“The code works. The answers look clean. But when I ask ‘Why does this work?’, the students struggle to explain.”
This observation from a veteran Indian educator captures a growing worry across the country. While standard Artificial Intelligence (AI) helps students finish homework faster, it is not always helping them understand the logic behind it. Students are using digital tools that give them the “answer” but hide the “explanation.”
However, a breakthrough at Tufts University suggests a massive technological pivot is coming. Researchers built a new kind of AI that achieved a 95% success rate on complex tasks while using 100 times less energy than the AI we use today. It is called Neuro-symbolic AI, and Indian scientists are already using it to turn AI into a true teacher rather than just a shortcut.
Difference between guessing and knowing
Most AI today (like ChatGPT) works like a very smart “auto-complete.” It predicts the next word based on patterns it has seen before. While brilliant, it doesn’t actually “know” the rules of math or physics—it just guesses what a right answer looks like.
Neuro-symbolic AI is different. It combines the visual “pattern-matching” of a human brain with the rigid “rule-following” of a calculator. Prof. Somak Aditya, Assistant Professor, Department of Computer Science, IIT Kharagpur, who leads the PrahelikaAI project, explains this using a classic challenge: placing eight queens on an 8×8 chessboard so they don’t attack each other.
“A student may visually interpret a chessboard,” Prof. Aditya says, “but solving the puzzle requires strict logical reasoning within well-defined rules.” He notes that the future of educational AI must combine the visual strengths of neural networks with the “reasoning capabilities of symbolic logic systems.” By doing this, the AI doesn’t just guess an answer—it follows the rules of the game.
Move beyond rote learning
“One of the biggest problems in the Indian education system is that we are motivating students to remember things,” says Prof. Aditya, whose project aims to break this cycle of memorization.
The idea came to life after his team visited anganwadis and local schools. “Many of the students do not have equitable access to quality education. Second, they lacked personalized teaching,” he observes, pointing to UN reports that highlight the severe lack of personalized learning in India.
The solution, he believes, lies in technology and especially a universal human joy: puzzles. “From Class 5 till your engineering, students are exposed to puzzles, and they stimulate the brain in completely different ways,” he says, noting the phenomenon even in his own five-year-old child. “Puzzles stimulate the brain because you cannot remember the solution; you have to understand the process.”
The Prahelika model
To bring this logic-first education to the average Indian student, the team at IIT Kharagpur is currently developing PrahelikaAI around three core pillars that move from data to the student’s screen.
First, the team is creating a large-scale puzzle dataset and translating problems into Bengali and Hindi, ensuring that the rules of reasoning are accessible to students in their own mother tongue. Second, they are developing advanced neuro-symbolic algorithms and next-generation “Vision-Language” models designed specifically to “see” a diagram and logically reason through it.
The final component is an intelligent user interface designed to act as a 24/7 digital tutor. Currently browser-based with a mobile version in development, this platform doesn’t simply hand over the answers; it watches how a student interacts with a problem to know exactly when to step in.
“When the AI observes a long delay in answering the problem, it gives a casual nudge saying, ‘Think in this manner,’” says Prof. Aditya. First, the AI simplifies the problem. If the student is still stuck, it gives an example. Only after exhausting all attempts does it reveal the final solution. By tracking this progress, the system builds a personalized learning profile, identifying recurring mistakes and mapping them to actual misconceptions in the student’s learning journey.
Why Big AI fails India’s classrooms
The push toward these logic-first tools becomes clear when looking at why massive, data-heavy models are a misfit for Indian schools. Leading AI researcher Dr. Amit Sheth, Founding Director of the Indian AI Research Organization (IAIRO), argues that “India’s classrooms are not Silicon Valley data centres.” General-purpose systems like GPT-4 fail on crucial structural counts, starting with a massive infrastructure mismatch. GPT-4’s staggering energy demand reflects “the energy demands of a small city, not a classroom tool,” which is an impossible fit for India’s rural and semi-urban schools where “only 47% of rural schools have functional computers” and high-bandwidth internet simply does not exist across large swathes of the country.
Furthermore, these English-dominant systems create a deep data availability trap. When confronted with India’s 22 constitutionally recognized languages and hundreds of regional dialects, a standard model “does not become less confident, it becomes confidently wrong.”
This approach worsens the pedagogy problem by amplifying rote learning. Because standard AI is just a “pattern-matching and text-generation system,” it generates fluent answers by retrieving statistical formulations rather than reasoning through concepts, meaning a student reading these answers “is no more capable of original scientific reasoning than before.”
Finally, the economics of deploying premium global models are prohibitive at India’s scale. While accessing a global frontier model cloud costs around $0.09 per request, compact, specialized models cost roughly $0.0004 per request, making it “over 200 times cheaper.” For a government deploying AI tutoring tools across 1.4 million schools, this extreme cost gap is not merely significant; “it is existential.” As Dr. Sheth summarizes, “The question for Indian education is not ‘What is the most powerful AI?’ but ‘What is the most appropriate AI?’”
Eradicating AI hallucinations
Beyond the economic crisis, the greatest threat standard AI poses to education is “hallucination”—or confident lying. Dr. Sheth warns that “the hallucination problem is not a software bug that will be patched in the next model release.” Instead, it is a permanent architectural feature of Large Language Models. Because an LLM “predicts the most statistically likely next token given the preceding context,” it lacks real grounding or semantics. When a student pushes it to its limits, the model confidently generates what sounds like an answer, “because sounding like an answer is precisely what it has been optimised to do!” In a high-stakes classroom, Dr. Sheth calls this a “trust-destroying, learning-corrupting failure mode.”
