The important question is no longer whether artificial intelligence will enter education. It already has. The question is what kind of relationship we will build with it.
We can design technology that produces increasingly fast answers, or technology that returns questions, hints, and alternative perspectives. The first removes friction. The second turns friction into learning. Both may look useful while a task is underway, but they do not produce the same result once the tool disappears.
This distinction is at the heart of GRAVYA: intelligence responds; judgment is formed.
Performance is not the same as learning
A randomized trial published in the Proceedings of the National Academy of Sciences worked with nearly one thousand high school students in Turkey across four mathematics sessions. It compared three conditions: standard learning materials, a ChatGPT-like interface, and a GPT-4 tutor with pedagogical safeguards designed with teachers.
While AI was available, the results looked remarkable. The open-interface group improved its assisted-practice performance by 48% relative to the control group; the safeguarded tutor group improved by 127%. But on a later exam completed without AI, students with open access performed 17% worse than students who had never used the tool. In the safeguarded-tutor group, that harm disappeared, although the researchers did not find a significant positive effect on the independent exam.
The lesson is not that AI inevitably damages learning. It is more precise: an aid that maximizes the immediate result can weaken the acquisition of the ability that result is supposed to represent. Design matters. Giving away the solution and accompanying the reasoning are not equivalent.
The study also suggested a mechanism. Students tended to use the open version as a crutch, requesting and copying solutions, while students using the protected tutor asked for help and attempted answers independently. They also appeared unaware of the learning loss. Ease can create a dangerous illusion: confusing fluency with mastery.
Positive evidence also has a human condition
Other experiments show that AI can expand learning when it occupies a supporting role.
In Nigeria, a World Bank randomized trial evaluated a six-week after-school program combining generative tutoring with teacher guidance. The overall effect was 0.31 standard deviations, a magnitude the authors compare with roughly 1.5 to 2 years of typical learning. This was not an isolated conversation between a student and a machine. It involved structured sessions, present teachers, and defined pedagogical goals.
Tutor CoPilot offers another signal. In a trial involving human tutors and students from historically underserved communities, students working with AI-assisted tutors were four percentage points more likely to master mathematics topics. Among initially lower-rated tutors, the difference was nine points. The system did not replace the tutor; it suggested guiding questions, examples, and strategies. Its greatest value appeared where AI helped a person practice their craft more effectively.
These findings do not justify universal promises. The contexts, ages, subjects, and intervention lengths differ. They do support a strong design hypothesis: educational AI works best when it increases the capacity of the human guide and protects the learner’s cognitive activity.
Five boundaries that turn an answer into learning
An AI system aligned with GRAVYA should:
- Ask before solving. Check what the learner understands and where the reasoning breaks down.
- Offer graduated hints. Provide the minimum help needed so the next operation remains human.
- Require retrieval. After an explanation, remove the support and ask the learner to rebuild the idea in their own words.
- Show uncertainty and sources. Separate what is verified, what is inferred, and what still requires checking.
- Return the decision. The tool proposes; the person compares, chooses, and accepts responsibility for the consequences.
UNESCO has called for a human-centered, safe, equitable, and meaningful approach to generative AI in education. That orientation cannot be solved by placing an ethical statement at the beginning of a project. It must appear in the interface, the rhythm of the conversation, what the tool refuses to do, and the way it returns agency to the learner.
The GRAVYA position
GRAVYA does not propose withdrawing from artificial intelligence. It proposes inhabiting it with judgment.
AI can widen the field of questions, offer contrast, adapt an explanation, or help a teacher notice patterns. But it should not appropriate the moment in which a person names a doubt, stays with uncertainty, attempts a response, and decides what they think.
That moment is not a delay to automate. It is where learning happens.
Question for the community: Does your AI tool help you understand more deeply, or does it simply help you finish sooner?
References
- Bastani, H. et al. (2025). Generative AI without guardrails can harm learning: Evidence from high school mathematics. PNAS.
- De Simone, M. E. et al. (2025). From Chalkboards to Chatbots: Evaluating the Impact of Generative AI on Learning Outcomes in Nigeria. World Bank Policy Research Working Paper 11125.
- Wang, R. E. et al. (2025). Tutor CoPilot: A Human-AI Approach for Scaling Real-Time Expertise. Stanford SCALE Initiative.
- UNESCO (2023, updated 2026). Guidance for generative AI in education and research.