Hired by the labs, closed by the universities: what philosophy is being paid for

In the outcomes-by-major table it published in February 2026, the Federal Reserve Bank of New York put the unemployment rate for recent philosophy graduates in the United States at 5.1 per cent, and the rate for recent computer science graduates at 7.0 per cent. "Recent" means aged 22 to 27, holding a bachelor's degree in the subject. The two numbers have been in circulation ever since, usually with an air of vindication, and usually on their own.

The rest of the row is less quotable. In the same table, 47.1 per cent of those philosophy graduates were underemployed — working in jobs that do not normally require a degree at all — against 19.1 per cent of the computer scientists. Median early-career pay was 52,000 dollars for philosophy and 87,000 dollars for computer science; by mid-career, 80,000 against 120,000. A philosophy graduate is slightly more likely to have a job and considerably more likely to be pulling pints in it.

Set that beside the other thing happening this year. On 6 May the University of Hertfordshire told staff it was withdrawing philosophy from its curriculum entirely — not shrinking it, withdrawing it, along with history, English literature, linguistics and creative writing. Offers already made to applicants were pulled. Once the teach-out is finished, philosophy will not be taught at any university in Hertfordshire. So the discipline is being bought and shut at the same time, and the interesting question is not which trend wins. It is what, precisely, is being bought, and what, precisely, is being shut, because they are not the same object.

The Fed's numbers are sound and the headline drawn from them is not#

Begin with the statistics, because a good deal of commentary this year has treated them as a moral fact rather than a survey estimate. The Fed's by-major table is built from the American Community Survey and refreshed each February; the February 2026 release is the current vintage, and the figures above are its figures, not a magazine's paraphrase of them.

The trouble is sample size. Roughly 5,900 bachelor's degrees in philosophy are awarded in the United States each year. In computer science the figure is around 94,000 — sixteen times as many. Subdivide a national household survey by a field that small and the confidence intervals become enormous. The Economic Innovation Group made this point sharply in its critique of the viral version of the chart: computer engineering's reported 7.5 per cent unemployment carried a 95 per cent interval running from about 4 per cent to about 11 per cent, which is the difference between a strong labour market and a recession. A gap of 1.9 points between philosophy and computer science does not survive that kind of noise.

What does survive is the pay and underemployment data, where the samples are larger and the gaps are not marginal but enormous. Nobody quotes those. The honest summary of the Fed table is that philosophy graduates are employed at roughly the same rate as anyone else and paid considerably less, and that computer science, after two years of contraction in entry-level hiring, has stopped being the exception it was. That is a real finding. It is not the finding that "philosophy beats computer science" implies, and this site's earlier note on the New York Times and the revenge of the philosophy major should be read with the caveat attached.

What a laboratory is buying when it buys a philosopher#

The hiring, reported here in August when Philosophy Now counted the firms involved, is real and it is not a gesture. Daily Nous keeps a running list, updated on 6 July, and the names on it are not obscure. Amanda Askell, Joe Carlsmith, Jackson Kernion, Ben Levinstein and Harvey Lederman are at Anthropic — Levinstein from Illinois at Urbana-Champaign in March, Lederman more recently, hired to work on alignment and character. Iason Gabriel, Adam Bales, Atoosa Kasirzadeh, Arianna Manzini and Julia Haas are at Google DeepMind. Robert Long runs Eleos AI. These are people with publication records in philosophy of mind, decision theory and political philosophy, and they left tenure or tenure-track posts to take the jobs.

The work itself is more specific than "ethics". Askell was lead author of Claude's constitution, published on 21 January 2026 and written, unusually, as a document addressed to the model rather than about it: an argument for why it should behave as it does, including an instruction to refuse assistance with the illegitimate concentration of power and, more strikingly, to push back against Anthropic itself where the company's requests look inconsistent with the values it has stated. Askell's stated rationale is worth quoting, because it is a claim about generalisation rather than compliance.

Instead of just saying, 'here's a bunch of behaviors that we want,' we're hoping that if you give models the reasons why you want these behaviors, it's going to generalize more effectively in new contexts.

OpenAI's equivalent is the Model Spec, most recently revised on 18 August 2026. It is a working document of considerable formal sophistication: a five-level chain of command running from root rules through system, developer and user instructions down to overridable guidelines, with hard constraints distinguished from defaults and a stated preference for pushing as much as possible down to the overridable level in order to avoid paternalism. Anyone who has read a canon law commentary, or the constitutions of a religious order, will recognise the genre immediately. It is casuistry: the systematic derivation of case judgements from stated principles, with an explicit hierarchy for resolving conflicts.

That is genuine philosophical labour, and it is closer to what most philosophers actually do than the sceptics allow. But notice two things. The Model Spec, by OpenAI's own account, has had text contributed by dozens of people across research, engineering, product, safety, policy, legal, communications and global affairs; philosophy is one voice in a drafting committee, not the author. And the job title is almost never "philosopher". It is member of technical staff. The discipline is being bought as a method, disaggregated from its institutional form, and the people doing it are answerable to a product schedule.

The arithmetic that closes a department#

The closures have a mechanism, and most commentary gets it wrong by reaching for decline-of-the-humanities sentiment when the answer is in the funding regulations.

English undergraduate tuition was capped at 9,000 pounds in 2012, rose with inflation to 9,250 in 2017, and was then frozen for eight years. It went to 9,535 pounds in 2025-26 and to 9,790 pounds for 2026-27, a rise of 2.71 per cent. The House of Commons Library puts real resources per student at about 12,600 pounds in 2012-13 and about 9,900 pounds in 2025-26. Meanwhile philosophy sits in price group D, the classroom-based band, which attracts no high-cost subject funding from the Office for Students at all. Laboratory and clinical subjects draw a teaching grant. Humanities draw the fee and nothing else. A philosophy department's income is therefore, to a first approximation, headcount multiplied by a number that has fallen by a fifth in real terms over a decade.

