Skip to content

India, and video, live and self-study teaching

Part of the AI coach reading library: 21 papers cited in the AI coach notes, each in plain language. Who was studied, how, what it found, why it matters for our students, and where our notes use it. Avanti’s own trials and Indian evidence, video and broadcast lessons in richer and poorer countries, evaluation, and India’s data law. Download links open the archived copy and need an Avanti Google login; the Source link is the public original.

The overviews are drawn from the extraction tables and literature reviews written when each paper was read, not from a fresh reading. Where a number here and in a note differ, the note’s section is the one to check.

India: Avanti’s own trials, classrooms and language

Section titled “India: Avanti’s own trials, classrooms and language”

Too Little, Too Late: Improving Post-Primary Learning Outcomes in India. Did Avanti’s blended after-school programme in Chennai move the Class X board exam?

Section titled “Too Little, Too Late: Improving Post-Primary Learning Outcomes in India. Did Avanti’s blended after-school programme in Chennai move the Class X board exam?”

Chiplunkar, Dhar and Nagesh 2020, working paper. Download · Source

  • Who was studied: Grade 9 students in 24 Chennai Corporation government schools (12 treatment, 12 control): 1,044 at baseline and 880 at endline. The programme was Avanti’s own.
  • How: Randomised controlled trial (RCT: schools assigned by public lottery). For one year, after school, students got 15 five-minute Tamil videos a week, peer worksheets and short facilitator lectures. Outcomes were the ASER basic reading and maths tests at the end of Grade 9 and the Class X board exam about a year later.
  • What it found: Large gains on basic skills: ASER maths +0.86 SD and reading +0.52 SD (SD: standard deviations, the usual unit for comparing test-score gains). No gain on the board exam: Class X total -0.28 SD (SE 0.150, p = 0.064, not significant at 5%); maths -0.01 and science -0.21, both not significant. No effect on non-cognitive skills.
  • Why it matters for us: Avanti’s own evidence that big gains on foundational skills need not reach a grade-level, high-stakes exam, so the coach must be judged on JEE-level outcomes.
  • Where we use it: What works, Summary, as evidence that Avanti’s digital content has not moved board scores; Meta-analysis section 1 (Introduction), section 3.13 and Appendix B (claim 1), for the same point; India context and Medium in poorer countries strand pages.

Explaining the Productivity Paradox: Experimental Evidence from Educational Technology. Did Avanti’s Sankalp classroom videos raise maths and science in Haryana?

Section titled “Explaining the Productivity Paradox: Experimental Evidence from Educational Technology. Did Avanti’s Sankalp classroom videos raise maths and science in Haryana?”

de Barros 2022, working paper (J-PAL). Download · Source

  • Who was studied: Grade 9-10 students in 240 Haryana government senior secondary schools (80 per arm), n = 24,584, in Avanti’s Sankalp programme.
  • How: Pre-registered three-arm cluster RCT (whole schools assigned by lottery). In the technology arm, teachers blended their teaching with 2.5-minute Hindi curriculum videos on a smart TV and tablet, plus workbooks, training and coaching. A second arm got the workbooks without technology; the control carried on as usual. Paper tests in maths and science after about 10-11 months.
  • What it found: Technology against control: maths -0.15 SD (SE 0.06) and science -0.03 SD (0.05); the paper’s text gives 0.16 for maths. Observed instructional quality fell by 0.46 SD, and time on instruction fell 6-7 percentage points in both treatment arms. Effects differed by up to 0.41 SD across districts. Take-up was high.
  • Why it matters for us: Avanti’s largest trial of video in the classroom found harm, not help, so the coach should not take time or attention away from the live teacher.
  • Where we use it: What works, Summary, as evidence that digital content in class has not helped; Meta-analysis section 1 (Introduction), section 3.13 and Appendix B (claim 1); Medium in poorer countries strand page.

From Proof of Concept to Scalable Policies: Challenges and Solutions, with an Application. Why did Teaching at the Right Level work in some Indian states and not others?

Section titled “From Proof of Concept to Scalable Policies: Challenges and Solutions, with an Application. Why did Teaching at the Right Level work in some Indian states and not others?”

