AI in Education: Customising Learning for All

Detailed Summary

Samantha Carter opened the session, introducing herself as the Policy Manager for AI Evidence at J‑PAL and noting her background in AI policy and development economics. She welcomed Thiago Rached to present the work of Letras, an AI‑driven literacy platform operating in Brazil.

1.1. The Literacy Challenge in Brazil

• Adult literacy: Roughly 90 % of Brazilian adults are not fully literate, unable to read high‑school textbooks or write a one‑page essay.
• School disparity: Public‑school students lag far behind private‑school peers on the national high‑school exam, which determines university admission.
• Core problem: Lack of regular writing practice and poor feedback—students receive only a score, not actionable guidance.

1.2. The Letras Solution

• AI‑based essay‑grader that provides personalised, in‑depth feedback on each student’s essay.
• Data pipeline: Collects writing samples, analyses them, and delivers structured insights to students, teachers, and system managers.
• Goal: Move the student to the centre of the learning process by supplying continual diagnostic data.

1.3. Why Conduct an Impact Evaluation?

• Internal data showed promise but was biased and not scientifically validated.
• In Brazil, government procurement is highly bureaucratic; an RCT (randomised controlled trial) was needed to prove impact and secure the first public‑sector contract.

1.4. The Randomised Evaluation (Espírito Santo)

Aspect	Details
Design	5‑month RCT, 178 schools (110 treatment, 68 control).
Intervention	Treatment schools received the Letras platform; control schools continued the status‑quo.
Outcomes Measured	• Number of practice essays • Quality of feedback • Essay scores on the national exam • Writing skills on alternative narrative tasks • Teacher time‑allocation
Partners	Letras, J‑PAL, FGV (Brazilian university), Lemon Foundation.
Key Findings	• 95 % of teachers assigned essays; 82 % of students submitted essays (high engagement). • 35 % increase in teacher‑student one‑on‑one engagement. • +17 points on writing test scores. • 9 % reduction in the public‑vs‑private gap on the national exam (closing the equity gap).
State‑Level Impact	In the study, public‑school students in treatment schools moved from below‑average to second‑place in the statewide ranking of average scores.
Recognition	Won the UNESCO Prize for Education Technology (2020) – the most prestigious global award for ed‑tech.
Policy Scale‑up	After the pilot, Espírito Santo climbed from 11th to 1st in the national exam ranking within a single school year and retained the top spot thereafter. Other states followed, adopting Letras as the official writing‑reading programme for both middle and high schools.
Reach	Currently deployed in > 1,600 schools, serving ≈ 2 million students cumulatively; the active cohort is ≈ 500 k students.

1.5. Lessons on Implementation

• Beyond the algorithm: Success stems from implementation expertise, not only from the AI model.
• Infrastructure flexibility: The platform works offline, digitises handwritten essays via photo capture, and adapts to schools lacking computers or internet.
• Curriculum integration: Content aligns with state curricula, ensuring teachers can pair AI feedback with textbook‑based lessons.
• Administrative dashboards: Provide real‑time data to education secretaries for targeted training, curriculum tweaks, and resource allocation.

1.6. Future Research

• A forthcoming RCT will explore long‑term economic outcomes: linking literacy gains to higher‑education enrollment and employment.
• Funding sources include USAID, IDB, and private philanthropy (Google.org).

Samantha Carter thanked Tiago and transitioned to the panel discussion.

---

2. Panel Discussion (≈ 45 min)

Moderator: Margaret Clarke (World Bank)

Panelists: Brigitte Hoyer‑Gosselink (Google.org), Rukmini Banerjee (Pratham), Namya Mahajan (Rocket Learning), with Sam (J‑PAL) providing occasional prompts.

The discussion centered on how AI can be leveraged at scale, bias mitigation, personalisation, and research support. Audience questions were interleaved throughout.

2.1. Pratham’s Data‑Driven AI Applications

• ASER & Simple Tool: Long‑standing assessment framework in India that collects paper‑based classroom data.
• Voice‑data collection via the Padhai app: Captures children’s pronunciation errors, enabling phonetic vs. visual mistake analysis.
  • Phonetic mistakes (confusing similar‑sounding letters) led to greater learning gains than visual mistakes (confusing similar‑looking letters).
• Adaptive instruction: Insights fed back to curriculum design (e.g., adjusting reading texts, phonological exercises).
• Tech‑Intal project: Ongoing data collection on teacher planning, classroom execution, and student progress, awaiting an RCT.

Key Insight: Data‑rich diagnostics can directly inform instructional adjustments, even before AI models are fully mature.

2.2. Google.org’s Emerging AI for Learning

• Model Development: Creation of LearnLM, incorporated into the Gemini family, designed to embody pedagogical best practices (curiosity‑driving, step‑by‑step guidance, avoiding cognitive overload).

• Evaluation Framework: Working with educators to define learning‑oriented metrics (e.g., curiosity prompts, scaffolded reasoning).

