Ultimately, strategies are conditional. Retrieval practice works but only if students actually retrieve. Interleaving works but only if the material and content is categorically distinct enough for discrimination to matter. Desirable difficulties are desirable only when the difficulty is calibrated to the learner’s prior knowledge. AI tools are no different in this instance. They work when they are deployed under conditions the research tested, with teachers who understand the pedagogy, in systems that measure outcomes rather than adoption rates.
The question is not whether AI belongs in education. It does. The question is whether we have the discipline to adopt it the way the evidence demands, slowly, conditionally, and with pedagogy and evidence in the lead.
References
Agarwal, P. K., Nunes, L. D., & Blunt, J. R. (2021). Retrieval practice consistently benefits student learning: A systematic review of applied research in schools and classrooms. Educational Psychology Review, 33(4), 1409–1453. https://doi.org/10.1007/s10648-021-09595-9
Huntington-Klein, N. (2026). Little impact of ChatGPT availability on high school student test score performance. arXiv preprint arXiv:2605.08812.
Bjork, E. L., & Bjork, R. A. (2011). Making things hard on yourself, but in a good way: Creating desirable difficulties to enhance learning. In M. A. Gernsbacher, R. W. Pew, L. M. Hough, & J. R. Pomerantz (Eds.), Psychology and the real world: Essays illustrating fundamental contributions to society (pp. 56–64). Worth Publishers.
Cheung, A. C. K., & Slavin, R. E. (2016). How methodological features affect effect sizes in education. Educational Researcher, 45(5), 283–292. https://doi.org/10.3102/0013189X16656615
Dunlosky, J., Rawson, K. A., Marsh, E. J., Nathan, M. J., & Willingham, D. T. (2013). Improving students’ learning with effective learning techniques: Promising directions from cognitive and educational psychology. Psychological Science in the Public Interest, 14(1), 4–58. https://doi.org/10.1177/1529100612453266
Hendrick, C., & Kirschner, P. A. (2020). How learning happens: Seminal works in educational psychology and what they mean in practice. Routledge.
Kasneci, E., Sessler, K., Küchemann, S., Bannert, M., Dementieva, D., Fischer, F., Gasser, U., Groh, G., Günnemann, S., Hüllermeier, E., Krusche, S., Kutyniok, G., Michaeli, T., Nerdel, C., Pfeffer, J., Poquet, O., Sailer, M., Schmidt, A., Seidel, T., … Kasneci, G. (2023). ChatGPT for good? On opportunities and challenges of large language models for education. Learning and Individual Differences, 103, 102274. https://doi.org/10.1016/j.lindif.2023.102274
Kirschner, P. A., Sweller, J., & Clark, R. E. (2006). Why minimal guidance during instruction does not work: An analysis of the failure of constructivist, discovery, problem-based, experiential, and inquiry-based teaching. Educational Psychologist, 41(2), 75–86. https://doi.org/10.1207/s15326985ep4102_1
Kraft, M. A. (2019). Teacher effects on complex cognitive skills and social-emotional outcomes. Journal of Human Resources, 54(1), 1–36.
Laski, E. V., Jor’dan, J. R., Daoust, C., & Murray, A. K. (2015). What makes mathematics manipulatives effective? Lessons from cognitive science and Montessori education. SAGE Open, 5(2), 1–8. https://doi.org/10.1177/2158244015589588
Education and Information Technologies (2026). Lost in the middle? Examining scoring reliability and position bias in LLM-based automated essay scoring. Education and Information Technologies. https://doi.org/10.1007/s10639-026-14019-8
Mayer, R. E. (2009). Multimedia learning (2nd ed.). Cambridge University Press.
Ophuis-Cox, F. H. A., Catrysse, L., & Camp, G. (2023). The effect of retrieval practice on fluently retrieving multiplication facts in an authentic elementary school setting. Applied Cognitive Psychology, 37(6), 1463–1469. https://doi.org/10.1002/acp.4141
Willingham, D. T. (2009). Why don’t students like school? A cognitive scientist answers questions about how the mind works and what it means for the classroom. Jossey-Bass.