The Forty-Eight Percent That Vanished

Six months before the MIT study, a research team ran a different experiment with nearly 1,000 high school math students. Half received access to a standard ChatGPT-style AI tutor. The AI tutor was effective. Students using it improved their problem-solving performance by 48 percent.^[6]

Then the researchers did something that most AI product demos never do. They took the tool away.

On a subsequent test without AI access, the students who had used the tutor scored 17 percent lower than a comparison group that had never had AI at all. The 48 percent gain was not learning. It was scaffolding. Remove the scaffold, and the students were worse off than if they had never been helped.^[7]

Chunhui Fan, a researcher whose 2025 study coined the term that would ripple through education research, called it "metacognitive laziness"^[8] – a reduced tendency for learners to self-regulate, plan, monitor, and evaluate their own thinking when AI is doing the cognitive heavy lifting.^[9] The assignments improved. The students did not. The output looked like competence. The cognitive residue was dependency.

The tool made the assignment better. It made the student worse. The 48 percent gain vanished. The 17 percent deficit remained.

The Mystique That Pays

If the evidence is this clear (if AI demonstrably weakens the cognitive processes it claims to support) why do students keep using it? Why do schools keep deploying it? Why does adoption accelerate while the research accumulates?

Stephanie Tully and colleagues at Stanford answered this in a 2023 study that Sidra Sidra and Claire Mason, researchers at Australia's national science agency CSIRO, called "one of the most underappreciated findings in the field."^[10] Tully found that individuals with lower AI literacy are more likely to perceive AI as "magical" – as performing tasks that appear to require human intelligence. This perception of magic mediates higher receptivity. People who understand AI less are more impressed by it, more willing to use it, and more likely to accept its outputs without evaluation.

Sidra and Mason, whose own study of 292 AI users was published in the International Journal of Human-Computer Interaction in 2025, noted the implication with academic restraint: "efforts to demystify AI, by enhancing transparency and educating users, may paradoxically reduce its appeal."^[11]

Through the lens of a business model, the restraint falls away. The companies that build AI tools have a financial incentive to keep them mysterious. Transparency reduces engagement. Understanding reduces receptivity. The mystique is not a bug in the user experience. It is a feature of the revenue model. Every student who perceives AI as magical is a student who will not check the AI's work, will not plan how to divide cognitive labour, will not notice when the scaffold is doing all the holding.

The wizard is more profitable when nobody looks behind the curtain.

Half the Room

Joseph Wright of Derby, "An Experiment on a Bird in the Air Pump" (1768). National Gallery, London. The audience watches the demonstration. Understanding the mechanism changes the relationship. Public domain.

Sidra and Mason's study contained a finding that most coverage overlooked. The researchers recruited 344 working professionals who used AI tools regularly. Before administering their scales, they asked three screening questions: does the AI tool "complement my skills," "engage in sustained interaction and two-way feedback," and do you and the AI "build upon one another's work"?^[12]

One hundred and forty-seven people (more than half) could not agree. They were screened out. These are working professionals who use AI daily, and more than half of them do not collaborate with it. They type prompts. They receive outputs. They move on.

The remaining 145 completed two new validated scales. The first, Collaborative AI Literacy,^[13] measures whether users know how to direct, contextualise, and refine AI outputs. The second, Collaborative AI Metacognition,^[14] measures whether users think about their own thinking while using AI – whether they plan, monitor, and evaluate the human-AI collaboration.^[15]

The statistical result was stark. When both scales were tested as predictors of benefit from AI use, metacognition held (β = 0.517, p < 0.001). Literacy dropped to non-significant (β = 0.093, p = 0.397).^[16] Knowing what AI does predicted almost nothing about whether people benefited from it, once you accounted for whether they thought about how they used it.

The MIT study showed what happens in the brain. The math study showed what happens to performance. The Sidra and Mason study showed what separates the people who benefit from the people who do not. It is not knowledge. It is metacognition – the thinking about the thinking.

The Teacher Who Stopped Grading Essays

Al Rabanera teaches math in California. In March 2026, he published a reflection on what AI had shown him about his own teaching.^[17] The revelation was not about the technology. It was about what the technology exposed.

"I didn't realize how much agency I had lost as a teacher," Rabanera wrote, "until AI showed me what I'd been missing – not because AI was disruptive, but because it exposed something harder to face: how comfortable I'd become with waiting."

