← The canon · AItopiaOrAImageddon?

The capability approach — poverty as capability deprivation

interpretation · Amartya Sen · 1979

A reading of events that a reading may need to name.

proposals.md filed this under idea, at line 956, inside the ## idea block that runs from line 195 to line 721. I am filing it as an interpretation, and the disagreement is close enough to be worth arguing rather than asserted.

The idea case is real and should be stated at full strength before it is declined. The canon's published definition of the kind is "a framing later work is built out of rather than argued about," cited "when the thing being watched descends from it and the descent explains its shape." The capability approach is the framing the Human Development Index (1990), the Multidimensional Poverty Index (2010) and an entire branch of official statistics were built out of, and the descent does explain their shape — you cannot read the MPI's indicator list without it. On that clause alone the proposal's filing looks right.

It fails the canon's own three tests anyway.

So the id names an argument, and the argument's implementations are other people's artifacts. That distinction has a practical consequence worth acting on: the thing that does pass the loop test is the Alkire–Foster dual-cutoff counting method (Journal of Public Economics 95:7–8, 2011), which is a computable procedure with an identification step, an aggregation step and published code, and which is what the MPI actually runs. It has no id in this canon and it should. Filing Sen as an idea would quietly annex Alkire and Foster's formalism to Sen's name — which, as it happens, is one of the misuses this entry has to warn against further down. A future run should propose alkire-foster-2011 and file it under idea, where it belongs.

descends_from is empty, and that is a gap rather than a fact. The capability approach's own ancestors are Aristotle's Nicomachean Ethics, Adam Smith's Wealth of Nations (1776) — Sen cites Smith's linen-shirt passage more often than any other single source — Marx, and John Rawls's A Theory of Justice (1971), which the 1979 lecture argues against at length and describes itself as extending. None of the four is in canon/. Rawls in particular is a real absence: half of this entry is a rebuttal to a book the canon does not hold, and a reading that wanted to cite the argument's other side would have nothing to point at.

What it is

On 22 May 1979, at Stanford University, Amartya Sen — then Drummond Professor of Political Economy at Oxford — gave a Tanner Lecture on Human Values titled Equality of What? It runs twenty-four printed pages and is the founding document of what became the capability approach.

The lecture's question is narrow and its answer is not. Everyone who argues about equality, Sen observes, is arguing about equality of something, and the choice of that something does most of the work before any distributive principle is applied. He takes the three answers then on offer — "(i) utilitarian equality, (ii) total utility equality, and (iii) Rawlsian equality" — and argues that each fails, that they fail differently, and that no combination of the three repairs the failure.

The case that does the damage

The engine of the argument is a single example, introduced on page 203 and returned to twice. Sen calls the two people a cripple and a pleasure-wizard — the language of 1979, kept here because it is quoted.

> For example, if person A as a cripple gets half the utility that the > pleasure-wizard person B does from any given level of income, then in the pure > distribution problem between A and B the utilitarian would end up giving the > pleasure-wizard B more income than the cripple A. The cripple would then be > doubly worse off: both since he gets less utility from the same level of > income, and since he will also get less income.

Utilitarianism, maximising the sum, sends resources to whoever converts them into utility most efficiently — and so takes them away from the person who converts them worst, which is very often the person who needs them most. Total utility equality does better on this case but fails elsewhere. And Rawls's difference principle, which judges advantage by an index of primary goods, "will give him neither more nor less on grounds of his being a cripple. His utility disadvantage will be irrelevant to the Difference Principle" (p. 215).

Sen's charge against the Rawlsian version is the sharpest sentence in the lecture, and it is the one to carry:

> Indeed, it can be argued that there is, in fact, an element of "fetishism" in > the Rawlsian framework. Rawls takes primary goods as the embodiment of > advantage, rather than taking advantage to be a relationship between persons > and goods. (p. 216)

Two pages earlier he has already named the cost of that error: "Judging advantage purely in terms of primary goods leads to a partially blind morality."