• Bias Auditing:

  • Ongoing efforts to ensure cultural, racial, gender fairness across 773 Indian districts and globally (1,000‑language “moonshot”).
  • Partnerships with Project Vani and Karya to assemble speech‑language datasets for under‑represented languages.
  • Emphasis on open‑data releases to accelerate inclusive model training.

• Application Outlook:

  • Personalisation (adaptive learning pathways).
  • Cost‑effectiveness: Prioritising low‑risk, cheap interventions that can be piloted rapidly while evidence accumulates.

Key Insight: AI models must be rooted in education science and continuously audited for bias to be trustworthy at scale.

2.3. Rocket Learning’s WhatsApp‑Based Early‑Childhood Scaling

• Target Audience: Anganwadi workers (government community health/education workers) and parents.

• Delivery Mechanism: WhatsApp (ubiquitous in India) to send bite‑size video lessons.

• Engagement Stats: 75 % of Anganwadi workers use the content weekly; 75 % of children reach school‑readiness, versus a 50 % national average.

• Cost: $1.50 per child per year – markedly lower than other interventions.

• AI‑Powered “Learning Buddy”:

  • Voice‑based data entry: Workers dictate observations (“Brigitte excelled in verbal but struggled in maths”), reducing manual entry burden.
  • Personalised alerts: System flags specific children for targeted activities in subsequent lessons.

• Learning Ladder: A hierarchical competency map (≈ 100 learning competencies derived from 10 core outcomes) used by an elephant‑avatar chatbot (Appu) to diagnose child levels via conversational interaction.

Key Insight: Combining low‑tech platforms (WhatsApp) with AI‑driven diagnostics yields high‑impact, cost‑efficient scaling in resource‑constrained environments.

2.4. Audience Questions & Panel Responses

Question	Speaker(s)	Core Points
Bias in AI models (cultural, gender, racial)	Brigitte	• Ongoing bias‑audit teams; • Need for multilingual data; • Collaboration with governments and civil‑society for open‑data.
Personalisation via WhatsApp (parent usability)	Namya	• Voice‑based entry reduces effort; • Appu chatbot guides parents; • No extensive training needed.
Future of AI in research (supporting hypothesis generation)	Brigitte	• AI Co‑Scientist builds knowledge trees from literature; • Deep‑Think models assist hypothesis crafting; • Emerging agentic tools execute scalable experiments (especially in hard‑science).
Human role in an AI‑augmented future (philosophical)	Rukmini	• Human planning, trust, and judgment remain central; • AI as augmentation, not replacement; • Emphasis on breaking rigid curriculum constraints to unleash creativity.

2.5. Concluding Remarks

• Consensus: AI can accelerate learning gains, but evidence, ethics, and contextual adaptation are non‑negotiable.
• Call to Action: Continue rigorous RCTs, bias monitoring, and partner‑driven scaling to ensure AI benefits all learners, especially the most vulnerable.

Key Takeaways

1. Evidence‑first approach – The Letras RCT demonstrated that an AI‑driven essay‑grader can raise writing scores by 17 points and shrink public‑vs‑private gaps by 9 %, securing government contracts and national policy adoption.

2. Implementation matters more than the algorithm – Offline‑first design, curriculum alignment, and administrative dashboards were critical to scale Letras across 1,600+ schools.

3. Data‑rich diagnostics enable rapid instructional improvement – Pratham’s voice‑data analytics identified phonetic vs. visual error patterns, informing targeted curriculum adjustments.

4. Bias mitigation requires multilingual data and open‑source collaboration – Google.org’s LearnLM / Gemini models are built with pedagogical metrics and undergo continuous bias audits across hundreds of languages.

5. Low‑tech platforms + AI deliver high impact at low cost – Rocket Learning’s WhatsApp‑based videos and AI “learning buddy” achieve 75 % school‑readiness at $1.50 per child per year.

6. Human judgement remains central – Across all speakers, AI is framed as an augmentation tool; teachers, parents, and policymakers keep the planning, trust, and contextual insight.

7. Future research pipelines – Upcoming RCTs will assess long‑term economic outcomes (higher‑education enrollment, employability), while Google.org explores AI Co‑Scientist for hypothesis generation and agentic research assistants.

8. Scalable, evidence‑driven AI in education demands a three‑pronged strategy: (i) Rigorous impact evaluation, (ii) Ethical, bias‑aware model development, and (iii) Contextual, incentive‑aligned implementation.

---

See Also:

• ai-2-0-reimagining-indian-education-system
• keynote-i-to-the-power-of-ai-an-8-year-old-on-aspiring-india-impacting-the-world
• equipping-the-judiciary-to-harness-the-power-of-ai
• evaluations-and-open-source-software-for-ai-for-social-good-at-scale
• safe-and-trusted-ai-the-ethics-and-governance-perspective
• building-indias-ai-governance-architecture-from-frameworks-to-implementation
• ai-horizons-building-safe-and-trusted-ai
• multi-stakeholder-collaboration-to-foster-ai-adoption-in-the-global-south
• power-protection-and-progress-legislators-and-the-ai-era
• shaping-the-ai-narrative-trust-outcomes-and-responsibility