Banerjee et al. 2017, NBER Working Paper 22931 / Journal of Economic Perspectives. Download · Source

  • Who was studied: Government primary-school classes in Bihar, Uttarakhand, Haryana and Uttar Pradesh, across a series of trials of Pratham’s Teaching at the Right Level (TaRL), which groups children by level rather than grade.
  • How: A series of RCTs, compared across states with different ways of putting TaRL into schools.
  • What it found: When teachers got training and materials but no dedicated time, 0 to 4% of regular Bihar classes grouped children by level. In Haryana, with a reserved hour and monitoring, over 90% of schools did.
  • Why it matters for us: A coach added on top of the day with no protected slot and no one checking is likely to go unused; it needs a timetabled slot and a named adult.
  • Where we use it: India context strand page, for the design rule that the coach needs a protected slot and someone checking; listed in SOURCES.md.

MILU: A Multi-task Indic Language Understanding Benchmark. How well do language models answer exam-style questions in Indian languages?

Section titled “MILU: A Multi-task Indic Language Understanding Benchmark. How well do language models answer exam-style questions in Indian languages?”

Verma et al. 2025, arXiv preprint (v3). Download · Source

  • Who was studied: Not students. 42 large language models, tested on 79,617 exam-style multiple-choice questions in English and 10 Indian languages.
  • How: Benchmark: each model answers the same question bank and is scored on accuracy, by language.
  • What it found: The best model, GPT-4o, scored 81.75% in English, 78.62% in Hindi and 73.44% in Marathi. Language-specific Indic models scored near chance.
  • Why it matters for us: Accuracy drops outside English, more for Marathi than Hindi, so the coach should default technical content to English and check a sample of Hindi or Marathi explanations; Claude was not tested and the items are general knowledge, not JEE physics.
  • Where we use it: India context strand page, for the language design rule; listed in SOURCES.md.

Video, live teaching and self-study in richer countries

Section titled “Video, live teaching and self-study in richer countries”

How Does Distance Education Compare with Classroom Instruction? Does teaching at a distance work as well as the classroom?

Section titled “How Does Distance Education Compare with Classroom Instruction? Does teaching at a distance work as well as the classroom?”

Bernard et al. 2004, Review of Educational Research. Download · Source

  • Who was studied: Learners from school to adult, mostly in North America, in studies from 1985 to 2002: 318 achievement outcomes, N = 54,775.
  • How: Meta-analysis (a statistical pooling of many studies’ results) comparing distance education with classroom teaching, split into live remote teaching (synchronous) and self-paced recorded work (asynchronous).
  • What it found: Essentially no difference overall: g+ = 0.01 [-0.007, 0.033] (g is a standardised effect size, read like SD). Live remote teaching came out slightly worse (-0.10 [-0.15, -0.06]) and self-paced work slightly better (+0.05 [0.03, 0.08]). The spread across studies was very wide (-1.31 to +1.41), and was explained more by method and teaching than by medium.
  • Why it matters for us: The medium on its own decides little; what the coach asks the student to do matters more than whether it arrives on a screen.
  • Where we use it: What works, Summary, as evidence that distance teaching matched the classroom; Meta-analysis section 3.13 (“the medium on its own changes little”) and Appendix B (claim 1); Medium in richer countries strand page.

Video Improves Learning in Higher Education: A Systematic Review. Does video help university students learn, and when?

Section titled “Video Improves Learning in Higher Education: A Systematic Review. Does video help university students learn, and when?”

Noetel et al. 2021, Review of Educational Research (read as the accepted manuscript). Download · Source

  • Who was studied: University students in many countries, mostly in health sciences: 83 randomised studies with 7,776 students.
  • How: Meta-analysis of RCTs only, comparing video that replaced other teaching with video added on top of it.
  • What it found: Replacing other teaching with video: g = 0.28 [0.14, 0.42]. Replacing a teacher: 0.18 [0.02, 0.34]. Replacing text: 0.51 [0.27, 0.75]. Adding video on top of teaching: 0.88 [0.62, 1.13] (the abstract says 0.80). Video’s advantage vanished when the comparison class was more interactive (-0.07 [-0.38, 0.25]). Publication bias was detected, and no study was rated low risk of bias.
  • Why it matters for us: Video helps most as a supplement and against plain text, and loses its edge against interactive teaching, which argues for a coach that adds to the teacher rather than replaces them.
  • Where we use it: Meta-analysis section 3.13, for “the medium on its own changes little” (g = 0.28) and “adding video beats replacing the teacher” (g = 0.80); Medium in richer countries strand page.

