The Calculator Was Cheating Too

A professor hid a word in white text and caught 32 of his 35 students. Every generation calls the new tool cheating, then quietly changes the test. Britain has just been handed both problems at once.

This issue is also available as a podcast. Listen on Spotify, Apple Podcasts or YouTube and tell me what you think.

A day late this week. I’ve been on holiday, and for once the newsletter went without me.

BOTTOM LINE UPFRONT: Using AI to write your essay is cheating, and that is by some distance the least interesting thing about it. A history professor in Mississippi buried an instruction in white text, and 32 of his 35 students handed it straight back to him. That tells you the students were lazy. It also tells you the assessment was already obsolete. We have been here before, repeatedly. The slide rule, the pocket calculator, the word processor and the search engine were each denounced as the death of a skill, and each time the resolution was the same: change the test, not the tool. AI breaks the pattern in one respect that matters. It does not just do the arithmetic. It does the thinking we were grading. Which is exactly why Andy Burnham’s technical education reforms, announced on Tuesday, are aimed at the right target, and exactly why aiming them at September 2028 is not good enough. Switzerland, Germany and Singapore did not build parity of esteem with a press release. Neither will we.

Thirty-two out of thirty-five is not a cheating scandal. It’s an assessment failure

Dr Jason Gibson teaches history and African American studies at Alcorn State University in Mississippi. He set his summer classes a midterm essay comparing the Industrial Revolution with the digital age. Buried in the prompt, in white text invisible on the page, was an instruction addressed to nobody human: put the word Madagascar somewhere in the answer, in a way that makes no sense.

Any student who read the question saw nothing. Any student who copied the question into a chatbot, and then copied the chatbot’s answer back without reading it, wrote themselves a confession.

Thirty-two of his 35 students confessed. The results included “Madagascar wore a toaster to a basketball game”, and a claim that the island floats sideways through afternoons. Gibson’s summary of it, in the TikTok video that has since gone round the world, was that “apparently they didn’t proofread it.”

To his credit, he did not enjoy it. He explained the trap to the class, invited appeals, and got two. One succeeded. He has said he probably won’t do it again, because catching cheats is not the job.

Here is what should worry us more than the cheating. The trap only worked because the task was one a machine could do end to end. Gibson had set a competent, conventional, entirely defensible essay question, and a free chatbot could complete it to a passing standard in nine seconds. The white text caught the students. It also caught the assignment.

Brown University supplies the control experiment. This spring the economics professor Roberto Serrano set a take-home midterm in welfare economics for the first time in nearly two decades, after a shooting on campus in December left students anxious about sitting exams in a room. Enrolment nearly tripled, from a typical 30 to 86. The take-home average came in at 96%, with 40 perfect scores, against a historical range of 65% to 80% on a paper Serrano had deliberately made harder.

He warned the class and moved the final in person. Eighteen students dropped out. Nine stayed enrolled and didn’t sit it. Of the 27 who never took the final, 22 had scored full marks on the midterm. Among the 59 who did sit it, the average was 48.6%, the lowest in the history of the course. Nineteen failed. Serrano’s verdict on where this leads was blunt: “We cannot choose to become idiots.”

Read the Brown numbers carefully, because they are worse than the headline. The cohort that sat the proctored final had already been filtered. The most likely cheats had removed themselves. What was left still collapsed by nearly half.

There is a neurological footnote, and it deserves its caveats. An MIT Media Lab team put 54 people in EEG headsets and had them write essays with ChatGPT, with a search engine, or with nothing. Brain connectivity scaled down with the amount of external help, and the group using the model showed the weakest coupling of the three. The paper is a preprint, the sample is small, and it has not been peer reviewed. Treat it as a hypothesis with electrodes attached rather than settled science. But it points the same way as the exam data, and the exam data is not a preprint.

I asked a version of this question in The Times in 2019, in a Thunderer column written before most people had heard of a large language model. The line I would stand behind today is this one: teaching young people how to retain information equips them with skills for a bygone age, and teaching them how to apply information empowers them. Jack Ma had put the case more economically. Computers, he said, never forget and never get angry.

Seven years on, the machines have moved from remembering to reasoning, and we are still setting the retention exam.

Every tool we now call essential was once called cheating

The slide rule was going to destroy mental arithmetic. The pocket calculator was going to destroy numeracy altogether. Word processors and spellcheck were going to destroy spelling and handwriting. Google and Wikipedia were going to destroy research. Each panic was sincere, each was partly correct, and in each case the profession eventually did the same thing: it stopped grading the part the machine had taken over.

The calculator is the closest parallel, and the real history is more interesting than the folk version. As early as 1975, America’s National Advisory Committee on Mathematical Education recommended that every maths student from the end of eighth grade should have a calculator available, and should be allowed to use it in all their work, tests included. In 1980 the National Council of Teachers of Mathematics went further in An Agenda for Action, calling for problem solving to become the focus of school mathematics, for “basic skills” to mean more than computational facility, and for calculators and computers to be used at all grade levels.

