AI-Powered Assessment and the Future of Student Success| India AI Impact Summit 2026

Contents

Executive Summary

This summit session demonstrated that AI for education is transitioning from promise to practice, with rigorous evidence now validating impact at scale. The Letris case study from Brazil showed that AI-driven writing assessment can dramatically improve learning outcomes and reduce educational inequality when combined with robust pedagogical design and implementation strategy—not just algorithmic sophistication. The broader message: AI's value in education depends on intentional evaluation, local context adaptation, ethical design, and partnership between technologists, educators, policymakers, and researchers.

Key Takeaways

Evidence + Implementation = Impact at Scale: Letris's path from internal data → RCT validation → government policy → 1,600 schools and 500,000+ students shows that rigorous evaluation is the bridge between innovation and public sector adoption, especially in resource-constrained settings.
Context and Pedagogy Trump Algorithmic Novelty: The most transformative AI for education combines simple technology (offline WhatsApp, offline apps, voice recording) with deep pedagogical design (TARL, curriculum mapping, teacher workflows) rather than relying on cutting-edge models alone.
Build Trust Through Transparency and Replicability: As Shobini framed it: "If a government algorithm makes a decision and no one can explain it, is it public policy or public mystery?" Responsible AI for public systems requires explainability, open evaluation, and documented methods.
Leverage Existing Incentives and Infrastructure: The most cost-effective interventions don't ask communities to adopt new behaviors; they align technology with existing motivations (parents wanting their kids to succeed, teachers wanting respect) and leverage existing touchpoints (community workers, mobile phones).
The Future of Learning Goes Beyond Test Scores: Rokamini's closing challenge—to move beyond exams and grades toward open-ended exploration of what children can become—signals that as AI handles routine assessment and feedback, the next frontier is helping educators imagine and enable deeper, more human forms of learning and development.

Key Topics Covered

AI-powered literacy assessment and feedback systems – Letris platform in Brazil using AI to analyze student writing and provide personalized feedback
Randomized controlled trials (RCTs) as validation tool – Five-month evaluation demonstrating significant learning gains and inequality reduction
Teaching at the Right Level (TARL) – Adaptive, learner-centered pedagogical approach enhanced by AI data collection
Foundational literacy and numeracy – Focus on early reading skills, phonetic vs. visual error analysis, and reading progression
EdTech implementation at scale – Infrastructure, curriculum integration, offline-capable apps, and teacher enablement
Large language models and learning – Google's LearnLM model and pedagogical principles encoded into foundation models
Personalized learning through AI – Addressing multi-age, multi-grade classrooms; adaptive pathways for diverse learners
WhatsApp-based AI for early childhood education – Rocket Learning's approach using conversational AI and simple mobile technology
Research and AI tools – AI co-scientist, hypothesis generation, deep thinking models for accelerating scientific discovery
Ethics, bias, and responsible AI – Multilingual datasets, cultural relevance, transparency, and avoiding reinforcement of disparities
Evidence-based policymaking – Embedding evaluation alongside innovation; moving from RCTs to faster interim measurement
Last-mile reach and equity – Using AI to serve underserved, vulnerable populations; cost-effectiveness and incentive alignment

Key Points & Insights

Letris case study outcomes: Over 9 months in Aspirus state, Brazil, the platform increased teacher one-on-one engagement by 35%, improved writing test scores by 17 points, and closed the public-private school gap by 9 percentage points on national exams. Students in treatment schools leapfrogged from below-average to second place in state-wide rankings.
Implementation complexity matters more than algorithm sophistication: Letris has worked for 10 years without GPT; success came from deep understanding of context, curriculum integration, offline functionality, teacher digital literacy support, and data dashboards for administrators—not just the algorithm.
RCTs are expensive and slow, but necessary for policy legitimacy: A five-month RCT was critical for government adoption in Brazil. However, newer models use faster interim measures (digitized data, proxy outcomes) alongside RCTs to reduce time-to-evidence without sacrificing rigor.
Pedagogical principles must be explicitly coded into AI models: Google's LearnLM work shows that foundation models perform better when trained against specific learning science principles (e.g., scaffolding, Socratic questioning, avoiding cognitive overload) rather than relying on general text patterns.
AI works best when embedded in existing incentive structures: Rocket Learning's success in early childhood leverages existing motivations (parents' financial incentive, ANGANWADI workers' desire for respect) and existing infrastructure (WhatsApp, community workers) rather than trying to create new ones. Cost: $1.50/child/year; 75% reaching school readiness vs. 50% national average.
Data-driven adaptive learning requires moving beyond grade-based grouping: Pratham's work reveals that even simple tools (ASER) become powerful when AI analyzes patterns (phonetic vs. visual letter errors) and informs immediate instructional adjustment, enabling teachers to address multi-level classrooms more effectively.
Language and cultural bias remain critical gaps: Models are weaker in non-English languages, minority languages, and culturally specific content. Solving this requires open datasets (Google's Thousand Language Moonshot, African language initiatives) and collaborative work with local experts—not just tech companies.
Access determines equity outcomes: The same AI tool can deepen inequality if deployed only to wealthy schools or without offline capability, teacher training, and curriculum alignment. Responsible scaling requires intentional design for the hardest-to-reach populations.
Evaluation must move alongside innovation: Because AI changes rapidly, the ecosystem needs both rigorous RCTs and faster feedback loops (interim measures, cost-effectiveness analysis, qualitative implementation data) to enable iterative improvement while maintaining accountability.
Human expertise remains irreplaceable: No panelist advocated "leaving it all to AI." Teachers, parents, administrators, and policymakers still drive planning, decision-making, and strategic direction; AI augments their capacity, not replaces it.