Is It Live or Is It Internet? Do students learn as much from online video of a lecture as from the same lecture live?

Section titled “Is It Live or Is It Internet? Do students learn as much from online video of a lecture as from the same lecture live?”

Figlio, Rush and Yin 2013, NBER Working Paper 16089 (Journal of Labor Economics). Download · Source

  • Who was studied: 296 volunteer students in introductory microeconomics at a large US research university.
  • How: RCT: students were assigned to attend the lectures live or to watch internet video of the same lectures; outcome was exam scores at the end of the course.
  • What it found: Live students scored 1.44 points out of 100 higher (SE 1.21, not significant). The gap was larger for low achievers (4.05 points), Hispanic students (11.28) and male students (3.48).
  • Why it matters for us: Replacing a live teacher with recorded lectures costs little on average but hurts weaker students, who are the students Avanti serves.
  • Where we use it: Meta-analysis section 3.13, as one of the trials showing weaker students lose most when live teaching is replaced; Medium in richer countries strand page.

Distance Learning in Higher Education: Evidence from a Randomized Experiment. Does letting students watch lectures by live stream change exam results?

Section titled “Distance Learning in Higher Education: Evidence from a Randomized Experiment. Does letting students watch lectures by live stream change exam results?”

Cacault et al. 2021, IZA Discussion Paper 12298 (Journal of the European Economic Association). Download · Source

  • Who was studied: First-year economics and management students at the University of Geneva, Switzerland: 23,766 student-course-weeks.
  • How: RCT in which access to live-streamed lectures was assigned at random by student and week; outcome was the share of exam questions answered correctly for each week’s material.
  • What it found: No average effect (intention-to-treat 0.001, SE 0.005; intention-to-treat means the effect of being offered streaming, whether used or not). Low-ability students lost 1.9 points in their share of questions correct (-0.18 among those who used it); high-ability students gained 2.5 points (+0.25 among users).
  • Why it matters for us: Giving weaker students a way to skip the live class lowers their results, so the coach should not become a substitute for attending.
  • Where we use it: Meta-analysis section 3.13, as one of the trials showing weaker students lose most when live teaching is replaced; Medium in richer countries strand page.

Zooming to Class? Does an online class teach college students as well as the same instructor in person?

Section titled “Zooming to Class? Does an online class teach college students as well as the same instructor in person?”

Kofoed et al. 2024, IZA Discussion Paper 14356 (American Economic Review: Insights). Download · Source

  • Who was studied: 551 West Point cadets in introductory economics, US, during COVID-19.
  • How: RCT: each of 12 instructors taught both a live Zoom section and an in-person section, and students were assigned at random.
  • What it found: Online classes lowered final grades by 0.215 SD (SE 0.084) and the final exam by 0.147 SD. Students in the bottom quartile lost 0.267 SD.
  • Why it matters for us: Even a live online teacher does worse than the same teacher in the room, and the loss is largest for weaker students.
  • Where we use it: Meta-analysis section 3.13, as the headline trial for “weaker students lose most when live teaching is replaced”; Medium in richer countries strand page.

Active Learning Narrows Achievement Gaps for Underrepresented Students in Undergraduate STEM. Does active learning close the gap for low-income and minority students?

Section titled “Active Learning Narrows Achievement Gaps for Underrepresented Students in Undergraduate STEM. Does active learning close the gap for low-income and minority students?”