Note what that reform actually did. It did not lower the bar. It moved it. Long division stopped being the thing you were graded on, and setting up the equation became the thing you were graded on. The tool absorbed the drudgery and the assessment climbed one rung up the ladder.

Search engines did the same to the essay. Once every fact was thirty seconds away, “when was the Battle of Waterloo” stopped being a question and “why did the coalition win it” became one. Retrieval was commoditised, so we started grading synthesis instead.

And that is precisely where the analogy stops being comforting. Every previous tool automated a lower rung. AI has climbed to the top of the ladder and taken the rung we retreated to. Synthesis was our answer to Google. It is not available as an answer to Claude or ChatGPT, because synthesis is the thing they are best at.

So the honest version of the calculator lesson is not “relax, we’ve seen this before.” It is narrower and harder. The lesson is that the tool always wins, the syllabus always follows, and the interval between the two is where a generation gets damaged. The interval for the calculator ran to roughly a decade. The interval for AI started in November 2022 and shows no sign of closing.

Which leaves two live options, and only two. Assess under conditions where the machine is absent, in person, viva voce, on paper, defending your own work in a room. Or assess the use of the machine itself, and grade the prompting, the verification, the judgement and the correction. Both are more expensive than an unproctored essay. That is the entire reason we’re not doing either.

Burnham has read the problem correctly and set the alarm for 2028

On Tuesday the Prime Minister stood in a train factory in Derby and said something no British premier has said with conviction in thirty years. From now on, Britain would “value the hard hat every bit as much as the graduation cap.”

The substance behind the soundbite is real. From Year 10, pupils in England will be able to combine core academic subjects with technical education tied to the jobs actually available where they live, with pathways shaped by mayors, colleges and local employers, and with Ofsted inspecting the vocational offer rather than ignoring it. Alongside it came a bursary worth up to £4,500 a year per household, free apprenticeship training for eligible under-25s from 1 August, up to £8,000 of support for small and medium employers taking on young apprentices, and a target of 50,000 new youth apprenticeships by the end of this Parliament. A further £287 million will fund more than 22,000 extra college places, weighted towards construction and technical subjects.

The bursary in particular is a genuinely thoughtful piece of policy, and it is worth understanding why. Families on Universal Credit have been losing money when their child takes an apprenticeship, because the apprentice wage comes in below the benefits it displaces. The Social Security Advisory Committee has been warning about that cliff edge for years. Burnham’s government has finally paid to remove it. That is the unglamorous, specific, plumbing sort of reform that actually shifts behaviour.

Now the numbers that explain the urgency. In the first quarter of this year 1.01 million people aged 16 to 24 in the UK were not in education, employment or training, 13.5% of that age group, the first time the figure has passed a million since 2013. Of those, 39% were unemployed. The other 61% were economically inactive, meaning they had stopped looking. The total rose by 89,000 over the year, and by 55,000 in a single quarter. The UK youth unemployment rate stands at 16.2%.

And the reforms begin in September 2028.

That is the sentence the government would rather you skimmed. A fourteen-year-old starting Year 10 in 2028 sits their A-levels or equivalent in 2033. The million young people currently outside work and study get nothing from this policy at all. They are not in Year 10. They are in their twenties, in their parents’ spare rooms, and the machine that displaced the entry-level graduate job they were aiming at was shipped three years before the first technical pathway opens.

I have some sympathy for the timetable. You cannot build a technical curriculum, train the teachers and sign up the employers in eighteen months, and pretending otherwise is how we got the last four failed vocational reforms. The revised national curriculum goes back out to consultation in spring 2027 for a reason. But a policy that is correct and five years away is not a response to a crisis. It is a response to the crisis after this one.

The Prime Minister said he could not say precisely when he would bring the NEET figure down. On present evidence, neither can anyone else.

Switzerland, Germany and Singapore did not do this with a press release

Every British government since the war has promised parity of esteem between academic and technical education. The phrase entered official English policy language in the 1943 Norwood Report. We have been failing at it for eighty-three years, which is long enough to suspect the problem is not the wording.

Look at what the countries that succeeded actually built.

Switzerland. About two-thirds of Swiss teenagers go into vocational education and training after compulsory schooling, choosing from roughly 250 federally recognised occupations, while a third take the academic baccalaureate route. Just over 90% of those VET students train through a company apprenticeship rather than a full-time school. The decisive feature is not the split. It’s the permeability. A Swiss student can change direction during training, and change career later, with relative ease, which is why choosing the workshop at fifteen does not close the university door at twenty-five. Burnham’s “no dead ends” is a direct translation of that principle. The Swiss spent decades building the plumbing that makes it true.

Germany. The duales System pairs the vocational school with the employer’s floor, with the chambers of commerce holding the standards. Depending on which body you count, there are somewhere between 250 and 330 recognised training occupations, and around 400,000 companies offering places. German youth unemployment sits at 6.3%, against an EU average of 14.7%. Set that against our 16.2% and the argument is over.

