AI for social good: Helping students learn

More children are in school than ever before, but not learning. In crowded classrooms, one third grade teacher might face fifty students with different learning levels, home lives, and languages. She has to choose where to focus her limited time and attention in a curriculum that her students often can’t follow. Some students fall behind, while others fail to reach their potential. Day after day of showing up and not learning, many kids will stop showing up at all. Unsupported, overwhelmed teachers follow suit.

This situation, repeated in millions of schools around the world, constitutes a global learning crisis. Can AI tools help? It has the potential to further personalize lessons, speed up student data analysis, deepen feedback on student performance, and make evidence-informed education programs easier to deliver at scale. 

Or, AI could become another “magic bullet” that fails to improve learning, leaving governments to spend scarce resources on ed tech that sits unused in schools. 

It’s happened before. In Peru, for example, after a ~$180 million investment in student laptops and ten years of iteration, an evaluation found that  students with laptops were no better off—in fact, they were marginally worse off. 

Multiple studies have shown that dropping technology into classrooms is not enough to improve learning; technologies have to integrate with teaching and learning practices. The danger of tech-based magic bullets is that they often come with a wave of hype: when the tech is the hammer, every problem in education starts to look like a nail. 

One click away on students’ browsers, AI has already arrived in many classrooms and students have essentially been participating in an unregulated experiment with their education. In Turkey, reliance on unregulated AI harmed student learning, by robbing students of longer-term critical thinking skills, as they used GPT to get quick answers to practice questions. 

To keep pace with AI,  policymakers, educators, and researchers should leverage existing evidence and avoid following every use case down the rabbit hole—especially those built on top of ineffective theories of change. Empowered with evidence, education systems can drive investments towards AI uses that will supercharge what we already know is likely to work. With this approach, we reduce the risk of wasting teachers’ time, students’ attention, and governments’ money. 

Early evidence on promising AI educational tools

AI offers promising ways to strengthen education systems and improve learning outcomes when integrated thoughtfully into tailored instruction, parent outreach, effective reading instruction, and other smart buys.

Assessing and providing feedback to students: The results of tests and quizzes can show teachers and students where students stand, but adding frequent, high quality feedback shows students how to reach the next rung. However, providing personalized feedback on piles of worksheets and papers takes time, during which student learning is stalled. 

Here, there are bottlenecks that AI can help clear. In Brazil, researchers evaluated an AI-powered essay grader which gives immediate feedback on students’ practice essays for a national exam. Students with the AI grader improved their writing through more practice reps, more useful comments, and more 1:1 attention from teachers, who could now spend less time parsing through students’ syntax. The AI program was trained on thousands of exemplar essays, several human grader scores and comments, and standardized grading criteria from the national exam coordinator. The program is now being scaled up in government schools across Brazil. 

AI voice recognition technology also has the potential to quickly assess a child’s reading skills in local languages. Ongoing evaluations of Facebook’s voice recognition software in South Africa and Pratham’s Padhi app in India will offer future insights. Many AI photo recognition technologies, like Photomath, are also widely available to check math work by scanning a photo and following the step-by-step explanations. However, we need to understand more about how students are using it: are they internalizing and learning from these quick checks, or are they simply copying and pasting an AI answer to complete the assignment?

Personalizing lessons to students: We know that tailoring instruction to a student’s learning level increases learning, but personalization takes teacher expertise and time. One ed tech advancement that is already building on tailored instruction evidence is student-facing Computer Adaptive Learning (CAL) software. CAL systems, like Mindspark, adjust which lesson/question comes next based on whether the student is answering correctly. 

CAL studies show substantial and consistent learning gains, particularly for lower-performing students, when schools use technology to complement classroom teaching and deliver targeted instruction. When schools use ed tech to replace classroom teaching, the results are mixed. 

While existing evidence on CAL has studied rule-based software with pre-written content, a growing body of research is evaluating AI tutors. These tools generate and personalize responses to students in real time, have also shown promise in early studies, such as after-school tutoring with Microsoft CoPilot in Nigeria, access to the Rori tutoring chatbot during study hall in Ghana, and the Khanmigo tutor in Canada. Ongoing research on Darsel’s math chatbot in Jordan and the mEducation chatbot in the Philippines will shed light on additional contexts and tools. Just as adult supervision is important for CAL’s effectiveness, emerging evidence suggests that AI instruction still needs a human in the loop to motivate and hold students accountable. 

As with many applications of AI, open questions remain about how much time users should spend with AI tools, as well as what kinds of human, social interactions they are replacing.

Leveraging detailed information to support parents: We know that sharing information that guide parents through school applications and alert them if their child is falling behind, allows them to better support their child. For example, several interventions of integrated chatbots, show that they can help with a range of tasks: from finding information about financial aid to finding mistakes in the college applications. Integrated AI chatbots may be able to help parents navigate a complex process, like finding a school for their child.

Untapped opportunities for AI in education

There are some intractable problems in education that don’t yet have good, scalable solutions. AI could help resolve some of these, but ideas are largely untested and more research is needed.

