What Does the Research Actually Say About AI in K-12 Classrooms?
Authors: Yerin Seung; info@ciddl.org
Artificial intelligence tools are entering classrooms faster than the research base evaluating them can keep up with. For educators, that gap creates a real dilemma: how do you make sound instructional decisions about AI when rigorous evidence remains so thin? This post summarizes a new report from Stanford's AI Hub for Education, The Evidence Base on AI in K-12: A 2026 Review (Fesler et al., 2026), which offers one of the clearest pictures yet of what we currently know and, just as importantly, what we don't.
A Small Slice of Strong Evidence
The Stanford team's Research Repository contained over 800 academic papers on AI in K-12 education as of October 2025. Of those, only 20 met the bar for strong causal evidence, which were designed to isolate the actual effect of an AI tool rather than simply describe its use through randomized controlled trials or quasi-experimental designs (Fesler et al., 2026). Notably, none of the student-facing causal studies were conducted in U.S. K-12 settings; most involved university students or international high school populations (Fesler et al., 2026). This means the findings below should be read as suggestive signals rather than settled conclusions for American classrooms.
Four Patterns Worth Watching
- AI helps in the moment, but the gains don't always stick. Several studies found that students performed better on math, writing, or programming tasks while they had active access to an AI tool. However, those gains weakened or disappeared once the tool was removed. In one widely cited example, high school students in Turkey who used a general-purpose AI chatbot to study for an exam scored about 17% worse on an unassisted final than peers who used no AI at all, even though they had outperformed those peers during AI-supported practice (Bastani et al., 2025). The concern here connects to a core learning science principle: transfer. If students are learning to work the tool rather than internalizing durable skills, performance won't generalize to contexts where the tool isn't available.
- Easier isn't always better. AI can meaningfully reduce cognitive load and make tasks feel more manageable, but that relief doesn't automatically translate into deeper learning. One study found that university students using a general-purpose AI chatbot for research produced lower-quality reasoning and argumentation than students using a traditional search engine, even though they reported the task felt easier (Stadler et al., 2024). This is a useful reminder for special educators in particular: reducing extraneous cognitive load is often the goal of good scaffolding, but productive struggle or what learning scientists call "desirable difficulties" still matters for retention.
- Tool design matters as much as tool access. Not all AI chatbots produce the same results. Tools built with pedagogical guardrails that nudged students toward an answer through hints rather than handing it over outright consistently outperformed general-purpose chatbots in the studies reviewed. In the same Turkish high school study cited above, students using a tutoring-specific chatbot performed on par with their textbook-only peers, while students using the unguarded, general-purpose version performed worse (Bastani et al., 2025). This is a meaningful finding for procurement and tool-selection decisions: a "free" general-purpose AI assistant is not pedagogically equivalent to a purpose-built tutoring tool.
- For educators, the evidence is more encouraging. Teachers given access to ChatGPT along with structured guidance spent roughly 30% less time on lesson preparation, with no detectable drop in lesson quality based on blind expert review (Roy et al., 2024). AI-generated feedback on classroom discourse increased teachers' use of higher-quality "focusing questions" by 20% (Demszky et al., 2025), and a real-time AI coaching tool for tutors produced especially large gains for less experienced and lower-rated tutors, a 9-percentage-point improvement in student topic mastery for that subgroup (Wang et al., 2025). For teacher educators, this suggests AI-supported coaching may hold real promise as a scalable complement to traditional, costly professional development models.
What's Still Missing, and Why It Matters for Our Field
The current causal literature includes no high-quality studies examining AI's effects on students with IEPs or 504 plans, and questions about fair access, including whether under-resourced districts can afford effective, education-specific tools, remain largely unanswered (Fesler et al., 2026). Similarly, the evidence base on AI's effects on social-emotional development and student wellness is thin, even as informal use of AI companions among children and teens is rising rapidly (Fesler et al., 2026; Robb & Mann, 2025).
For educators working at the intersection of general and special education, this is an invitation rather than a dead end. The available evidence, limited as it is, points toward design principles that align well with what we already know about effective scaffolding: AI tools that provide graduated support within a student's zone of proximal development, rather than complete answers, appear more likely to support durable learning (Fesler et al., 2026). Until more U.S.-based, special-education-specific research exists, that principle is a reasonable starting point for evaluating any AI tool brought into the classroom.
Join the conversation in our community!
CIDDL is committed to providing high-quality resources to support the growing knowledge, adoption, and use of a range of educational technologies for educators, related services, or leadership preparation programs. For more resources, including videos and blogs, subscribe to our newsletter and follow us on YouTube, Facebook, and LinkedIn. The most important part of our CIDDL community is YOU. Join our community and share the innovative ways you are using technology, ask a question about technology integration, or participate in our bi-weekly live AI Community Chats. We look forward to seeing you in our community!