The Prompt You Write

In June 2024, first-year senior secondary students from nine public schools in Benin City, Nigeria, stayed after class to study English with a machine.² Floods in the rainy season, teacher strikes and after-school jobs broke their attendance.¹ Microsoft Copilot, running on GPT-4, waited on every screen and answered whatever each student typed. One student, Omorogbe Uyiosa of Edo Boys High School, called the tool a tutor that takes any shape, “depending on the prompt you write.”¹

After six weeks, a pen-and-paper test showed gains of about 0.3 standard deviations, which World Bank researchers equate to almost two years of typical learning.^1,2 Against a database of randomized trials in developing countries, the program outperformed four in five education interventions. Girls, who started behind the boys, posted the largest gains.^1,2

Uyi’s six words carry the central bargain of this age: the machine answers anyone, and the prompt decides what each person carries home. For three decades I reported the news after it happened, in newspapers and on television in Guyana and Canada. Today I report on a future that can still break two ways.

Machine intelligence opens a new era of human flourishing. Cost, skill and habit hold billions of people in a no man’s land between an old world of reaction and a new world of authorship. Global leaders can lay boards across that mud, and they must start with schools.

01The evidence for flourishing

Benin City joins a growing file of measured gains. Economists Erik Brynjolfsson, Danielle Li and Lindsey Raymond tracked 5,179 customer support agents as their employer rolled out a generative AI assistant. Agents resolved 14 percent more issues per hour on average, and the novice and low-skilled among them gained 34 percent.³ The authors found suggestive evidence that the assistant carried the best agents’ practices to the newest hires.³

At Harvard, lecturers Greg Kestin and Kelly Miller and their colleagues built an AI tutor on research-based teaching principles and tested it against active learning led by experienced instructors. In a randomized trial with 194 physics students, learners gained more than twice as much from the tutor, in a median of 49 minutes against a 60-minute class.⁴

In Toronto, St. Michael’s Hospital runs CHARTWatch, a system that reads patient data such as blood pressure and heart rate. It warns nurses and doctors hours before a patient slides toward crisis. A study in the Canadian Medical Association Journal found that unexpected deaths on the ward fell 26 percent after the system’s launch.⁵

Demis Hassabis, who shared the 2024 Nobel Prize in Chemistry for AI that predicts protein structures, sized the coming change for The Guardian: “10 times bigger than the Industrial Revolution.”^6,7 Microsoft counts more than 1.2 billion people who picked up AI tools in under three years, a pace that beat the internet, the personal computer and the smartphone.⁸

Anyone with a connection can now call a tutor, a researcher, an editor and a programmer into one conversation. An idea that once waited years for a firm or a grant can now reach a working draft before lunch. The promise rests on measurement, and the measurements keep arriving.

02Same machine, split outcomes

Kenya supplies harder evidence, from a five-month field experiment that researchers at Harvard Business School and the University of California, Berkeley ran with 640 Kenyan entrepreneurs. A random draw gave some of them a GPT-4 business mentor on WhatsApp.⁹ Across the whole group, revenues and profits showed no clear average effect, yet beneath that average the owners split. By the researchers’ estimate, high performers gained more than 15 percent, while low performers did almost 10 percent worse than owners without the mentor.⁹

The split came neither from the questions the owners asked nor from the advice they received. It came from which advice each owner chose and how each one put it to work.⁹ Strong performers used the mentor to find tailored fixes, such as another source of power during blackouts and a new car-wash detergent in demand. Weaker performers leaned toward generic moves such as cutting prices and buying ads.⁹

A team led by Harvard Business School researchers ran a sharper test with 758 consultants at Boston Consulting Group. On 18 realistic tasks inside the machine’s abilities, consultants using GPT-4 finished 12.2 percent more work, 25.1 percent faster, at more than 40 percent higher quality. Below-average consultants gained 43 percent against 17 percent for stronger peers.¹⁰ The researchers then picked a task just outside the machine’s reach. Consultants using AI reached the correct answer 19 percentage points less often than consultants who worked without the machine.¹⁰

The authors call that boundary a “jagged technological frontier,” because two tasks of similar difficulty can fall on opposite sides of the line. Elite professionals followed the machine past that line.

