1:1 device programs: what the evidence shows

Few school reforms have ever moved money as fast as the laptop. From the mid-2000s onward, states, districts and whole countries bought a computer for every child. Maine did it. Uruguay did it. Peru bought nearly a million machines. The pitch was simple and emotional: the digital world is coming, and a child without a device is a child left behind.

Notice what the pitch was not. It was not a claim about reading scores or maths scores. It was a claim about fairness and the future — and against that yardstick, buying hardware always succeeds, because the box arrives. Whether the box teaches anyone anything is a separate question. That question sat unanswered through the biggest purchasing wave in education technology history.

The answer now exists, and it is unusually crisp. This article walks through it. First the meta-analysis — a study that pools many studies — of school laptop programs. Then the giant randomized trial in Peru that tested hardware alone, and the home-computer experiments that back it up. Then the reason writing improved more than anything else, the wider map of what education technology actually moves, and what all of this means for the next ministry with a budget line called “devices”.

The biggest bet ever placed on hardware

The scale of the wager deserves a moment. One Laptop per Child alone shipped millions of its little green XO machines across the developing world. Peru was its flagship: roughly 900,000 laptops, bought by a national government for its poorest schools (Cristia et al., 2017). In the United States, 1:1 programs spread from Maine’s pioneering statewide rollout to thousands of districts. The pandemic then finished the job — device-per-child became the default almost everywhere that could afford it.

Through most of this, evidence played a small role. Early evaluations were mostly surveys and case studies. They measured enthusiasm, attendance, sometimes computer skill — rarely learning against a comparison group. The researcher who watched this most closely, Mark Warschauer, spent years inside laptop classrooms and put the pattern plainly: the machines amplified what schools already did, for better and for worse (Warschauer, 2006).

That word — amplify — turns out to be the key to the entire literature. Hold onto it. Every result below, positive and null, is a version of it.

The meta-analytic verdict

In 2016, Zheng, Warschauer, Lin and Chang gathered a decade of research on one-to-one laptop programs — close to a hundred published papers. From that pile they pulled out the small set with usable comparison groups (Zheng, Warschauer, Lin & Chang, 2016). Only about ten studies were rigorous enough to pool. That ratio is itself a finding. A reform adopted by thousands of schools had produced barely a handful of controlled tests.

The pooled results were positive. Modest, but positive. Achievement rose by roughly 0.15 standard deviations in English, around 0.16 in maths, about 0.20 in writing, and approximately 0.25 in science (Zheng, Warschauer, Lin & Chang, 2016). A standard deviation is the researcher’s common yardstick for learning gains — 0.2 on it is small but real, the sort of gain a good curriculum change produces. So: no revolution. A useful nudge.

The synthesis part of the review explains where the nudge came from. Students in laptop classrooms wrote more, revised more, got more feedback on drafts, and used the machines heavily for research and science projects. Teachers who integrated the devices into daily lessons saw the gains; teachers who parked them at the back of the room did not (Zheng, Warschauer, Lin & Chang, 2016). The subjects that improved most — writing and science — are exactly the subjects where the device changed what students actually did.

One caution before the numbers harden in your mind. The studies in the pool were mostly quasi-experimental — matched schools rather than coin flips — and mostly American. The authors are open about this. Read the effect sizes as the best available estimate for well-run programs in funded systems, not as a law of nature. What happens when the teaching layer is absent? For that, the literature has an answer of rare quality.

Peru: the perfect test of hardware alone

Most policy questions never get a clean experiment. This one did. Peru’s national OLPC rollout was evaluated by randomly assigning which rural primary schools received laptops first. Cristia and colleagues followed 319 schools — machines against no machines, decided by lottery (Cristia et al., 2017). It remains the largest randomized trial of an education technology program ever conducted.

The program delivered exactly what it promised, physically. Computers per student in treated schools jumped from about 0.12 to about 1.18 — a tenfold increase in access. The machines came loaded with books, a word processor, games and music tools. Most survived. Many went home with children. Access, the thing the program was named for, was achieved (Cristia et al., 2017).