Data-driven decision-making: AI can analyze data quickly and make accurate predictions with good data. This could help teachers identify learning gaps and sort students into groups to provide tailored instruction. In Italy, researchers are evaluating whether an AI algorithm can close gender bias in STEM by giving teachers more informed recommendations for math tracks for incoming high schoolers. (Results are forthcoming.) 

AI may also help districts allocate resources more effectively, such as targeting scholarships or administering school-based health interventions. Targeting is often the most expensive part of these sorts of interventions, so AI-powered predictions of who is most in need could help systems if they rely on trusted, unbiased sources. 

Adapting structured pedagogy and tailored instruction for greater scale: Scaling programs requires lots of tweaks. Resources are often needed to translate materials into local languages, adjust curricula to different school calendars, and adapt old systems. In Uganda, curriculum developers used ChatGPT to translate teacher guides into local languages, cutting costs without losing quality. 

In Kenya, EIDU is rolling out a new remedial learning tool that uses AI to identify learning gaps and create tailored lessons for small groups. Partnering with EIDU, researchers are evaluating the most effective ways to implement the new tool in the classroom, like combining it with dynamic grouping similar to Teaching at the Right Level. 

Making materials easier to understand and more relevant: When students don’t understand what a question is asking, they may guess, feel defeated, and give up. This is especially true for language learners and students with learning disabilities. One emerging program, M7E AI, helps curricular developers revise math problems into clear, accessible language, but has not been evaluated at scale. Similarly, a good tutor uses analogies that resonate for their students. Generative AI can potentially unlock student’s interests and creativity. However, we need more evidence to make sure models that develop new, untested content are reliably accurate and effective.

Assessing and coaching teachers: Teachers need personalized, continuous feedback too. As new pedagogical programs and reforms roll out, education systems need a way to train and support teachers at scale, but in-person coaching is expensive. Research has not yet cracked the code on the most cost-effective ways to train teachers–though one evaluation in South Africa found that virtual coaching was less effective than in-person coaching at improving and sustaining student learning outcomes. When coaches are in the room, teachers get practical, observation-based feedback and build the trust and accountability needed to try new approaches. Whether AI can deliver the same value remains to be seen. 

One ongoing evaluation uses an AI program to analyze recordings of student-tutor interactions, and offers a way to deliver feedback to the teacher that was previously hard to observe from afar, like how to use warmer language with a student. 

Selecting, allocating, and retaining teachers: There are many unanswered questions when it comes to teacher hiring and retention. Quasi-experimental evidence from the United States suggests that selective evaluation systems that incentivize effective teachers and remove ineffective ones can improve student learning outcomes. However, many allocation systems rely on favoritism and random placements, which means teachers are often assigned to classrooms without regard to whether they are a good fit for the students. 

Some school systems that have tried to develop merit-based hiring systems have struggled to accurately predict who would make a good teacher. AI has the potential to improve predicting who will be a good teacher and increase student learning, and thus more efficiently match teachers with schools, but this approach is untested.

Filling gaps in school-provided mental health: AI may have the potential to offer emotional support tools where school counselors are scarce, particularly in rural areas. However, there are also risks: When AI substitutes humans as children’s companions, they could worsen a child’s mental health. More research is needed to find the right balance here. 

Cautions and caveats: AI is not a silver bullet

There are several risks specific to the education sector that policymakers and researchers should keep in mind when considering AI solutions. 

Erosion of critical thinking: Students may rely on AI for quick answers, reducing opportunities to develop reasoning, problem-solving, and creativity. Many AI systems prioritize efficiency over pedagogy, offering solutions instead of guiding inquiry. Without guardrails, they risk undermining deep learning.

Loss of social interaction: Increased screen time with AI tutors or companions can crowd out peer collaboration and teacher-student engagement, weakening socio-emotional skills like how to collaborate with others. 

Safeguarding: Chatbots designed to mimic intimacy can blur boundaries, especially for adolescents, and have been shown to produce inappropriate or unsafe responses. Additionally, AI systems handling sensitive student data require strong protections. Without safeguards, data could be misused and students could be exposed to harmful content.

Overburdened teachers: If AI tools are poorly integrated, they can add complexity rather than reduce workload, increasing stress without improving outcomes.

Equity: Unequal access to devices, connectivity, and AI tools could widen existing disparities between students and schools. Models trained on narrow datasets may provide examples or guidance that are irrelevant—or even harmful—in diverse contexts.

Dependence on unverified sources: AI models drawing on the open internet may provide inaccurate or misleading information unless aligned with trusted curricula.

Looking forward

Both J-PAL’s Learning for All Initiative and Project for AI Evidence fund new evaluations of AI solutions that aim to address these unanswered questions. We’ll continue to share new research results as more findings emerge. And be on the lookout for J-PAL’s upcoming AI Evidence Playbook, which will summarize what we know and don’t know about AI’s role across sectors and share perspectives on building on these lessons.