Theobald et al. 2020, PNAS. Download · Source

  • Who was studied: US undergraduates in science, technology, engineering and maths: 9,238 students in 15 studies, comparing underrepresented and low-income students with others, with the same instructor teaching both ways.
  • How: Meta-analysis of student-level data, comparing the exam gap under lecturing and under active learning.
  • What it found: The exam gap was -0.62 SD [-0.69, -0.55] under lecturing and -0.42 SD [-0.48, -0.35] under active learning, a 33% reduction. Only high-intensity active learning narrowed the gap.
  • Why it matters for us: How a lesson is run matters more than the medium, and active work helps the students who start furthest behind, as long as there is a lot of it.
  • Where we use it: Meta-analysis section 3.13, for “what happens in the lesson matters more than the medium” (the exam gap cut by a third); Medium in richer countries strand page.

Video and broadcast lessons in poorer countries

Section titled “Video and broadcast lessons in poorer countries”

Engaging Teachers with Technology Increased Achievement, Bypassing Teachers Did Not. Do expert videos work better played by the teacher in class or on students’ own tablets?

Section titled “Engaging Teachers with Technology Increased Achievement, Bypassing Teachers Did Not. Do expert videos work better played by the teacher in class or on students’ own tablets?”

Beg et al. 2022, NBER Working Paper 25704 (American Economic Journal: Economic Policy). Download · Source

  • Who was studied: Government middle-school students in Punjab, Pakistan: Grade 8 in 59 schools (2,622 test observations) for the classroom arm, and Grade 6 in 71 schools (3,058) for the tablet arm.
  • How: Two parallel cluster RCTs over about 4 months. In one, 9-minute expert videos were projected in class with the teacher present. In the other, the same type of content went on students’ own tablets. Outcomes were a researcher-designed maths and science test and the provincial PEC standardised exam.
  • What it found: Classroom videos: +0.30 SD on the project test (0.296, SE 0.130) and +0.27 SD on the PEC exam (0.269, SE 0.103). Tablets: -0.43 SD (-0.426, SE 0.149). Our papers quote 0.22 and -0.42, from the specification with fewer controls. Gains were largest in low-baseline schools and with less-experienced teachers. Cost about $9 per student for the classroom arm against $131 for tablets.
  • Why it matters for us: Technology the teacher runs can help, while content that bypasses the teacher can do harm, so the teacher should stay in the coach’s session.
  • Where we use it: What works, Summary, as evidence that lessons on students’ own tablets lowered scores; Meta-analysis section 3.13 (“adding video beats replacing the teacher”) and Appendix B (claim 1, “technology the teacher runs can help”); Medium in poorer countries strand page.

Effectiveness of Interactive Satellite-Transmitted Instruction. Can live lessons from a studio teacher raise learning in rural primary schools?

Section titled “Effectiveness of Interactive Satellite-Transmitted Instruction. Can live lessons from a studio teacher raise learning in rural primary schools?”

Johnston and Ksoll 2017, CEPA Working Paper 17-08 (Economics of Education Review, 2022). Download · Source

  • Who was studied: Grade 2-5 pupils in 147 rural primary schools in Ghana (70 treatment, 77 control): 12,848 student-round observations.
  • How: Cluster RCT. A teacher in an Accra studio taught live, two-way, by satellite, with a trained local facilitator in each classroom, replacing the usual subject lesson. Standard maths and reading tests over 1-2 years and one year after the programme ended.
  • What it found: Maths rose by 0.23 SD (SE 0.03), and was still 0.21-0.22 SD a year after the programme ended. Reading fluency did not improve (0.02, SE 0.03).
  • Why it matters for us: Broadcast lessons help where the alternative was weak teaching; that is not the position at a Centre of Excellence with strong live teachers.
  • Where we use it: Meta-analysis section 3.13, for “broadcast lessons help where there was little teaching”; Medium in poorer countries strand page.

Do School Electrification and Provision of Digital Media Deliver Educational Benefits? Did exam-preparation videos raise Tanzania’s national secondary exam?

Section titled “Do School Electrification and Provision of Digital Media Deliver Educational Benefits? Did exam-preparation videos raise Tanzania’s national secondary exam?”