What he proposes instead, and how little of it he claims

Section 4 of the lecture, on page 217, is headed Basic Capability Equality. The proposal itself:

> It is arguable that what is missing in all this framework is some notion of > "basic capabilities": a person being able to do certain basic things. The > ability to move about is the relevant one here, but one can consider others, > e.g., the ability to meet one's nutritional requirements, the wherewithal to be > clothed and sheltered, the power to participate in the social life of the > community. (p. 218)

And the reason the shift matters, which is the sentence a reading will most often want:

> …there is evidence that the conversion of goods to capabilities varies from > person to person substantially, and the equality of the former may still be far > from the equality of the latter. (p. 219)

Then he takes most of it back, and this is the part that almost every summary of the lecture leaves out. The indexing problem is named and deferred (quoted in full at the top of this entry). Any application "must be rather culture-dependent, especially in the weighting of different capabilities" (p. 219). Basic capability equality is "a partial guide to the part of moral goodness that is associated with the idea of equality" (p. 220). And the lecture's last sentence detaches the proposal from the argument entirely:

> This is my main thesis. I have also made the constructive claim that this gap > can be narrowed by the idea of basic capability equality… I should end by > pointing out that the validity of the main thesis is not conditional on the > acceptance of this constructive claim. (p. 220)

The main thesis is negative: utility, total utility and primary goods are each insufficient, and so is any combination. The capability proposal is offered as a detachable extra. Sixty-seven words at the end of a lecture, explicitly disclaimed, became a field.

The vocabulary as it settled

Over the following two decades the apparatus acquired its standard terms, chiefly through Commodities and Capabilities (1985), the Dewey Lectures of the same year, Inequality Reexamined (1992) and Development as Freedom (1999), whose chapter 4 is titled "Poverty as Capability Deprivation."

The relativity that is not relativism

A separate paper closes the loop that most later readers assume is open. In "Poor, Relatively Speaking" (Oxford Economic Papers 35:2, 1983, 153–169), Sen argues that poverty has, in his phrase, an "irreducible absolutist core" — if there is starvation, there is poverty, whatever the relative picture — while the commodities required to reach a given capability are relative to the society. Adam Smith's day-labourer needs a linen shirt not because linen is a need but because in eighteenth-century England you could not appear in public without shame without one. The capability (appearing in public without shame) is absolute; the commodity bundle that delivers it rises with the society.

Absolute in the space of capabilities, relative in the space of commodities. That formulation is forty-three years old and it is the answer to a question METHODOLOGY.md currently records as unresolved. More on that below, under the fourth citation occasion.

The empirical sibling

Poverty and Famines: An Essay on Entitlement and Deprivation (Clarendon Press, 1981) is not the capability approach, but it is where the same instinct was first made to do empirical work. Sen argues that major twentieth-century famines — Bengal 1943 above all, with something on the order of three million dead — occurred without an aggregate collapse in food availability, and were instead collapses in particular groups' entitlements: the bundles of food a person could lawfully command through production, trade, labour or transfer. Food was present. Rural labourers, fishermen and artisans could no longer command it, as wartime inflation ran ahead of their wages.

This claim is contested and the entry says so. Peter Bowbrick (1986, and again in the Journal of International Development in 2024) has argued the entitlement account of Bengal is refuted on its own evidence; Mark Tauger's work on the 1942 Midnapore cyclone and the Helminthosporium oryzae outbreak reconstructs a real crop failure, and he and Sen exchanged directly in the New York Review of Books in March and May 2011; Cormac Ó Gráda's position is that pure entitlement famines with no food-availability dimension are scarce enough that modern scholarship probably underrates supply. The conceptual move — availability is not access — survives regardless of how the Bengal tonnages resolve, because it does not depend on them. A reading citing this entry for the concept is on firm ground; a reading citing it for "there was plenty of food in Bengal" is walking into a live dispute and should say so.

The Royal Swedish Academy of Sciences announced on 14 October 1998 that Sen would receive that year's Sveriges Riksbank Prize in Economic Sciences in Memory of Alfred Nobel, for his contributions to welfare economics; the prize itself was conferred that December. He is 92, at Harvard, and still gave interviews to the Indian press this year.