International recruitment is the other half, and here the lazy version — collapse — is wrong. Main-applicant study visas in the year to September 2025 were up 7 per cent on the previous year, at 419,558; that is a rebound, not a rout, and the total remains well above pre-pandemic levels. The damage is more specific. A BUILA survey in November 2025 found international enrolment down about 6 per cent overall, with Chinese numbers down 17 per cent and Indian numbers down 9 per cent, concentrated in taught postgraduate programmes, which are where the cross-subsidy lives. Then Home Office figures for the first quarter of 2026 showed sponsored study visas 32 per cent below the same quarter of 2025. A levy of 925 pounds per international student is due from August 2028. Universities plan against forecasts, and the forecasts turned.

The Office for Students published its sector assessment on 14 May 2026: 35.8 per cent of providers in deficit in 2024-25, 42.7 per cent forecasting deficit in 2025-26, restructuring costs up 20.7 per cent to 218.2 million pounds, and a warning from its director of regulation, Philippa Pickford, against persistent over-optimism in institutional projections. Against that background, a subject-level contribution calculation is run: fee income attributable to the subject, minus staff costs, minus an allocated share of overhead. Philosophy, taught in small groups by expensive permanent staff, with modest enrolment and no research grant income to speak of, loses that calculation almost everywhere it is run.

What is striking is that nobody sees the calculation. Hertfordshire's UCU branch reported that affected staff had not seen any of the metrics used in the decision process, nor been asked to contribute to it. The university's stated reason was sustained low recruitment and declining demand; its vice-chancellor announced 200 job losses and a 20 million pound savings target in June. At Dundee, where a 35 million pound deficit drew 62 million pounds of Scottish government money in 2025 and the philosophy programme was partly rescued in August after a campaign that ran through the summer, the pattern held in a more revealing form. Single honours philosophy was suspended, along with joint degrees pairing it with politics, film and history. What was approved was a new undergraduate and postgraduate provision in philosophy and artificial intelligence, endorsed by Luciano Floridi and Shannon Vallor. Philosophy survives at Dundee in the form that carries the word AI. Our earlier survey of the wider pattern of UK closures has the fuller list.

Two different activities share a word#

Floridi has called the movement of philosophers into the technology sector a haemorrhaging of philosophy departments. Brian Leiter's response was blunt — that this is obviously nonsense, since philosophy PhDs have outnumbered academic posts by a wide margin for decades, and a few dozen departures cannot drain a labour market in permanent oversupply. Leiter is right about the arithmetic. Floridi is pointing at something the arithmetic does not capture, which is not the number of philosophers but the direction of the questions.

A laboratory hires a philosopher to make a product defensible: to specify its behaviour precisely, to anticipate objections, to build the evaluation that shows the objection has been met. This is skilled and often admirable work. It is also work that begins downstream of a decision already taken, which is that the system will ship. A department, when it is functioning, trains people to ask a question one step earlier — whether the thing should be built, on whose authority, and at whose cost. The second question has no client.

The strongest case against that distinction is historical, and it deserves stating properly rather than as a straw man. Applied ethics has almost always done its serious work inside institutions rather than outside them. The Warnock committee, which reported in July 1984 and produced the framework that became the Human Fertilisation and Embryology Act 1990, was chaired by a philosopher working within a government inquiry, to terms of reference she did not set, under a deadline. Research ethics committees, clinical ethics consultations and data protection boards have shaped practice far more than any monograph. Mary Warnock herself denied that philosophers had any special standing in such rooms: in matters of life and death, she wrote in 1985, no-one is prepared to defer to judgements made on the basis of a superior ability in philosophy, and there cannot be moral experts. She was the most influential philosopher in British public life in her generation, and she thought the authority claim was empty. That is an argument for embedded work, not against it.

Nor is academic philosophy's own record a clean one. The discipline paid very little attention to computing for the first fifty years it existed, and a good deal of what it published in the decades when it might have been useful went unread by anyone outside it. Departments that now appeal to public value are appealing to a standard many of them declined to be judged by when funding was easy. The defence of the humanities is weaker for having been made so late.

Who pays for the questions nobody is commissioning#

Here is the claim, stated as carefully as the evidence allows. Industry funds the questions it needs answered, and it needs a great many of them answered: how to specify a behaviour, how to rank competing obligations, how to test whether a model has understood an instruction, whether a system of this kind can have interests. Those are real questions and the money spent on them is not wasted.

The questions industry does not need answered are of a different kind. Whether a general-purpose model of this capability should be trained at all. Whether the labour arrangements that produce the training data and the safety annotations are just. Whether the concentration of this technology in five firms is compatible with democratic self-government. Whether the correct response to a persuasive machine is regulation, refusal or something we have not thought of. No firm will fund the fourth answer if it is refusal, and no firm should be blamed for that; it is simply not what a firm is for. It is what a university is for, and in England the university is now funded in a way that makes the subject which asks those questions the cheapest thing on the balance sheet to remove.

If the first set of questions is well funded while the second is defunded, the discipline does not shrink. Head counts may even rise; the Fed's unemployment figure may keep looking respectable. What changes is who decides which questions are worth asking, and the answer moves from a committee of academics, with all their parochialism and their slowness, to a product roadmap. The test is not whether philosophers are employed. They are, at better salaries than the universities offered them. The test is whether anyone is still paid to ask the questions their employer would rather were not asked, and on that measure the trend line and the headline point in opposite directions.