MyTh

In 2021 I was given £36,500 by the Global Challenges Research Fund to work on a problem that sounded, at the time, faintly eccentric: could a machine automatically redraw a children’s picture book for a different culture?1 Not translate it. Redraw it. A child in Kerala reading a story set in a British seaside town does not know what a fish-and-chip shop is; a child in Yorkshire does not know a dosa stall. Publishers already solve this by hand, with an illustrator, one book at a time, which is why it reaches almost none of the world’s languages.

Five years later the same question has a fashionable name — cultural alignment — and everyone has an opinion about it. So here is mine, which is less comfortable than the one you would expect from someone who has spent six years on this.

The tax you pay for your alphabet

Start with the least arguable part, because it is arithmetic rather than politics.

Language models do not read characters. They read tokens, and the tokenizer that produces them is fitted to whatever the model saw most. English got the efficient encoding. Everything else got what was left. Petrov and colleagues measured this properly and found that the same text, translated, can take up to 15 times as many tokens depending on the language it is written in.3 A later study across more than 200 languages found non-Latin scripts and morphologically rich languages routinely paying three to five times the English rate.4

That is not an aesthetic complaint. Tokens are the unit of billing, the unit of latency, and the unit of memory. So a speaker of an unlucky language pays more money for the same answer, waits longer to receive it, and can fit less of their problem into the model’s context before it starts forgetting the beginning.3 The same query, the same model, the same day — and a structurally worse service, priced higher, for the people least able to absorb either.

I find this the most clarifying fact in the whole debate, because nobody chose it. No committee decided that Telugu speakers should pay a premium. It fell out of an engineering decision about compression, made by people optimising for the corpus in front of them. Which is exactly how this kind of bias usually arrives: not as prejudice, but as a default that nobody had a reason to question.

Data, not speakers, decides who the model serves

The obvious assumption is that big languages are well served and small ones are not. It is wrong, and the counterexample is instructive.

Stanford’s work on the language divide points out that Swahili has roughly 200 million speakers but too little digitised text for models to learn from properly, while Welsh — with a fraction of that number — does comparatively well, because it has been documented, digitised and institutionally preserved for decades.5 Models work well for the 1.52 billion people who speak English, less well for 97 million Vietnamese speakers, and worse again for the 1.5 million who speak Nahuatl.5

So the variable is not how many people speak your language. It is how much of your language someone bothered to write down in a machine-readable form, and whether anyone funded that. Which means the map of who AI serves well is, to an uncomfortable degree, a map of historical investment in linguistic infrastructure — colonial, national, academic. The model inherits that map without ever being told it is a map.

Whose values got called “alignment”

Now the harder claim. A model trained overwhelmingly on one part of the world does not represent that part as a culture. It represents it as the absence of culture — the neutral case, from which everything else is a departure requiring a modifier. Ask for a wedding and you get one specific wedding. Ask for a family, a breakfast, a professional, a beautiful house, and the defaults have a postcode.

The literature calls this WEIRD — Western, Educated, Industrialised, Rich, Democratic — borrowing a term Henrich, Heine and Norenzayan coined in 2010, when psychology noticed that its findings about “human nature” rested on samples drawn overwhelmingly from Western university undergraduates — among, as they put it, “the least representative populations one could find for generalizing about humans”.96 The parallel is close enough to be embarrassing. A discipline generalising from an unrepresentative sample, and calling the result universal, is not a new failure mode. We have simply automated it and put it behind an API.

What makes the word alignment worth pausing on is that it names a destination without naming it. Aligned to what? The honest answer, most of the time, is: to the preferences of the annotators who were hired, under the guidelines they were given, at the companies that could afford to run the process. That is a legitimate thing to optimise for. It is not a neutral thing, and the vocabulary invites us to forget the difference.

The finding that ruins the easy version of this argument

Here is where I part company with most people who agree with me so far.

If the problem is Western defaults, the obvious fix is less Western default. Zhou, Constantinides and Quercia tested that directly across GPT-3.5, GPT-4, Llama-3, BLOOM and Qwen. Models with lower alignment to WEIRD values did produce more culturally varied responses — and were also 2–4% more likely to generate output that violated human rights, particularly on gender and equality. The examples are not subtle: endorsements of the ideas that a man who cannot father children is not a real man, or that a husband should always know where his wife is.6

I do not think this vindicates the current defaults, and I want to be careful here, because that is exactly how the finding will be used. Human rights are not a Western cultural preference, and treating them as one is a move with a long and disreputable history. But it does demolish the lazy formulation — that cultural representativeness is straightforwardly good and Western skew is straightforwardly bad. Sometimes a model reproducing local majority opinion faithfully is a model reproducing something a great many people in that locality are fighting against.