How it was turned into numbers, by other people

The Human Development Index, 1990. Mahbub ul Haq built the first Human Development Report around a single composite of longevity, education and income. Sen was, by every account including his own, reluctant — the objection being precisely that a single index buries the weighting problem he had flagged in 1979. Haq's answer is the most quoted sentence in the literature: "We need a measure of the same level of vulgarity as GNP — just one number — but a measure that is not as blind to social aspects of human lives as GNP is." Sen was persuaded on those terms, which is to say he was persuaded that a crude index that pointed at the right things would beat a precise one that pointed at the wrong things.

The Multidimensional Poverty Index, 2010. Sabina Alkire and James Foster's dual-cutoff counting method — one cutoff within each indicator to decide whether a person is deprived in it, a second across indicators to decide whether enough deprivations make a person poor — gave the approach the computation Sen had declined to supply. UNDP and the Oxford Poverty and Human Development Initiative have published a global MPI annually since.

The current instrument, from the 2025 global MPI (17 October 2025), counts ten indicators in three equally weighted dimensions:

A person is poor if deprived in at least one third of the weighted indicators. On that rule, 1.1 billion people (18.3%) out of 6.3 billion across 109 countries are in acute multidimensional poverty, 887 million of them exposed to at least one climate hazard. No 2026 edition had appeared as of 20 August 2026.

Why a reading would cite it

Occasion one: a price is offered as a claim about how people live. This is the standing case and the reason the entry is worth its file.

On 10 June 2025 Sam Altman published The Gentle Singularity, which contains two sentences this canon will be citing for years: "As datacenter production gets automated, the cost of intelligence should eventually converge to near the cost of electricity" and "Intelligence too cheap to meter is well within grasp." Take the argument at full strength, on its own premise, and Sen's objection lands without needing a hostile reading. Electricity is not universal. The Tracking SDG 7 report of 25 June 2025 puts 666 million people without access to it in 2023, 85% of them in sub-Saharan Africa. The ITU's Facts and Figures 2025 (17 November 2025) puts 2.2 billion people offline, 96% of them in low- and middle-income countries, and finds mobile broadband still unaffordable in around 60% of low- and middle-income countries. A price converging to the price of a thing that 666 million people cannot buy at any price is a statement about supply cost, not about who can do anything with it. That gap is the entire content of the capability approach, and Sen wrote the sentence for it in 1979: the conversion of goods to capabilities varies from person to person substantially.

The ITU makes the same argument without the vocabulary. Its Secretary-General describes divides increasingly "defined by speed, reliability, affordability, and skills", and the report's sharpest number — "a typical user in a high-income country now generates nearly eight times more mobile data than one in a low-income country" — is a conversion-factor finding in everything but name. When a reading needs to explain why a connectivity percentage and a capability are different quantities, this is the citation.

Occasion two: the same offer, at the same price, delivering different lives. The project's own most recent reading contains a clean instance, dated 19 August 2026 and recorded in digests/2026-08-20-00.md: Google gave US college students a free year of AI Pro and students in more than 140 other markets a free year of the smaller AI Plus, and the reading noted that "the US tier includes Gemini in Gmail and Docs and 5TB of storage, the international tier is a smaller plan with 400GB, and the gap is worth noting in a lens about how people meet these systems, because where you live decides which one you meet." Two offers, both priced at zero, producing two different capability sets sorted by geography. That is LENSES.md's Concentration lens — which names "price and access" in the same breath and is right to distinguish them — and it is the smallest, cleanest worked example of Sen's point the record currently holds. The same window's Alexa+ going free on Fire TV "on one surface only" is the same shape at a different scale.

The general rule this licenses: a free tier is a price fact. What it is a capability fact about depends on the electricity, the connection, the device, the language, the literacy and the account you need to use it — and those are exactly Sen's personal, social and environmental conversion factors.

Occasion three: a poverty number moves and nobody's life changed. On 5 June 2025 the World Bank raised the international poverty line from $2.15 (2017 PPPs) to $3.00 (2021 PPPs), a roughly 40% increase driven by an inflation adjustment, revised Indian survey methodology, and a rise in the median national poverty line among low-income countries to $3.04. The 2022 count of people in extreme poverty went from 713 million to 838 million — about 125 million people added to global poverty by a definitional change, with the Bank stating plainly in the same document that "changes in the poverty lines do not change the day-to-day realities of people living in poverty."

