Fast AI vs Slow AI

The question nobody is asking

Most conversations about AI in schools focus on whether students are using it. The more important question is how.

Fast AI means using AI to get quick answers, skip the struggle, and produce polished outputs rapidly. Paste the question, read the answer, move on. It feels efficient. It produces results that look like learning.

Slow AI means using AI to think harder, not less. To get a question answered with another question. To have your reasoning challenged. To receive a hint rather than a solution. It can feel harder, but well-designed guidance can support deeper learning.

The OECD's Andreas Schleicher puts it plainly: "GenAI is clearly not a magic wand. It is capable of magnifying good pedagogy and bad." The way a tool is used can change its learning outcomes.

Why struggle matters

Productive struggle is the effort of working through something that is genuinely difficult. Not impossible: that is just frustration. Not easy: that is not learning. The zone in between, where a student is challenged but not overwhelmed, is where the most durable learning happens.

Derek Muller, who runs the Veritasium science channel and holds a PhD in physics education, draws the connection directly: "There's a danger in using AI or any technology in education if it causes us to skip the effort." Learning, he argues, requires converting slow deliberate thinking into automatic fast knowledge through practice. Skipping the practice skips the learning.

The generation effect is a well-established finding in memory research: people often remember material better when they actively generate it rather than merely read it. A 2007 meta-analysis of 86 studies reported an average effect size of 0.40 standard deviations, not a 20–40% increase in retention. When students copy AI-generated answers without thinking, they may miss some of this beneficial practice.

This is not an argument against AI in classrooms. It is an argument for understanding what happens cognitively when students use it.

More problems, less understanding

A field experiment with Turkish high school math students, published in PNAS in 2025, found a result that every teacher should know.

Students using ChatGPT without guardrails completed 48% more problems correctly during practice sessions. On paper, that looks like improvement. But when those same students were tested on conceptual understanding, they scored 17% lower than students who had learned without AI assistance.

They got more problems right during practice by leaning on AI. They understood the material less as a result.

The critical note: the same study found that a well-designed "GPT Tutor" with appropriate guardrails (one that used Socratic questioning rather than direct answers) largely eliminated the negative effects. The tool is the same. The design is different. The outcome is different.

Cognitive offloading: what it looks like

Cognitive offloading is what happens when a person outsources mental effort to an external tool. A calculator offloads arithmetic. GPS offloads navigation. AI, used as an answer machine, offloads thinking.

A 2025 study published in Societies reported a negative correlation between AI-tool use and critical-thinking scores among 666 participants; correlation alone cannot show that AI caused the difference. In a separate study of 45-day retention, the AI-assisted group scored 57.5% and the comparison group 68.5% under that study's conditions, an 11-percentage-point difference—not a general prediction for every classroom.

Anthropic's own research found that students asked Claude for direct answers almost half the time, offloading higher-order cognitive functions at exactly the level where learning is supposed to happen.

Harvard's Christopher Dede captures the classroom implication: "If a student uses AI to do the work for them, rather than to do the work with them, there's not going to be much learning."

The most important number in this research

A 2025 meta-analysis by Gu and Yan, published in a SAGE journal and drawing on 19 studies, found an overall positive association between generative-AI interventions and academic performance, with a pooled effect size of 0.683. Outcomes varied by study and instructional setting.

The breakdown is what matters. With teacher support: effect size 1.426. Without teacher support: effect size 0.077.

In that set of studies, interventions with teacher support had substantially larger estimated effects than interventions without it. That does not prove teacher presence was the sole cause or that students can never learn independently with AI. It does underline the importance of instructional design and teacher guidance.

For educators concerned about replacement, the finding highlights an important role for teachers in designing tasks, monitoring understanding, and helping students use AI deliberately. It is a strong reason to include teachers in AI implementation, not a universal prediction about every classroom.

What Slow AI actually looks like in practice

The distinction is not about which tool you use. It is about how the interaction is structured. Here is the same student task handled two ways.

The Bellwether report from June 2025 frames it well: "AI's role in accelerating or weakening learning largely rests on how well it can turn the dial of productive struggle up or down." Not all friction is beneficial, and not all ease is harmful. The teacher's job is to set the dial deliberately.

Section 3 covers the specific tools and prompts that make Slow AI work in practice.