The same split runs across the planet. In early 2026, Microsoft’s AI Economy Institute found generative AI in use among 27.5 percent of working-age people in the Global North. The Global South reached 15.4 percent, while adoption in the North grew more than twice as fast as adoption in the South.¹¹ The International Telecommunication Union counts 2.2 billion people without the internet. In high-income countries, 94 percent of people use the internet, against 23 percent in low-income countries.¹²

Kristalina Georgieva, the IMF’s managing director, introduced the Fund’s 2024 analysis of AI and jobs with a warning: “In most scenarios, AI will likely worsen overall inequality.”¹³ Even Sam Altman, whose company built ChatGPT, warned in 2024 that without enough infrastructure AI could turn into “mostly a tool for rich people.”¹⁴

Carry that logic forward ten years. The machine answers every caller, but people who already ask well pull ahead each quarter. People who never learned to ask fall further behind with every unchecked answer they accept. Habit alone builds that default future.

03No man’s land

On 29 October 1917, near Hooge in the Ypres salient, a camera caught Australian artillerymen walking single file on a duckboard track through the shattered trees of Chateau Wood. The photograph carries Frank Hurley’s name, though historians now debate who pressed the shutter, and on which day.²⁸ Duckboards laid safe footing across ground that shellfire churned into mud. Between the opposing trenches of that war lay no man’s land, ground neither army could hold.

Billions of people now stand in a no man’s land of a different kind. Behind them lies an old trench of habit, a life spent reacting to whatever the day throws. Ahead lies the open ground the machine cleared. Three barriers churn the ground between them into mud.

The price

The first barrier costs money, even though the price of machine intelligence collapsed at the low end. Stanford’s AI Index tracked the cost of running a model as capable as GPT-3.5, the engine of the first ChatGPT. The price fell from US$20 per million tokens in November 2022 to 7 cents by October 2024, a more than 280-fold cut.¹⁵

The frontier still charges a toll, since Anthropic and OpenAI each sell a top consumer plan at US$200 a month.¹⁶ The World Bank classes a country as low income when its gross national income per person falls at or below US$1,175 a year.¹⁷ In such a country, a standard US$20 plan costs at least a fifth of annual national income per person, and the top plan costs more than twice that sum. Connection adds its own toll: the ITU finds mobile broadband unaffordable in about 60 percent of low- and middle-income countries.¹²

The skill

The second barrier lives in the head and the hand: the skill to ask the machine a precise question and to judge its answer. Uyi named the first half of that skill, and Kenya exposed the second half. Owners who lost ground asked much the same questions as the winners and faltered at the choosing.⁹

The machine does its strongest work in territory we cannot map for ourselves, and it keeps answering beyond that territory in the same calm voice. A person who cannot see the edge cannot tell a sound answer from a fluent mistake.

In a survey of more than 48,000 people across 47 countries, the University of Melbourne and KPMG found 66 percent using AI on a regular basis. Just 39 percent reported any AI training. Among people who use AI at work, 66 percent admitted relying on its output without checking accuracy, and 56 percent reported mistakes in their work caused by AI.¹⁸ Canada, my home, ranks 44th of those 47 countries for AI training and literacy, and 24 percent of Canadians report any AI training.¹⁹

The habit

The third barrier runs deepest: a culture that trains people to react. Microsoft’s 2025 telemetry found a meeting, an email or a chat breaking into its busiest users’ core workday about every two minutes.²⁰ A day like that trains the mind to answer the next ping instead of authoring the next move.

In July 2014, in my Stabroek News column, I urged Guyanese to stop trading insults and scapegoats and to roll up their sleeves as “solution-generators.” I warned that a nation that reacts instead of planning keeps stumbling into a default future.²¹ Twelve years later, the same reflex meets a machine.

The reflex also shapes how people meet AI: with fear instead of skill. Pew Research Center found that half of American adults report more concern than excitement about AI in daily life, against 10 percent who report the reverse.²² Pew also found 56 percent of AI experts expecting a positive effect on the United States over the next 20 years, against 17 percent of the public.²³ Canada’s new national AI strategy reports that half of Canadians regard AI as a threat to humanity.⁵ Specific risks call for specific safeguards. Vague dread teaches nobody to write a prompt or check a claim, and it leaves millions frozen in the mud.