Cite this when a headline number moves and the reading has to say whether the world moved. It is also the strongest single argument for the approach that Sen never had to make himself: the monetary line is not a stable ruler, it is a convention that gets revised, and each revision is somebody's judgment about what a life costs.

Occasion four, and it is inward-facing, so the boundary is stated first. METHODOLOGY.md lists among the three things actually decided: "Poverty measured as access, not dollars a day. Sen's capability approach; the Multidimensional Poverty Index is the existing instrument." The owner's own words, from 2026-08-18, are recorded there: "I don't think 0% poverty is 2.15 a day. It's measured more in people have access to resources, technology, food. They tried to evaluate this in dollars per day." That is Sen's argument, arrived at independently, which is a reason to think it is the natural one rather than a clever one.

The boundary: this is a note for a session working on the project's design, not a licence for a reading. Nothing in this file is evidence. It places no needle and moves none. A reading may cite Sen to explain why a price claim is not an access claim; a reading may not cite this project's own methodology as a fact about the world, and rule 7 exists because that loop is invisible from the inside.

With that said, the design note is worth writing down, because it closes a question METHODOLOGY.md marks open. §2 records the ratchet as unresolved: access thresholds rise — electricity was not deprivation in 1850, internet access increasingly is — so either the +10 threshold is pinned to a fixed absolute standard that will look quaint, or the pole is asymptotic and never reached. The file states both options are defensible and does not choose.

Sen chose, in 1983, and the choice dissolves the dilemma rather than picking a horn. The threshold is fixed in the space of capabilities — appear in public without shame, be adequately nourished, take part in the life of the community — and floats in the space of commodities, because the commodity bundle that delivers a fixed capability rises with the society. There is no ratchet in the standard; there is a ratchet in the shopping list. A project that measures the capability and lets the basket move gets a series that is comparable across a century without pretending 1850 and 2026 need the same goods. METHODOLOGY.md's own worry — that a fast ratchet makes the poverty rate rise while everyone is better off, rendering as regression on the chart — is a symptom of pinning the threshold to goods. It does not arise when the threshold is pinned to what the goods are for.

The same section's other decided item, hours of median work for the basket, is the natural partner and the file already says so: access is a threshold, hours is a conversion. Sen supplies the argument for why you need both, in the term "conversion factor" itself.

What it got right, and what it got wrong

Not required for interpretation, but the canon carries it wherever there are dated claims to grade, and there are.

Right

The prediction implicit in the whole argument — that a monetary line and a direct deprivation count would keep disagreeing — is borne out, and the size of the disagreement is the finding. As of 2026 the two leading global instruments say: 838 million monetarily poor (2022, at the $3.00 line, World Bank) and 1.1 billion multidimensionally poor (2025 global MPI, 109 countries). These are not the same populations, not the same definitions, not even the same country coverage, and they cannot be subtracted. That is the point. Two serious instruments, both trying to count the poor, differ by roughly a third and overlap imperfectly — which is what you expect if income is a proxy for capability and, as Sen said, a poor one.

The conversion-factor claim became routine. It is now unremarkable for official statistics to report that identical incomes buy different lives — the ITU's affordability and skills framing, the MPI's separation of electricity from sanitation from schooling, the SDG framework's indicator 1.2.2 for national multidimensional poverty. In 1979 this was a philosopher's objection; in 2026 it is how the UN counts.

Institutional adoption is the strongest evidence and it is thirty-six years long. HDI since 1990, published annually. Global MPI since 2010, published annually, 109 countries. Whatever one thinks of the indices, the argument won the question of what development reporting is about.

Wrong, or at least unfinished

The indexing problem Sen named in 1979 was not solved. It was decided by fiat, and the fiat has a price tag. Martin Ravallion's "Troubling Tradeoffs in the Human Development Index" (Journal of Development Economics 99:2, 2012) worked out what the HDI's aggregation rule implies about the relative worth of its components, and found that after the 2010 switch to a geometric mean the index's implicit valuation of an extra year of life expectancy ran from about $0.51 per year in Zimbabwe to nearly $9,000 in the richest countries — the Zimbabwe figure being under 0.3% of that country's own mean income in 2008. He also found the implied returns to extra schooling many times higher than the measured economic returns. Nobody chose those tradeoffs. They fell out of an aggregation formula.