Learning was not. After roughly fifteen months, maths and language scores in laptop schools were statistically indistinguishable from schools without them. The point estimates sat within a whisker of zero. No effect on enrollment or attendance either. The one bright spot was general cognitive skill: on Raven’s-style reasoning tests, treated children scored roughly 0.11 standard deviations higher — a modest gain, and one that never showed up in school subjects (Cristia et al., 2017).

Why nothing? The follow-up data is unsparing. Teachers received little training in using the machines to teach. The software did not map to the curriculum. Internet access was largely absent. Usage logs showed the laptops used often for word processing, games and music, rarely for structured practice in reading or maths (Cristia et al., 2017). Peru bought the amplifier and had nothing plugged into it.

It is hard to overstate how useful this null result is. It is not a study of a bad idea done badly. It is a study of the pure hardware theory — give children machines and learning follows — run at national scale, with random assignment, by a government that genuinely delivered the machines. The theory failed cleanly.

The home-computer experiments agree

Perhaps school was the wrong place to look. A parallel literature tested the same theory at home, and it triangulates the Peru result almost perfectly.

In Lima, a companion randomized trial gave XO laptops to children for home use. Computer proficiency rose. Academic achievement did not move (Beuermann et al., 2015). In California, Fairlie and Robinson ran the cleanest possible version: over a thousand schoolchildren without home computers, half given one free, outcomes tracked in school records. Ownership and use rose sharply. Effects on grades, test scores and attendance were tight, precise zeros — no gain, and no harm (Fairlie & Robinson, 2013).

Romania supplies the darker data point. A government voucher program subsidized home computers for low-income families, and its income cutoff created a natural experiment. Children just below the line got computers; near-identical children just above did not. Winners gained computer skill — and their school grades slipped. Homework time fell and game time rose. The negative effects were smaller in homes where parents enforced homework rules (Malamud & Pop-Eleches, 2011).

Three countries, three designs, one conclusion. Give a child a computer and you reliably produce one outcome: a child who is better at using computers. That is worth something. It is not the outcome the budget line promised.

Why writing moved most

Go back to the positive half of the split, because its internal pattern is just as telling. In the meta-analysis, writing shows the most consistent gains of any core skill (Zheng, Warschauer, Lin & Chang, 2016). Warschauer’s classroom studies had predicted exactly that, years earlier (Warschauer, 2006).

The reason is mechanical, not mystical. A laptop changes the economics of writing. Drafting is faster. Revision is cheap — no re-copying by hand. Work can be shared, commented on, and returned quickly, so feedback cycles that took a week now take a day. Students in laptop classrooms simply wrote more, in more genres, for more readers (Warschauer, 2006). The device removed friction from a high-value learning activity, and the activity expanded to fill the space.

That is what “amplification” looks like when it goes well. The laptop did not teach writing. It made an existing instructional loop — draft, feedback, revise — run faster and more often. Where no such loop existed, there was nothing to accelerate. Peru is the same sentence with the sign flipped.

The wider ed-tech map

Zoom out once more. In 2020, Escueta and colleagues reviewed the full experimental literature on technology in education — every randomized or quasi-random study they could find, across rich and poor countries (Escueta et al., 2020). Their map has four territories, and the borders are sharp.

Access alone — devices at school or at home, the territory this article has covered — shows little to no effect on learning. Computer-assisted learning software, by contrast, is one of the most promising interventions in the review, especially adaptive practice in maths, where several large trials report solid gains. Cheap behavioral nudges — texts to parents, reminders, information — deliver small effects at trivial cost. And fully online coursework tends to underperform in-person teaching, particularly for weaker students (Escueta et al., 2020).

An economics handbook survey of the same terrain, by Bulman and Fairlie, lands in the same place (Bulman & Fairlie, 2016). Across the quasi-experimental record, spending on hardware and connectivity by itself shows little achievement effect. Well-designed instructional software, by contrast, shows real promise. The pattern could not be more consistent. Technology helps when it carries instruction. It does nothing when it merely carries electricity.