04Cheap answers, costly questions

Put the studies side by side and a pattern surfaces. Where a system carried expert practice to the novice, the novice made the biggest gains. The support assistant brought top agents’ habits to new hires; the Harvard tutor carried research-based teaching into every exchange; the Nigerian program paired the machine with teacher support.^1,3,4 Where a person faced the machine alone, the strong pulled ahead and the weak lost ground.⁹ So the divide runs through the structure around a person, not through the machine.

Machine intelligence cut the price of an answer toward zero. It raised the price of a good question.Shaun Michael Samaroo

A blank chat box rewards the person who arrives with a clear question, a way to test the reply and a reason to act. The same box hands a person with a vague wish a fluent mistake.

A sharp critic will answer that prices keep falling and interfaces keep improving, so the market will close the gap without schools. Kenya answers that critic: every owner in the treatment group held the same assistant on the same app, and the gap between strong and weak performers still widened.⁹ Access puts a person at the edge of the new ground. Judgment carries that person across the mud.

The Brookings Institution raises a harder objection: its Center for Universal Education consulted 505 students, teachers, parents, education leaders and technologists across 50 countries over a year. The center concluded that, on the current trajectory, AI’s risks to students outweigh its benefits. The same authors wrote that the trajectory can still bend.²⁴ I read that finding as a verdict on drift, not on the machine. A student alone with a chat box can outsource the thinking. Students whose teachers train them to question the machine learn faster, as Benin City and Harvard showed.

I stake Qualped on that reading. In my May 2026 white paper on world-building, I argued that the next founders will come from the ranks of creative writers, gamers and imaginative problem-solvers.²⁵ Such people see a human pain, picture the world that heals it and use AI to build that world.

I write this Note with QwaiAI as my cognitive executive. It retrieves sources, compares options, tests claims and asks the question I avoid.²⁷ It never picks the question worth asking; the author keeps that choice. Qualped calls this the Age of Authorship: a mandate for every person to author, create and design a future instead of inheriting a default.

05The school leaders must build

Global leaders hold the lever: they can lay duckboards across no man’s land, and the evidence shows where each board goes. Microsoft’s data marks the path: countries that invested early in digital infrastructure, AI skills and government adoption, such as the United Arab Emirates and Singapore, lead the world in use.^26,11

1. Teach the brief. Every citizen should learn to hand a machine a brief: a purpose, the context, the limits, an example and a test for the answer. In Benin City, students in the program outscored their peers on AI knowledge and digital skills after six weeks.¹
2. Teach judgment. Every curriculum should train people to check a claim against a primary source, to probe the jagged frontier and to reject a confident error. Canada’s strategy describes literacy the same way: people who can recognize AI, weigh its output and decide what place it earns in their lives.⁵
3. Put teachers beside the machine. The Nigerian team tied its results to careful use with teacher support, and the Harvard tutor ran on proven teaching methods.^1,4 A leader who hands children a chat box without a teacher invites the Kenyan split into every classroom.
4. Treat access as public infrastructure. Fund connection, public computing and subsidized plans the way earlier generations funded libraries and roads. Canada’s strategy already names public libraries as partners to carry AI literacy into every community, with a focus on rural, remote and northern regions.⁵
5. Replace dread with specifics. Name each risk, teach its safeguard and measure progress in skills rather than headlines.

In June 2026, Canada took a first step. Its national AI strategy promises free entry-level AI training for all Canadians, AI literacy content for one million post-secondary students and training for more than 3,000 educators. Its authors put the logic in one line: “For Canadians to benefit from AI, they must first learn to use it.”⁵ Three thousand educators make a start, but every classroom and every library should get the same training, in Georgetown and Benin City as well as Toronto.

Any reader can start this week without a minister’s permission. Write the problem in one sentence before you open the machine. Ask the machine to question you before it answers. Check one claim in every reply against a primary source. Finish one small piece of work each week: a letter, a budget, a lesson plan or a fix for a real problem on your street.

06Lay the boards

Nothing in that scene calls for a genius; it calls for a taught prompt, a practised judgment and a machine the public funds. Uyi’s rule still holds: the prompt decides the harvest. Leaders owe every child a path across that mud.

Build the school. Hand the next child the keyboard, and let her write the first line.

Shaun Michael Samaroo ◆ Ontario, Canadawith QwaiAI, cognitive executive ◆ 10 September 2026

This work uses SPELJ, Shaun Michael Samaroo’s genre that fuses short story, personal essay and literary journalism. QwaiAI took the role of cognitive executive for this Note.