Two honest caveats, because this is where British policy papers usually stop reading. The German headline rate flatters the picture: many school leavers are not counted as unemployed because they are parked in transitional schemes, waiting lists and unskilled work, which delays rather than prevents the problem. And the export record is poor. Most evaluations of attempts to transplant the German system abroad show little long-term effect, and only a handful of central European countries have adopted it. A system built on eight centuries of guild structure and a particular relationship between employers and the state does not arrive in a shipping container.

Singapore. The most useful case for Britain, because Singapore started from where we are. Its Institute of Technical Education was widely regarded as the destination for children who had failed at everything else. The government did not argue with that perception. It bought its way out of it, building lavish campuses, wiring them into multinational employers, and funding the whole thing as industrial strategy rather than welfare. Employment rates for ITE graduates six months after finishing now run close to 90%. SkillsFuture, launched in 2015, extended the same logic across the working population on the assumption that any static qualification decays.

The common thread across all three is money, patience and employer obligation, in that order. Not one of them achieved parity of esteem by declaring it. They achieved it by making the technical route demonstrably lead somewhere better than the alternative, and then waiting a generation for parents to notice.

Britain’s specific disease is the last part. Esteem here has never attached to the skill. It attaches to the social origins of the people who hold it. That is why every rebrand from City & Guilds to GNVQ to Diploma to T-Level has been absorbed and downgraded within a decade. You cannot legislate your way out of a class system. You can only outspend it, which is roughly what Singapore did, and £287 million is not that number.

The Signal Tech & AI Layoff Tracker

A quiet week by the standards of this summer, which is what the last week of July usually delivers. The significant activity was disclosure rather than announcement, as second-quarter earnings forced companies to put numbers against restructurings they had already trailed.

BP, ~700. Announced 30 July, attributed to organisational simplification, debt reduction and a renewed focus on core oil and gas. Not a technology story, and logged against the cross-sector line only. Worth watching all the same, because energy majors are now competing directly with hyperscalers for the same grid capacity.

Magic Leap, ~200. Announced 28 July. The augmented reality pioneer that raised billions is completing its retreat into component supply, pivoting to selling waveguides to other people’s headsets. A reminder that the losing side of a platform bet does not vanish. It becomes a supplier.

K&L Gates, around 10% of non-lawyer staff. Announced 27 July, headcount not disclosed. This one matters more than its size. Professional services firms have spent two years insisting AI would augment rather than replace, and the first cuts are landing exactly where the theory predicted: not on the fee earners, on the people who support them.

Intel, undisclosed. Reports late in the previous week, still developing, that its data centre and AI group is next. That division posted first-quarter revenue of $5.05 billion, up 22% year on year. Intel’s global headcount has fallen from roughly 132,000 in 2022 to around 81,000. Cuts in your fastest-growing unit are not cost control. They are a statement about what the unit needs to look like.

The macro picture is the one to hold onto, because it connects directly to everything above. Financial Times analysis puts US tech job cuts at close to 140,000 so far this year, with Amazon, Oracle, Meta and Microsoft accounting for nearly 50,000 between them, roughly 6% of their combined workforce. The same four are expected to spend as much as $725 billion in capital expenditure this year.

Two findings from that analysis are worth pinning up. Enrico Moretti, the Berkeley economist, argues that blaming AI is largely a way for executives to avoid admitting they overhired, an “easy way out.” And investors appear to agree. In the 30 trading days after an announcement, companies attributing cuts to AI lagged the Nasdaq by nearly 10%, against about 4% for those citing other reasons. Amazon and Microsoft have both since stated plainly that AI was not the driver of their latest reductions.

Watch list for Week 31: Intel’s data centre and AI group headcount, which may surface in earnings detail; and professional services generally, where K&L Gates has just given every managing partner in London a template.

Final Thought 🚀

Hold the week’s two stories against each other.

In Mississippi, a professor proved that a machine can pass the test we set our young people, and that our young people know it. In Derby, a Prime Minister announced that Britain will start teaching something a machine cannot do, beginning in September 2028.

He is right about the direction and wrong about the clock, and the clock is the whole argument. The entry-level jobs are being removed now, one restructuring memo at a time, by companies spending three-quarters of a trillion dollars on the machines doing the removing. A million young people are already outside work and study, and 61% of them have stopped looking. The technical pathway that might have caught them opens when they are pushing thirty.

We have run this experiment before and we know how it ends. The tool arrives, the institutions call it cheating, the syllabus eventually catches up, and the cohort caught in the gap pays for the delay with the first ten years of their working life. The calculator gap cost us a decade of maths teaching. This gap is wider, it is moving faster, and it sits on top of a class system that has spent eighty-three years pretending to respect the toolmaker.

Seven years ago I wrote that the ancient Greeks freed their elite for philosophy through the inhumane use of slaves, and that the next generation could master the machines and free themselves from the same servitude without the cruelty. I still believe that. But emancipation is not automatic, and it does not arrive on the government’s timetable. It arrives when we stop testing children on the things we have already automated, and start testing them on the things we have not.

Madagascar wore a toaster to a basketball game. Thirty-two students submitted that sentence without reading it. Before we decide what that says about them, we should be honest about what it says about the question we asked.

Until next Sunday,

David