Anyone selling you a clean answer to that has not read the tension carefully. I certainly do not have one. What I have is the conviction that the trade-off should be visible and argued about, rather than resolved silently in a guidelines document by whoever happens to be writing it.

And we cannot currently tell whether any of it works

The methodological news is worse than the political news.

Most cultural alignment evaluation works by giving the model survey questions — typically from instruments like the World Values Survey — and comparing its answers to a country’s human responses. Khan, Casper and Hadfield-Menell examined the three assumptions this rests on, and found all three fail.7 Cultural alignment is not stable: response shifts caused by trivial changes in question formatting frequently exceed the actual differences between cultures. It is not extrapolable: alignment measured on a few dimensions predicts alignment on held-out dimensions about as well as chance. And it is not steerable: prompting a model to adopt a culture produces erratic patterns that no human population resembles.

Read that again, because it undercuts a great deal of published work, including work I am sympathetic to. If reformatting your question moves the result more than swapping Japan for Germany does, then you are not measuring a property of the model. You are measuring your own questionnaire. A field cannot claim to have reduced cultural bias using an instrument that unreliable, and the honest position is that many confident claims in this area — in both directions — are currently unsupported.

Self-awareness, not substitution

Six years on the same problem has moved me from the position most people start at to one that sounds duller and is, I think, correct.

The goal is not to find the right culture to encode. There isn’t one, the attempt to pick one is how we got here, and the evidence suggests that swapping the default out wholesale trades one set of harms for another. The goal is for the system to know that it is standing somewhere — that its defaults are defaults, that the person in front of it may not share them, and that this is a thing to be established rather than assumed. That is the argument behind CALM: Culturally Self-Aware Language Models, which we published at NeurIPS last year,2 and it is the same instinct as the picture-book problem: not “what should this child’s story be”, but “does the system know whose story it is telling”.

Three things follow, and none of them are technical.

• The tokenizer tax is the easiest win and nobody is taking it. It is a measurable, unambiguous, purely engineering inequity that makes AI more expensive for the world’s poorer speakers. It requires no philosophy to fix. It persists because the people paying it are not the people in the room.
• Digitisation is cultural policy, and it is being done by accident. Whether your grandchildren’s language works with these systems is being decided now, by who funds corpus-building. Swahili and Welsh should be the standing example in every ministerial briefing on AI. It almost never is.
• Fix the measurements before trusting any claim, including mine. On the current evidence, “our model is culturally aligned” is not a verifiable statement. Until the evaluations are robust, that phrase belongs in scare quotes — on marketing copy, on benchmark leaderboards, and in my own papers.

When UNESCO's statistics institute last put a headline number on this, in 2017, it counted 617 million children and adolescents not reaching minimum proficiency in reading and mathematics.8 It is an old figure and I quote it with that caveat, but it is the one that was in front of me at the start, and it is why I began on picture books rather than benchmarks, and it is the thing I try to keep in view when this subject becomes an argument about definitions. Somewhere in that number is a child who would read more if the story looked like somewhere they had been. Whether the technology of this decade reaches them is not going to be settled by a leaderboard.

I have worked in universities on both sides of the world, and I have never known a year in which so many good academics were quietly asking whether they should still be doing this. Not because the work stopped being interesting — it did not — but because four separate pressures arrived at once and started interacting. Individually each is survivable. Together they change what the job is.

What follows is an attempt to be accurate rather than cathartic. Some of it is grimmer than the sector admits. One part of it is genuinely better than the discourse allows, and I will say so plainly, because a critique that cannot register good news is just a mood.

The wage that quietly halved its ambition

Start with the least glamorous number in the sector. The 2025–26 national pay round opened with a “full and final” offer of 1.4%.6 The 2026–27 round improved on it: 2% from 1 August 2026, against a union claim of 3% plus RPI — roughly 7%, or £3,000, whichever was greater. Universities in difficulty may defer even that award by up to eleven months, without back pay.5

Set 2% against the prices actually being paid. CPI ran at 3.3% in the twelve months to March 2026, easing to 2.6% by June.9 Average private rent across the UK rose 3.3% over the year to June 2026, to £1,388 a month — and 6.3% in the North East, where several of the universities making the deepest cuts happen to sit.10 A 2% award in that environment is not a pay rise that feels small. It is a pay cut that has been given a nicer name.

Do that for fifteen years and you get the number the sector has learned to skim past. When UCU launched its pay modeller in October 2021, it put the cumulative real-terms fall at 20% since 2009.7 The joint unions' 2026/27 claim now puts it at 30% since 2011.23 Employers dispute the deflator — the figure is smaller on CPI than on RPI, and the baseline year does a good deal of work in both versions — but nobody seriously argues the direction. Something close to a third of the value of the job has gone, one “affordable in the circumstances” settlement at a time.