Rabanera had been waiting for curriculum updates, for district mandates, for professional development that told him what to do next. The students using AI in his classroom were doing the same thing – waiting for the output, waiting for the answer, waiting for the machine to tell them what they thought. The passivity was not caused by AI. It was revealed by it.

Across the country, Pam Amendola, a high school English teacher, faced a different version of the same problem. When ChatGPT arrived, her students asked a question that the AI industry would prefer not to hear: "Why should I complete a worksheet when AI can do it for me?"^[18] The question was honest. The worksheet was designed to produce a finished product. The AI could produce the finished product faster and better. The student's logic was sound.

Amendola's response was to stop assigning worksheets. She redesigned her Macbeth unit so that the process (the reading, the arguing, the writing, the revising) was the assessment, not the output. The essay became secondary. The thinking became primary. She could not assess the thinking if it was invisible. She needed to see it.

The teacher's question is not "did the student use AI?" The question is: "can I see how the student thinks?" If the tool hides the process, the teacher is blind. If the tool shows the process, the teacher can teach.

The Architecture of Visible Thinking

Winslow Homer, "The Country School" (1871). Saint Louis Art Museum. The teacher watches the room. The work is visible. The process is in the open. Public domain.

The MIT brain study, the math tutoring experiment, the metacognitive laziness research, and the Sidra and Mason validation study all point to the same architectural requirement: the thinking must be visible.

Not the output. The thinking. The planning. The monitoring. The evaluation. The moments where the student pushed back on the AI, changed direction, chose to ignore a suggestion, or accepted one uncritically. The process that produces the essay is more important than the essay. The process that solves the math problem is more important than the answer.

No consumer AI tool makes this visible. ChatGPT shows a text box and a response. Claude shows a text box and a response. Gemini shows a text box and a response. The metacognition (the planning, monitoring, and evaluating that Sidra and Mason proved predicts benefit) happens invisibly, if it happens at all.

Sage.Education was built on the premise that the process must be on the table for everyone to see. The Sage.is AI-UI platform provides conversation maps – branching, visual records of every interaction, showing the full tree of prompts, responses, revisions, and decisions.^[19] A teacher using Sage can see not just what the student produced but how: which prompts they wrote, which responses they accepted, where they pushed back, where they revised, where they gave up. This map makes the metacognition visible – to the student, to the teacher, and to anyone evaluating whether the collaboration produced learning or just output.

Sage is an open source platform without the consumerist scale of ChatGPT or the marketing budget of Google Classroom. The architecture is the argument: AI tools for education should show the work, not hide it behind a text box that displays only the result.

The Curtain

W.W. Denslow, "The Discovery of Oz, the Terrible" (1900). Toto has knocked over the screen. The Tin Man confronts the little old man. The machinery is visible. Public domain.

In the Wonderful Wizard of Oz, Dorothy and her companions travel to the Emerald City to meet a wizard they believe is all-powerful. When Toto knocks over the screen, they discover an ordinary man operating machinery. The wizard is not diminished by the revelation. The companions get what they came for: courage, a heart, a brain, a way home. They just get it honestly.

The AI tools in classrooms today are the curtain. They produce impressive outputs. They generate summaries, essays, lesson plans with a fluency that feels like intelligence. The mystique is powerful. The mystique is also the mechanism by which metacognitive laziness enters. The user, awed by the output, stops thinking about the process that produced it.

Pulling the curtain back does not make the tools less useful. It makes the students more capable. Sidra and Mason proved it: metacognition, not literacy, predicts benefit. The MIT study proved the cost of not doing it: cognitive debt accumulates, neural engagement declines, and after four months the person cannot quote their own work.

For educators, the task is clear: Show students how the AI works; make the collaboration visible; assess the process, not just the product; and build tools (or choose tools) that put the thinking on the table instead of behind the curtain.

The wizard is a machine. The machine is useful. Seeing how it works does not reduce its value. It increases yours.

That is what education is for.

The views expressed are those of the editorial board. Sage.is AI-UI and Sage.Education are products of Startr LLC. The author has no financial relationship with MIT, CSIRO, or any AI company referenced. Full disclosure and transparency is a feature, not a bug.