Seo 2017, IGC Working Paper E-40308-TZA-2. Download · Source

  • Who was studied: Grade 11 (O-level) students in 164 secondary schools in northern Tanzania: 10,171 students.
  • How: Six-arm cluster RCT. Schools got solar power and TVs, with or without English or bilingual videos of lectures solving 10 years of past biology and geography exams. Outcome was the national O-level exam about a year later.
  • What it found: No clear effect. Solar-enabled programmes pooled: 0.052 SD (SE 0.039), 95% CI [-0.024, 0.129]; the paper rules out effects larger than 0.13 SD. Schools with solar and English videos watched 4.2 more hours of TV a week, with no exam gain. Cost $6.41 per student.
  • Why it matters for us: The design closest to exam-preparation video, at Grade 11, did not move the exam, which cautions against a coach built around watching worked papers.
  • Where we use it: What works, Summary, as evidence that exam-preparation videos moved the national exam by no more than 0.13 SD; Meta-analysis section 3.13; Medium in poorer countries strand page.

How Big Are Effect Sizes in International Education Studies? What size of learning gain is normal in poorer countries?

Section titled “How Big Are Effect Sizes in International Education Studies? What size of learning gain is normal in poorer countries?”

Evans and Yuan 2020, CGD Working Paper 545. Download · Source

  • Who was studied: 130 education RCTs with learning outcomes in low- and middle-income countries.
  • How: Meta-review: collects the effect sizes those trials report and describes their spread.
  • What it found: Median effect 0.10 SD; 90th percentile 0.38 SD; South Asia median 0.08 SD. Studies with over 5,000 students have a median of 0.05 SD, against 0.10 SD for smaller ones.
  • Why it matters for us: A realistic bar for the coach is about 0.1 SD, not the best results from small trials.
  • Where we use it: India context strand page, to set what a pilot can expect and the success bar (design rule 12); listed in SOURCES.md.

Using Randomization in Development Economics Research: A Toolkit. How many students does a trial need to detect an effect?

Section titled “Using Randomization in Development Economics Research: A Toolkit. How many students does a trial need to detect an effect?”

Duflo, Glennerster and Kremer 2007, NBER Technical Working Paper 333. Download · Source

  • Who was studied: Not a study. A methods toolkit for randomised evaluation.
  • How: Sets out the formula for the minimum detectable effect (MDE: the smallest true effect a trial of a given size can reliably find) and the design choices that change it.
  • What it found: The MDE scales with one over the difference in take-up between groups; using baseline scores as a covariate raises precision; spillovers between treated and control students bias the estimate.
  • Why it matters for us: By this formula a one-centre pilot with about 40 students can detect only about 0.7 SD, so the pilot should be designed to learn about use, not to prove impact on JEE.
  • Where we use it: India context strand page, as the basis for the power numbers and the within-student design; listed in SOURCES.md.

The Big-Fish-Little-Pond Effect on Academic Self-Concept: A Meta-Analysis. Do students in high-achieving schools think less of their own ability?

Section titled “The Big-Fish-Little-Pond Effect on Academic Self-Concept: A Meta-Analysis. Do students in high-achieving schools think less of their own ability?”

Fang et al. 2018, Frontiers in Psychology. Download · Source

  • Who was studied: 33 studies with 56 effects and N = 1,276,838 students.
  • How: Meta-analysis of how a school’s average ability relates to a student’s academic self-concept, holding the student’s own ability fixed.
  • What it found: Mean beta = -0.28 (a standardised slope: higher school average, lower self-concept); -0.32 in high school and -0.35 in Asian samples (7 effects). It measures self-concept, not exam scores.
  • Why it matters for us: JNV CoE students are selected twice, so any feature that compares them with their own cohort is likely to push self-concept down.
  • Where we use it: India context strand page, for the rule never to show within-centre ranks (design rule 9); listed in SOURCES.md.

What You Don’t Know… Can’t Hurt You? Does telling university students their rank change their grades?

Section titled “What You Don’t Know… Can’t Hurt You? Does telling university students their rank change their grades?”