To eliminate this pedagogical failure, neuro-symbolic technology creates a structural separation between what the creative neural component generates and what a strict symbolic component verifies. Dr. Sheth’s team pioneered the CREST framework, which deploys structured “knowledge graphs”—essentially digital maps of authoritative facts, like the NCERT science curriculum—to act as an unyielding factual guardrail. If a standard model tries to mathematically guess a historical date or formula incorrectly, the symbolic layer instantly catches the error against the graph.
This architecture changes the classroom experience by introducing “user-level explainability”—explanations written “not for researchers, but for the actual user, in this case, a student or teacher.” Because the symbolic layer leaves a clear audit trail of its logic, it can tell a student exactly which chapter and principle it used to solve a problem. Conversely, when the system cannot find a verified fact to back up its claim, it safely restricts itself. In an educational setting. Dr. Sheth notes, “’I don’t know, let’s ask your teacher’ is infinitely preferable to a confident fabrication.”
This logic-first architecture is already moving into active classrooms through his team’s Personal Adaptive Learner (PAL) system. Built as an intelligent learning manager, PAL leverages these neuro-symbolic guardrails to transform static school lessons into interactive experiences, asking students questions at adaptive levels of difficulty without the risk of fabricating feedback.
Inclusive growth
To understand what this technical shift looks like in practice, one must look at the students left behind by Silicon Valley. Dr. Sheth frames this through a specific profile: a 19-year-old female student enrolled in a diploma program at a government polytechnic in a small town in Odisha. “She has a second-hand Android smartphone with a 2G connection that drops regularly. Her medium of instruction is Odia.” She cannot afford tutoring, and standard cloud-dependent models were simply not designed for her.
Neuro-symbolic tools built on the compact, hardware-conscious principles of Dr. Sheth’s C3AN framework change this calculation entirely by allowing edge deployment, meaning the AI travels with the student. Because a domain-specific model has a defined scope, it is highly compressible, can be downloaded once onto a device, and will “function entirely offline.” It also changes the game for regional languages. While English-dominant models generate text that is conceptually imprecise, a neuro-symbolic system “does not merely translate English AI responses into Odia, but reasons in Odia from the ground up.” Driven by voice interfaces, it preserves a cultural and historic fidelity to India’s past that Western-trained models intrinsically lack.
Furthermore, it introduces intelligent diagnosis. If a student struggles with engineering concepts, the system can reason back through explicit representations of concept dependencies to see exactly which underlying principle is misunderstood. “The causal chain of prerequisite concepts is encoded symbolically, enabling a form of intelligent diagnosis” that purely statistical models cannot reliably perform.
Finally, it allows for teacher integration. Because the system’s reasoning is transparent, a polytechnic teacher “can see exactly how the AI system arrived at an answer, which textbook section was referenced, [and] which rule was applied,” allowing them to supervise and trust the tool rather than viewing it as an opaque black box. True educational sophistication in India, Dr. Sheth argues, “is not measured by model parameters but by contextual alignment, matching the right pedagogical tool to the specific student, language, and hardware.” The ultimate value of this homegrown tech sovereignty lies in “empowering a diploma student in rural Odisha, enabling her to master thermodynamics in Odia at midnight on a secondhand device, completely offline. This is the authentic architecture of inclusive growth.”
The ‘explainable AI’
This combined approach may well align India’s National Education Policy (NEP) 2020, which champions experiential, inquiry-driven learning over rote learning.
Neuro-symbolic AI fixes this by being fundamentally “explainable.” By practicing interactive puzzles, students build the exact logical foundations that mathematicians use to solve complex equations and prove theorems. Furthermore, because these rule-based systems are lightweight, they don’t need massive computing power plants to run. “Our goal,” says Prof. Aditya, “is to make these technologies lightweight, affordable, and accessible—so that powerful educational tools can eventually run even on smartphones.”
While the application currently targets students with smartphones and an internet connection, parallel research is underway to compress these models further. “It will be our next step to offer the application without an internet connection, benefiting rural students,” Prof. Aditya notes.
Ultimately, this technology changes the classroom dynamic by removing the fear of making mistakes. “A single computer or tablet available in a school will make a whole lot of difference,” says Prof. Aditya. “Students can track their own progress and they will start discussions freely with other students as puzzles are fun. It breaks away their inhibitions.”
Some caveats
Bringing a critical perspective from the frontlines of pedagogy, Dr. S. Thangarajathi, Professor and Head of the Department of Education at Bharathiar University, highlights the deep structural and human limitations of deploying these systems in real-world classrooms. “Its implementation at the grassroots level particularly in schools and educational institutions faces several practical constraints. Many institutions still struggle with basic digital infrastructure, teacher training, data quality, and equitable access to technology. Without addressing these foundational issues, even sophisticated AI models may remain limited to pilot projects rather than becoming scalable educational solutions.”
“One important concern is that AI-supported educational systems may unintentionally increase the cognitive and administrative burden on teachers. It leads to technological overload,especially when educators are already balancing teaching, assessment, emotional support, and administrative responsibilities. Unless systems are designed with simplicity, contextual flexibility, and adequate training, AI may risk becoming an additional layer of work rather than a meaningful support mechanism,” she points out.
“AI systems—including Neurosymbolic AI—can certainly assist in identifying learning patterns, recurring conceptual errors, behavioural trends, and possible indicators of academic difficulty,” she says.
She further adds, “However, diagnosing a child’s psychological or conceptual blocks is a deeply human and contextual process. Children’s learning difficulties are often influenced by emotional states, family environment, motivation, peer interaction, socio-cultural experiences etc that may not be fully captured through data-driven systems. AI can support early identification and personalized learning recommendations, but it cannot completely replace the nuanced observation, empathy, and professional judgment of teachers, counsellors, and psychologists. Human interpretation remains essential, particularly in educational and mental health contexts. I believe that no one can replace the teachers.”
Remove