In May, the government offered every resident aged fourteen and over a free premium chatbot. A two-hour online course unlocks a year of ChatGPT Plus or Microsoft Copilot, worth about €240. Ministers called agreeing OpenAI a world first, and rightly so. Three months later, The Economist reported the first extensive study of what such tools do to schoolchildren. Its findings deserve a careful reading in Castille and at the Education Ministry.
David Strömberg of Stockholm University wrote the paper with Victor Lei and Yanhui Wu of the University of Hong Kong. They followed 26,811 secondary pupils in one county of central China for thirty months. Approximately 80 per cent of pupils used chatbots for homework, while the remainder made up the comparison group. Within six months, the users' homework scores rose 18 per cent. Their time on each assignment fell from 64 minutes to 45. Then came the closed-book examinations, where the same pupils scored 20 per cent below classmates who worked unaided.
The damage did not fade as the months passed. After two years, AI users scored 24 per cent lower in the entrance examination for high school. At the university entrance examination, they trailed their classmates by 18 per cent. Homework marks once predicted examination success in that county. Among AI users, the highest homework scores now signalled the weakest examination results. Teachers saw rising marks and had no reason to suspect a problem.
Honest readers will note the limits of the study. Pupils chose for themselves whether to use AI, and the data covers one Chinese county. Researchers elsewhere must replicate the work before anyone treats its numbers as law. Yet the pattern inside the data reaches beyond the headlines. The examination penalty fell on pupils who rushed their homework. Those who used AI but spent as long on assignments as non-users lost little. Effort, and not abstinence from the technology, protected their learning.
A second study points in the same direction. Zara Contractor and Germán Reyes from Middlebury College conducted a randomised study with 211 American college students. Students who had access to AI performed better on tests, and the improvement remained a week later. The benefit went to students who asked the chatbot to explain difficult concepts. Those who let it write for them kept nothing once the tool disappeared. The chatbot succeeds as a tutor and fails as a ghostwriter.
Malta has more at stake in that distinction than any other European country. Euro barometer found this year that 39 per cent of Maltese use generative AI daily for work or study. The average for the European Union is 20 per cent, with no member state exceeding this figure. PISA 2025 reported that 91 per cent of Maltese fifteen-year-olds use AI chatbots. The same survey placed those pupils below the OECD average in mathematics, reading and science. Enthusiasm for the tool has outrun mastery of the basics.
Now consider how Malta examines its teenagers at sixteen. Since 2025, school-based assessment has supplied 30 per cent of every SEC result. Schools set and mark that work across Years 9, 10 and 11. MATSEC, the national examinations board, does not moderate those marks. The remaining 70 per cent comes from a closed-book examination, which MATSEC sets and marks. Malta has built the Chinese experiment into its own school-leaving certificate. Chatbots lift the first component, and the study shows they depress the second when pupils rush.
The timing of the state's gift compounds the risk. A pupil turns fourteen in Year 9 or Year 10, as school-based assessment counts. At that moment, a premium chatbot is handed over by the taxpayer. One of the three core modules covers AI for learning, a fraction of a two-hour course. The government announced no uptake targets and no measure of success for the scheme. No one has disclosed the number of pupils under eighteen who have subscriptions. A programme of this size deserves better bookkeeping.
The Education Ministry noticed the danger but then ignored it. In April, Clifton Grima, then minister, launched a National Homework Policy. In the foreword, Director General Jude Zammit cautions that generative AI may lead to over-dependence and diminished authentic thinking. I read the numbered actions in the policy and found none that mentioned artificial intelligence. The warning sits in the preface, and the remedy features nowhere in the measures. Rigid time limits also went from the policy, which declined to equate more time with more learning. That concept holds value for young children overloaded with worksheets. The Chinese data indicate that, for teenagers using a chatbot, time spent on the task is the most important factor.
Some have requested the ministry provide more than a foreword. The Malta Union of Teachers told the political parties before the election that AI belongs in the compulsory curriculum. Its proposals also asked schools to limit the dependency of pupils on digital devices in order to encourage more independent thinking. Ireland issued national guidance on AI in schools in October 2025. Its examination commission wrote rules for AI in coursework. I searched for a Maltese equivalent for compulsory schooling and found none. The Economy Ministry distributed the tool, and the Education Ministry has yet to write the manual.
Delay carries a cost, which shows in figures we already hold. One in three candidates failed the SEC mathematics this year. Inflated school marks will hide such weakness until the examination hall exposes it. Parents who see high marks in February deserve to know what the summer holds. A certificate that rewards borrowed work misleads employers and sixth forms as well.
Byron Camilleri, minister since June, needs no ban and no new strategy. Four decisions suffice, and none requires a foreign consultant. First, schools must hold school-based assessments in class, under the eyes of a teacher. Second, the ministry must issue clear rules on AI in homework during this scholastic year. Third, the Malta Digital Innovation Authority is required to disclose the number of minors who claimed a subscription. Fourth, teachers need training to show pupils the method of the tutor. That method has four steps: attempt the problem, ask, check the answer and explain it back.
The evidence holds both a promise and a warning. A child improves learning speed by struggling with a problem first and seeking help from the chatbot later. Another who copies the answer finishes early and learns nothing. Malta gave its children the machine before it taught them the difference. The examination hall, where no chatbot enters, will deliver the verdict.