Notable Quotes or Statements

"Mastering the written language is the most unique and important trait of humankind. But this skill is a privilege of a few. In Brazil, 90% of adults are not considered fully proficient."
— Thiago Rashid, Letris CEO, framing the equity problem

"Having an RCT was incredibly crucial for us to close our first contract and build the reputation we needed to expand our program."
— Thiago Rashid, on why rigorous evidence was essential for government adoption

"If a government algorithm makes a decision and no one can explain it, is it public policy or public mystery?"
— Shobini Mukharji, JPAL South Asia Executive Director, on the imperative for transparency in public AI systems

"What allow us to reach the results that we have are focus on implementation, especially when we talk about scale and diversity… our capacity to really understand context and adapt our technology to that context."
— Thiago Rashid, emphasizing implementation and adaptation over algorithmic sophistication

"The basis for scale has to be somebody wants what you're doing. The incentive cannot be built using technology."
— Namya Mahajan, Rocket Learning CEO, on the primacy of human motivation over technological innovation

"AI is transforming research and scientific discovery by helping us generate better hypotheses, run parallel experiments, and point human attention at the right questions at the right time."
— Bridget Orr Gosling, Google.org, on AI's emerging role in accelerating research

"We really need to use the non-linear capabilities of technology… break out of constraints and let children explore how far they can go with themselves, their friends, their books, their teachers, and with AI."
— Rrookmini Banerjee, Pratam CEO, calling for reimagining education beyond traditional exams

"As AI handles routine assessment and feedback, the next frontier is helping educators imagine deeper, more human forms of learning."
— Implicit across multiple panelists' remarks

Speakers & Organizations Mentioned

Featured Speakers

Thiago Rashid – Co-founder and CEO, Letris (Brazil)
Sam [Policy Manager, JPAL] – Researcher at Abdul Latif Jameel Poverty Action Lab (JPAL), led Letris RCT
Rrookmini Banerjee – CEO, Pratam Education Foundation (India)
Namya Mahajan – Co-founder and CEO, Rocket Learning (India)
Bridget Orr Gosling – Director of AI and Social Impact, Google.org
Margaret Clark – Lead Education Specialist, World Bank
Shobini Mukharji – Executive Director, JPAL South Asia

Organizations & Institutions

Letris – AI-driven literacy and writing assessment platform (Brazil)
JPAL (Abdul Latif Jameel Poverty Action Lab) – MIT-affiliated research institution; JPAL South Asia
Pratam Education Foundation – Major Indian nonprofit focused on foundational learning
Rocket Learning – Nonprofit using AI for early childhood education (India)
Google.org – Philanthropy arm of Google; funder and AI research partner
UNESCO – Recognized Letris as best education technology globally (2020)
World Bank – Education and development policy institution
IDB (Inter-American Development Bank) – Funding partner for Letris research expansion
USAID – Funder of research initiatives
Ministry of Education, Espírito Santo and São Paulo states (Brazil) – Government partners
Government of India – Implicit stakeholder in TARL, foundational learning policy
Anthropic – AI research partner on "Anytime Testing Machine" with Pratam
Google & affiliated research units – LearnLM, language datasets, speech models

Technical Concepts & Resources

AI & Machine Learning

LearnLM – Google-developed foundation model optimized for learning using pedagogical principles (curiosity, scaffolding, cognitive load management)
Gemini models – Google's core foundation models now incorporating learning-optimized features
Foundation models / Large Language Models (LLMs) – General-purpose models that can be adapted for specific tasks
Neural networks – Modeled after brain learning; used in pattern recognition and knowledge integration
Agentic AI / AI agents – Tools that can act autonomously to test hypotheses and optimize solutions in parallel
AI co-scientist – Tool for hypothesis generation and literature synthesis across scientific domains
Deep Think model – Google research tool for advanced hypothesis generation and early solution exploration

Educational AI Tools & Platforms

Letris platform – Analyzes student writing, provides personalized feedback, tracks progress for teachers and administrators
Paradhygi app – Pratam's voice-based literacy assessment tool; analyzes phonetic and visual reading errors
Shiksha Sadhadi ("Learning Buddy") – Rocket Learning's AI assistant for ANGANWADI workers; voice-based data entry and personalized classroom recommendations
APU (Apu the elephant) – Conversational AI character that assesses early childhood learning levels
Merlin & bio-acoustic models – AI tools for biodiversity identification (used by Shobini in bird watching analogy)
ASER (Annual Status of Education Report) – Pratam's simple, one-page literacy/numeracy assessment tool, now enhanced with AI analysis