What separated the programs that worked

Put the two halves of the literature side by side and the recipe almost writes itself. The programs that produced gains shared three features, visible across the synthesis studies (Zheng, Warschauer, Lin & Chang, 2016).

First, the device carried specific learning activities — writing with feedback, science investigation, structured practice software — rather than open access. Second, teachers were trained and supported in using it, and used it in daily lessons, not as a Friday treat. Third, someone watched what happened and adjusted. Maine’s long-running program, the best-studied US case, invested heavily in teacher development from the start; its writing gains are among the clearest in the record (Zheng, Warschauer, Lin & Chang, 2016).

The failures share the mirror-image features: hardware first, purpose vague, teachers untrained, usage unmeasured. The economists’ reviews make the same point in cost terms — the returns sit in the software and teaching layer, which is also the cheap layer (Escueta et al., 2020). A ministry that spends 95 percent of its budget on boxes and 5 percent on what runs inside them has built Peru again and should expect Peru’s results.

What the evidence doesn’t show

The device literature is unusually clear, but clarity has edges. Here is where the claims in this article stop.

• Few true school randomized trials. The positive pooled estimates rest mostly on matched-comparison designs, not lotteries; the cleanest randomized evidence is precisely the hardware-only kind that found nothing (Zheng, Warschauer, Lin & Chang, 2016).
• The pool is dated and US-heavy. Most pooled studies predate tablets, cheap broadband and modern learning software; how far the estimates travel to other systems and decades is uncertain (Zheng, Warschauer, Lin & Chang, 2016).
• Implementation is inferred, not randomized. No trial has randomly assigned “devices with teaching” against “devices without” head to head; the split this article draws comes from comparing across literatures (Escueta et al., 2020).
• Long-run outcomes are unknown. Follow-ups end within a year or two; whether childhood device programs change graduation, employment or later skills is essentially unmeasured (Bulman & Fairlie, 2016).
• Cognitive gains are a puzzle. Peru’s reasoning-test improvement is real but modest, and nobody has shown it transfers to school learning or persists (Cristia et al., 2017).
• Cost-effectiveness is rarely computed. Even positive programs seldom report gains per dollar, the number a ministry actually needs (Escueta et al., 2020).

Where the evidence stops

1. 1Few true school randomized trials
2. 2The pool is dated and US-heavy
3. 3Implementation is inferred, not randomized
4. 4Long-run outcomes are unknown
5. 5Cognitive gains are a puzzle
6. 6Cost-effectiveness is rarely computed

© 2026 FUTURE PROOF™

Buying devices by the evidence

Read as one body of work, the literature converts into a short procurement manual — for a school board, a head teacher, or a ministry.

Budget for the instructional layer first. Decide what the devices will carry — which software, which subjects, which daily routines — before a single box is ordered. The returns live there, and that layer is cheap relative to hardware (Escueta et al., 2020). If the plan is “access”, the honest forecast is Peru (Cristia et al., 2017).

Aim the device at friction. The clearest school gains came where laptops removed friction from a proven learning loop — drafting, feedback, revision in writing; investigation in science (Zheng, Warschauer, Lin & Chang, 2016). Ask, activity by activity: what does the machine make faster or more frequent? If the answer is nothing, the machine is furniture.

Train teachers as the main line item, not the afterthought. Programs with sustained teacher development produced the gains; programs without it produced the nulls (Zheng, Warschauer, Lin & Chang, 2016). A workable rule of thumb from the successful cases: plan as much recurring spend on people and software as one-off spend on machines.

Put structure around home use. Unstructured home access moved grades nowhere in the US and downward in Romania — except where parents imposed rules (Malamud & Pop-Eleches, 2011). Devices that go home should carry assigned work, and parents should be able to see it. Guardrails are part of the intervention, not an accessory.

Measure use and learning, not distribution. Peru’s program hit 100 percent of its delivery targets and 0 percent of its learning ones (Cristia et al., 2017). Track what runs on the devices weekly, and test learning against a comparison group from day one. Boxes delivered is a logistics metric. It is not an education metric.