The consequence is not that academics are poor in any absolute sense. It is that the trade has changed. The old bargain was: accept below-market pay, receive in exchange security, autonomy, and time to think. Two of those three are now being withdrawn while the pay term stays where it is. UCU counts more than 15,000 job cuts across the sector in the current round of restructures.8 HESA recorded academic staff numbers in UK higher education falling for the first time in more than a decade in 2024–25 — 244,755 academic staff excluding atypical contracts, down 1% on 246,930, breaking an upward trend running since 2014/15.1213 And 29% of academic staff — 69,875 people — remain on fixed-term contracts, with a further 57,365 on “atypical” ones.11

The generational effect is the part that worries me most. Wonkhe's read of the same data shows the share of academics under 35 sliding from nearly a third a decade ago to roughly a quarter today.13 You do not notice a missing cohort in the year it fails to arrive. You notice it fifteen years later, when there is nobody to hand the field to.

“Falling student numbers” is the wrong description of a real problem

Here is where I part company with the standard account. Domestic demand for UK higher education is not falling. In the 2026 admissions cycle, 338,940 UK 18-year-olds applied by the January deadline, up 4.8% on the previous year, with total applicants reaching 619,360.16 Acceptances of UK undergraduates through UCAS grew 3.1% in 2025 against the same point in 2024, as reported by the Office for Students.3 On the raw headcount, more young people want a degree than ever.

Two things are nevertheless true at the same time, and the sector's finances live in the gap between them.

• The growth is below forecast, and forecasts had already been spent. That 3.1% rise came against sector plans assuming 4.1%.3 A percentage point of shortfall on a fee-funded budget is not a rounding error; it is a hiring freeze.
• The students who subsidise the system are leaving one door, not every door. International undergraduate applications through UCAS actually rose 5.1% in the 2026 cycle, to 124,830, with Chinese applicants up 10%.16 But undergraduates are not where international fee income comes from. The one-year taught master's is, and that is the floor that gave way: HESA recorded a 6% fall in international enrolments, and in a BUILA survey 42 of 69 responding universities — 61% — reported postgraduate commencements down for September 2025.18 Study-visa issuance sits well below its 2023 peak,19 and the Home Office received roughly a third fewer sponsored-study applications from main applicants in early 2026 than in the same months of 2025.20

Domestic fees do not cover the cost of teaching a domestic student, so a British undergraduate is, in accounting terms, a loss. The cap was frozen at £9,000–£9,250 for most of a decade while prices rose; it reached £9,535 for 2025–26 and £9,790 for 2026–27, now indexed to inflation for providers meeting quality conditions.2122 Indexing from a base that already lost roughly a fifth of its value does not restore the fifth. It just stops the bleeding at the current wound size.

So the true sentence is not “fewer students”. It is: record domestic demand for a product sold below cost, minus the postgraduate international students whose fees closed the gap. The result is what the OfS now reports with striking calm — 124 English providers, 45% of those analysed, projected to run a deficit in 2025–26, and nearly one in six holding less than 30 days of liquidity.3 The 2026 report has that easing only slightly, to 41% in deficit the following year.4 This is a policy choice about migration and fee regulation that has been reclassified as a management problem for individual vice-chancellors.

And the demographic reprieve is temporary. The current bulge of UK 18-year-olds peaks around 2030. What follows is a smaller cohort arriving into a sector that will by then have cut the capacity it is now cutting.

The machine that writes, and now marks, and now reviews

I work on language models. I am not going to pretend to be a neutral observer, and I am not going to perform alarm I do not feel. But the speed of what happened to academic practice in three years deserves to be stated without euphemism.

Among UK undergraduates, 95% now use generative AI in at least one way, and 94% use it for assessed work. The share pasting AI-generated text directly into submitted work has gone from 3% in 2024 to 8% in 2025 to 12% in 2026.2 Two-thirds of students say assessment on their course has already changed significantly because of it.

The instinct is to call this a cheating problem. I think that is a category error, and a self-flattering one. If a task can be completed to a passing standard by a general-purpose model in nine seconds, the interesting question is not who used the model. It is what we thought we were measuring. A great deal of what universities assess turns out to have been a proxy for effort rather than a measure of understanding, and the proxy has just been commoditised. That is uncomfortable, but it is diagnostic information we were never going to get any other way.

