Will AI in education succeed?

In May 1959, the first PLATO (Programmed Logic for Automated Teaching Operations) system was unveiled at the University of Illinois. This means that last month was the 67th anniversary of the birth of Computer-Assisted Instruction (CAI). Such an event merits a look at technology in education today.

Many people are currently either for or against generative artificial intelligence (AI) in education—a stark binary that misses the complexity and inevitability of the current technological revolution. The success or failure of using technology for education (EdTech) is only partly about the technology itself. What matters more are the conditions underlying EdTech. EdTech is a tool, not a standalone solution, and without a proper support system in place, it will not succeed.

Within the context of low- and middle-income countries (LMICs), what conditions have determined whether past EdTech iterations succeeded or failed—and are those conditions different for AI-driven interventions today?

Three conditions emerge consistently from over two decades of rigorous evidence. None is unique to AI, but each is now being tested again by current AI rollouts.

1. Whether the technology was deployed to substitute for teachers or support them

Muralidharan, Singh, and Ganimian (2019) found that the Mindspark personalized learning program in Delhi raised math scores by 0.37 standard deviations and Hindi by 0.23, but did so as an after-school complement with a teaching assistant—not as a replacement for classroom instruction. An earlier evaluation in Gujarat by Linden (2008) showed the same software flipping from a loss when used in place of teacher-led time to a gain when used as an out-of-school complement. The difference came down to a single design choice: Whether the software replaced teaching time or supplemented it. Early AI evidence points the same way. Two rigorous LMIC randomized controlled trials (RCTs) of generative AI tutors in Nigeria and Ghana both demonstrated gains when teachers remained present and involved. The question for AI is not whether the model is good enough to teach alone; it is whether the school is organized enough to use it effectively as an instructional complement.

2. Infrastructure and equity

Peru’s One Laptop Per Child program distributed about 900,000 devices over a decade but produced no measurable effect on math or reading achievement, enrollment, or lasting attainment (Cristia et al. 2017; Cueto et al. 2025). The problem wasn’t the hardware; it was the absence of electricity, connectivity, teacher capacity, and curricular integration. AI raises that infrastructure bar rather than lowering it, requiring reliable bandwidth, sufficient processing power to run AI models, and capable devices. In 2022, only about 40% of primary schools and 50% of lower secondary schools globally had internet access (UNESCO GEM Report 2023). The schools that fell furthest behind during COVID-19 closures are the same schools AI deployment will struggle to reach.

3. Whether the field has been disciplined about measuring effectiveness

The Global Education Evidence Advisory Panel’s 2023 review of more than 13,000 studies classified eight LMIC interventions as “Great Buys” or “Good Buys” for improving learning. None is an EdTech reform. Providing more devices and platforms without complementary changes in teacher practice and curriculum actually sits in the “Bad Buys” tier. The discernible pattern is not that EdTech cannot work; it’s that EdTech has been deployed at a national scale based on inputs delivered or the number of users reached, rather than evidence of learning gained. AI rollouts are accelerating ahead of evidence in the same way. The relevant question here is whether ministries and funders have learned to require pre-registered, independent evaluation as a condition of procurement—or whether vendor pilots will again outrun what is actually known.

So, what is the best path forward?

We must acknowledge that AI cannot replace teachers; we must ensure sufficient infrastructure and capacity, especially in hard-to-reach areas; and we must properly evaluate tech reforms’ real impact on learning. We offer three additional recommendations.

• Develop clarity around the purposes of education. In any location, education must balance several, sometimes competing, purposes such as social mobility, economic development, democratic equity, and nation-building. Until there is sufficient stakeholder consensus on what education is for, it will be difficult to adopt the right EdTech approach and align it with curricular and assessment changes that fit identified education priorities. Without coordinating the parts of an education reform, AI will further fragment the landscape it’s meant to improve.
• Build sufficient decisionmaking capacity. Teachers, families, and students need digital literacy and support to use EdTech effectively, but so do government leaders. The recent “Millions Learning study on EdTech decisionmaking in South and Southeast Asia”—the written brief of which will be published next month—found many policymakers wishing for more time, expertise, and useful evidence as they decide how to move their systems toward a digital future and choose appropriate EdTech applications to pilot and scale. Facing significant political, economic, and marketing pressures to embrace AI quickly, they want to get it right—but will struggle until given sufficient time, expert support, and relevant data.
• Recognize that AI may be best used for institutional support work, not classroom instruction. AI’s true value may lie in carrying out tedious, repetitive tasks like information management and education administration, freeing up teachers and school leaders to focus on students and collaboration. We believe that the core of successful education remains committed professionals working closely with learners and families in supportive environments. AI’s real promise may lie in making that more possible than ever.