Azmat, Bagues, Cabrales and Iriberri 2016, IZA Discussion Paper 9853 (Management Science, 2019). Download · Source

  • Who was studied: 977 students at a Spanish university.
  • How: RCT: treated students were told their decile rank in their cohort every six months for three years.
  • What it found: In the second year, GPA fell by 0.05 SD and students passed 0.4 fewer modules (0.1 SD). The effect faded by graduation. It was driven by students who had underestimated their rank. Satisfaction rose.
  • Why it matters for us: Showing rank is not neutral and can lower short-run performance, so the coach should express standing against cutoffs and the student’s own trend.
  • Where we use it: India context strand page, for the rule never to show within-centre ranks (design rule 9); listed in SOURCES.md.

Top of the Class: The Importance of Ordinal Rank. Does a pupil’s rank in class affect later achievement?

Section titled “Top of the Class: The Importance of Ordinal Rank. Does a pupil’s rank in class affect later achievement?”

Murphy and Weinhardt 2013, CEP Discussion Paper 1241 (Review of Economic Studies, 2020). Download · Source

  • Who was studied: Over 2 million pupils in England, from administrative records.
  • How: Observational: compares pupils with the same test score but a different rank within their primary school.
  • What it found: Moving from the 25th to the 75th percentile of rank within primary school raises later test scores by 0.2 within-pupil SD, working through self-concept.
  • Why it matters for us: Local rank itself affects later achievement, and a within-centre leaderboard gives half the cohort a low rank.
  • Where we use it: India context strand page, for the rule never to show within-centre ranks (design rule 9); listed in SOURCES.md.

The Digital Personal Data Protection Act, 2023. What does Indian law require before processing a child’s data?

Section titled “The Digital Personal Data Protection Act, 2023. What does Indian law require before processing a child’s data?”

Government of India 2023, Act No. 22 of 2023 (statute). Download · Source

  • Who was studied: Not a study. India’s data-protection law.
  • How: Primary statute, read for the sections on children.
  • What it found: A child is anyone under 18 (s.2(f)). Section 9 requires verifiable parental consent before processing a child’s data, bars processing likely to harm a child’s well-being, and bars tracking or behavioural monitoring of children. Breach of s.9 carries a penalty of up to Rs 200 crore.
  • Why it matters for us: The coach’s logging, and any plugin that records what students do on other sites, fall inside s.9.
  • Where we use it: India context strand page, for consent, logging and dropping the browsing-tracking plugin (design rules 3-5); listed in SOURCES.md.
Section titled “Digital Personal Data Protection Rules, 2025 (G.S.R. 846(E)). How must parental consent be verified, and what is exempt for schools?”

Ministry of Electronics and Information Technology 2025, regulation. Download · Source

  • Who was studied: Not a study. The rules under the DPDP Act, notified 13 Nov 2025.
  • How: Primary regulation, read for Rule 10, Rule 12 and the Fourth Schedule.
  • What it found: Rule 10 requires measures to check that the consenting parent is an identifiable adult, through details already held, details the parent gives, or a token such as DigiLocker. The Fourth Schedule, Part A, exempts an educational institution from s.9(1) and s.9(3) only for tracking and monitoring for its educational activities or child safety. Part B exempts creating an account only when it is used for email. The duty not to harm a child’s well-being has no exemption.
  • Why it matters for us: Whether Avanti or the JNV counts as the “educational institution” decides who may log what, and that needs a legal opinion.
  • Where we use it: India context strand page, for consent and the exemption question (design rules 3-4) and the conflict with the spec’s identity plan; listed in SOURCES.md.

Notification G.S.R. 843(E): commencement dates under the DPDP Act. When do the children’s-data rules come into force?

Section titled “Notification G.S.R. 843(E): commencement dates under the DPDP Act. When do the children’s-data rules come into force?”

Ministry of Electronics and Information Technology 2025, gazette notification. Download · Source

  • Who was studied: Not a study. The notification setting when parts of the DPDP Act take effect.
  • How: Primary notification, dated 13 Nov 2025.
  • What it found: Section 9 and Rules 3 and 5 to 16 come into force 18 months after 13 Nov 2025, around 13 May 2027.
  • Why it matters for us: The law does not yet bind a November 2026 pilot, but it will bind any scale-up in academic year 2027-2028.
  • Where we use it: India context strand page, for the compliance timeline; listed in SOURCES.md.