What genuinely troubles me is the supply side, where the same tool is being used by people who are supposed to be doing the judging. At ICLR 2026 — one of the largest machine-learning conferences in the world, in my own field — detection analysis flagged 15,899 peer reviews, 21% of the total, as fully AI-generated, and found some degree of AI involvement in over half of them — with researchers reporting hallucinated citations and confidently wrong objections in reviews of their own work.1415 Peer review is not a bureaucratic formality. It is the only mechanism by which the literature distinguishes itself from a very large pile of assertions.

There is a grim symmetry to it. We are exhorting students to use these tools responsibly while a fifth of the reviews in a flagship venue are generated by them. The academic who outsources a review has usually not become lazy; they have been asked to referee eight papers in a semester in which their department shed a fifth of its staff. Squeeze one produces squeeze three. That is the part the “AI in education” conversation keeps missing: the misuse is a workload symptom before it is an ethics failure.

And students notice the asymmetry. Sixty-eight per cent say AI skills are essential for the world they are entering; fewer than half — 48% — think their teaching staff are helping them build those skills.2 That gap is a straightforward indictment, and it is ours, not theirs.

What a degree is now worth — and the surprise in the data

Here is the good news I promised, and I want to give it its full weight rather than a grudging clause.

In the 2026 Student Academic Experience Survey — 10,065 full-time undergraduates — 45% rated their course good or very good value for money, up from 37% a year earlier and the highest figure in over a decade. Sixty-six per cent are happy with their choice of course and institution, up from 56%. The proportion who have considered withdrawing fell to 22%, the lowest in recent years.1 After a decade in which “value for money” was the stick used to beat the sector, students themselves have started answering the question more warmly. Anyone arguing that undergraduates have lost faith in universities is arguing against the evidence.

But look at what those same students are doing with their week. Sixty-five per cent are in paid employment during term, averaging nearly 14 hours. Their total weekly commitment — study plus work — is 44.2 hours, against a national average working week of 36.6.1 Full-time study is no longer full-time study. It is a full-time degree with a part-time job welded onto it, because maintenance support does not meet the rent.

Meanwhile the destination is getting harder to reach. The Department for Science, Innovation and Technology's own snapshot puts the UK hiring rate down 14% year-on-year to April 2026, with 30 of 38 tracked entry-level occupations in decline — accountant −29%, graphic designer −28%, software engineer −27%. The steepest falls are concentrated precisely in information-processing roles.17 To their credit, the authors state explicitly that this “should not be considered causal evidence of AI's impact”; the labour market has several things wrong with it at once, and disentangling them honestly is not yet possible.17 I would rather cite that caveat than write the headline everyone wants.

Put the two findings together and you get something more interesting than either alone. Students value the experience more than they did, while the economic case for it looks worse than it did. That combination should demolish the framing that has governed English higher education policy for fifteen years — that a degree is an individual investment purchased for a private financial return, and that a course is failing if its graduate salary premium is thin. Students are telling us, with their satisfaction scores and their 14 hours of shift work, that they are buying something else as well. The policy machinery has no column for it.

Four squeezes, one system

What makes 2026 different is not the severity of any one pressure but the fact that they now feed each other. Deficits produce redundancies. Redundancies produce workload. Workload produces the AI shortcut — in marking, in reviewing, in the parts of the job nobody defends in public. Degraded teaching and degraded review erode the thing a university actually sells, which is credible judgement. And an institution whose judgement is not credible has no answer at all for the eighteen-year-old asking why they should spend £9,790 a year plus three years of foregone earnings on it.

I do not have a five-point plan, and I distrust essays that end with one. But I hold three things fairly firmly.

• The funding model is the root cause, and it is a political choice. A regulated fee below the cost of provision, cross-subsidised by an international market that immigration policy then suppressed, is not a market failure. It is a design. Everything downstream — the redundancies, the 2% offers, the mergers — follows from refusing to say out loud who pays for teaching.
• AI is not the crisis; it is the contrast agent. It shows up wherever the system was already weak. Assessment that measured effort rather than understanding. Peer review sustained by unpaid goodwill that austerity had already exhausted. The tool did not create either gap. It made both impossible to keep ignoring.
• Students are more loyal to universities than universities' funders are. The value-for-money numbers went up in the worst financial year the sector has had. That is not a mandate for complacency; it is a mandate we are in danger of squandering.

I remain in this job, for what it is worth, and I would still recommend it — with a candour I would not have needed a decade ago. The work is as good as it ever was. It is the conditions around the work that have been quietly renegotiated, in fifteen annual instalments, by people who never had to describe the cumulative effect in a single sentence.

So here is the single sentence: we have spent fifteen years asking universities to do more, for less, for more people, and we are now surprised that the arithmetic caught up.