A professor I know timed her own midterm the first time she entered it into ChatGPT. It took six seconds. With the work displayed, the model returned a neat, readable response. She refrained from gasping. She didn’t send her dean an email.
She simply sat in an office on the third floor of an old humanities building that was a little too warm, staring at the screen in the same way that you stare at a leaky pipe that you have been ignoring.
| Topic | Co-designing AI with students rather than for them |
| Core shift | From AI-proofing assignments to designing assignments that demand human thinking |
| Reported performance lift | 30% average gain in personalized AI-driven learning (McKinsey, 2024) |
| Adoption signal | 75% of countries piloting or integrating AI in education (UNESCO, 2024) |
| Teacher behavior | 68% using generative AI weekly to prep lessons (Global Education Forum, 2025) |
| Public funding cited | €1.3 billion committed via the AI4Education program |
| Dropout impact | Finland cut school dropout 18% in three years using learning analytics |
| Accessibility horizon | 90% of children with disabilities reachable by 2030 (UNICEF estimate) |
| Tools mentioned | Squirrel AI, Century Tech, Gradescope, Copilot for Education, NotebookLM |
| Why it matters | The metric of teaching is shifting from time spent to learning produced |
This year, faculty offices all over the world are witnessing that small scene. Silent experiments. quieter insights. The task wasn’t particularly difficult. Simply put, there wasn’t much that a machine couldn’t accomplish more quickly already. Talking to teachers these days gives me the impression that the old reflex—lock it down, outlaw the tool, create a more stringent curriculum—has run its course. AI is already present in the room unless you are proctoring students in person and without a connection. They received it via their phones. It entered through their web browsers. This one speaks back, but it entered through the same back door that their calculators used decades ago.
Who gets a seat at the design table is gradually changing. AI for students—adaptive platforms, automated grading, dashboards that alert students to a declining grade before they even realize it—has been the topic of discussion for the past two years. useful and occasionally eerie. With real-time lesson adjustments, Squirrel AI’s deployments in China currently reach about 100 million students each week. In a single night, Gradescope can process thousands of papers. Microsoft estimates that grading saves about 45% of the time. These figures are not insignificant. These are the kinds of figures that the finance committee observes before the pedagogy committee.
However, the teachers who stopped creating assignments centered around AI and began creating them with their students, including AI, are the ones who appear to be the most at ease right now. There is a quantifiable calmness to them, a sort of sleep-well quality. They execute loops of iteration. Students are asked to turn in their prompts as well as their responses. Instead of grading the artifact, they grade the conversation. According to one teacher I spoke with, it’s like “watching them think out loud, finally.” She might be romanticizing it. She might also be onto something that the policy memos haven’t yet addressed.
Where the data is available, it favors her. Up until you sit with it, McKinsey’s 30% increase in personalized learning seems like a marketing ploy. It is more difficult to ignore Finland’s 18% dropout reduction over a three-year period, as shown by the Learning Analytics 360 project. A visually impaired child in Norway uses Seeing AI to read aloud from a page. A refugee student in Spain uses real-time translation to follow a lesson that was unavailable four years ago. These are no longer pilots. It’s Tuesday.
Student voice in the build is what’s lacking and what the more experienced teachers are constantly pointing out. Although they are rarely trained with students, adaptive systems are trained on them. There is always a child who plays the platform, and they have knowledge that the engineers do not. There won’t be an A/B test if the student who dislikes the chatbot’s tone provides feedback. The next generation of educational AI will be evaluated more on whether the people learning from it had any influence on its development than on benchmark scores, according to a quiet argument that is emerging, mostly in conference rooms and Substack drafts.
It’s difficult to ignore how frequently the word “friction” is used in settings where it was once considered offensive. Reflection is the result of friction. Revision is what friction entails. Friction occurs when a student must justify a decision that the model made for them, override it, or provide an explanation for why they didn’t. No algorithm can outsource that portion. Not just yet. Perhaps never. And if, as one French essayist recently stated, the measure of education is actually changing from time spent to light transmitted, then the teachers who are already holding the candle a little differently are the ones who are worth keeping an eye on.