Pedagogical Frameworks

Teaching at the Right Level (TARL) / Kamal (in Indian languages) – Evidence-based approach: assess by learning level (not grade), group heterogeneously, provide targeted activities, adapt in real time
Tech in TARL project – Data collection initiative to understand classroom dynamics, teacher planning, and implementation barriers
Anytime Testing Machine (ATM) – Proposed framework for moving beyond rigid curricula and exams toward open-ended learning exploration
Reading progression framework – Structured sequence of reading skills from letter recognition through fluency

Data & Research

RCT (Randomized Controlled Trial) – Gold standard for impact evaluation; Letris conducted 5-month RCT across 178 schools in Espírito Santo
Digitized data collection – Using apps, voice recording, photo capture to enable faster feedback and analysis than manual methods
National standardized exams (Brazil & India) – Used as external validation metrics and policy-relevant benchmarks
Qualitative implementation data – Teacher interviews, observation, planning documents to understand "how" and "why" change happens
Cost-effectiveness analysis – Letris (~$X/student in pilot → government contract); Rocket Learning ($1.50/child/year)
Interim outcome proxies – Faster measures (essays written, feedback quality, practice frequency) to approximate long-term learning impact before final test scores emerge
Pattern recognition & data interpretability – Using AI to surface non-obvious patterns (e.g., phonetic vs. visual error types) in large datasets

Language & Multilingual Efforts

Thousand Language Moonshot (Google) – Initiative to improve AI performance across 1,000 languages
Project Vani – Collaboration between Google and Indian government to build speech and language datasets across 773 districts
Language bias & representation – Open datasets in sub-Saharan Africa, East Africa, and India to improve model quality for underrepresented languages
Phonetic vs. visual error analysis – Technique for diagnosing specific reading challenges (e.g., similar-sounding letters, visually similar letters) in non-Roman scripts

Implementation Infrastructure

Offline-capable apps – Letris works without internet; can digitize handwritten work via photo
WhatsApp for education – Rocket Learning's primary platform (70% parental penetration); byte-sized videos for parents/ANGANWADI workers
Laptop & device provisioning – Government provision of devices to schools as prerequisite for edtech adoption
Curriculum mapping & integration – Ensuring AI tools generate activities aligned with state/local curriculum, not generic content
Teacher training & digital literacy – Practical, ongoing support for non-digitally native educators

Additional Context & Implications

Policy & Scale

Letris scaling: From pilot RCT (2019) → first government contract Espírito Santo (2022) → 10x growth in students, now 1,600+ schools, 500,000+ students, partnerships with multiple states
No government program using Letris has been cancelled or not renewed; all partners have improved performance
Brazilian state rankings: Espírito Santo jumped from 11th place (pre-Letris) → 1st place in writing (first year) → sustained top position
UNESCO recognition (2020) as best edtech globally accelerated policy adoption

Equity & Access

Private vs. public school gap in Brazil: Letris reduced gap by 9 percentage points on national exam
Early childhood: Rocket Learning reaches 1 in 4 villages in intervention areas; 75% school readiness vs. 50% national average
Foundational literacy: Pratam's work shows that without intensive intervention, gaps widen; TARL + AI diagnostics can equalize learning across caste, gender, income
Language barriers: Most AI models still underperform for Indian languages, minority languages, and low-resource contexts—critical gap for last-mile equity

Tensions & Open Questions

Speed vs. rigor: Fast tech change creates pressure to deploy before full RCT evidence; ecosystem using interim measures and cost-effectiveness to balance
Standardized exams vs. deeper learning: Letris improves writing exam scores but also tests for non-exam writing (poetry, articles, biography) to avoid teaching-to-the-test
Curriculum as constraint vs. guide: How much should AI personalization override local curriculum requirements?
Teacher autonomy vs. algorithmic guidance: How to ensure AI recommendations empower rather than micromanage teachers?
Costs of evidence: RCTs are expensive and slow; smaller, resource-constrained organizations often lack funding for rigorous evaluation, creating blind spots
Generalizability: Does Letris work equally well in other languages, cultural contexts, education systems?

Future Directions (Mentioned or Implied)

Expansion of Letris RCT to measure long-term economic impacts on college enrollment, employability, earnings
Broader application of AI to non-exam writing and deeper literacy skills
TARL + AI integration across more Indian states and contexts
Rocket Learning scaling beyond early childhood to primary education
Research on how AI can help humans better understand learning itself (reverse engineering learning from neural network behavior)
Ethical AI frameworks for public education systems with transparent, auditable algorithms
Multilingual and culturally adapted models as prerequisite for equitable AI scaling

Overall Significance: This summit demonstrates a maturing AI-for-education ecosystem that has learned that **impact requires evidence, local adaptation, pedagogical rigor, and intentional equity work—