This is the exact failure Sen predicted and declined to prevent. He wrote in 1979 that indexing was serious, that weighting would be culture-dependent, and that it was not the occasion to go into it. It was never made the occasion. The weights were set by whoever was building the index, and a rule that prices a Zimbabwean life-year at fifty-one cents is what "not the occasion" eventually cost. A reading that cites the capability approach approvingly and then quotes an HDI or MPI rank without this caveat has taken the good half and left the bill.

The refusal of a canonical list left the selection to the institutions the approach exists to hold to account — and the current list has a hole this project should care about. The 2025 global MPI counts cooking fuel, sanitation, drinking water, electricity, housing and assets. It has no connectivity indicator at all. Its assets indicator marks a household deprived if it "does not own more than one of these assets: radio, TV, telephone, computer, animal cart, bicycle, motorbike or refrigerator, and does not own a car or truck" — a computer having been added only in a 2018 revision. That is a test of device ownership, not of reach: a household with a radio and a bicycle and no network of any kind is not deprived on this measure.

So the leading capability-based poverty instrument, in 2025, is blind to the capability this project spends every reading watching. Recorded flatly, without a recommendation: the MPI is not wrong to be conservative about adding indicators, since a measure that changes its list every year cannot show change over time. But the gap is real, it is checkable, and any reading that reaches for the MPI as the access instrument for AI capability should say that the instrument does not currently contain one.

The entitlement empirics are contested. Recorded above under Poverty and Famines; repeated here because it belongs in the grading. Sen's Bengal food-availability claim has drawn sustained technical refutation for forty years and has not been settled in his favour. The conceptual move survives the dispute; the specific historical claim should not be quoted as though it did not have one.

Commonly misused as

Not required for interpretation — the canon requires it for limit — but there are five misuses here that recur, and one of them is a trap specific to this project.

1. "Income doesn't matter." It is the most common misreading and it inverts the argument. Sen's claim is that income is a means whose conversion rate into capability varies, not that it is irrelevant. Development as Freedom is explicit that low income is among the principal causes of capability deprivation. The approach demotes income from the definition of poverty to a cause of it; that is a different operation from dismissing it, and a reading that uses Sen to wave away a wage figure has misused him.

2. "Poverty is relative." Sen argued the opposite and published a paper to say so. "There is an irreducible absolutist core in the idea of poverty. If there is starvation and hunger, then — no matter what the relative picture looks like — there clearly is poverty." The relativity is in the commodity bundle, not in the standard. Anyone citing Sen for the proposition that poverty is whatever a society says it is has the 1983 paper against them.

3. "The MPI is Sen's index." It is not. Sen supplied the framing and explicitly declined to supply the list, the weights or the cutoffs. The list is UNDP's and OPHI's; the method is Alkire and Foster's; the 1/3 weights and the k = 1/3 poverty cutoff are choices those authors made and defend on their own grounds. Attributing them to Sen makes his refusal disappear, which matters because the refusal is the thing his critics attack. Praise for the MPI is not evidence for Sen, and Ravallion's critique of the HDI's tradeoffs is not a refutation of Sen. Keep the layers apart.

4. The word collision, and this one is the project-specific trap. This canon uses capability constantly, in the LENSES.md sense: what a model or an agent can do. Sen's capability is what a person is free to do and be. The two senses are not related and a sentence that lets them touch produces confident nonsense — "capability is rising" would then mean both that GPT-class systems got better at olympiad problems and that more people can eat, which are different facts that can move in opposite directions in the same window. A reading citing this entry must say which sense it means, every time. The honest formulation usually needs both words: model capability rose; whether human capability rose is a separate question with its own evidence, and the second does not follow from the first. That non-inference is, in one line, the reason this entry is in the canon.

5. "A framework for declaring people poor because they lack a chatbot." The capability approach is a reason to ask whether a technology has changed what people can do and be. It is not a licence to write frontier-model access into the deprivation list by assertion. Sen's own test is whether the absence of the thing prevents a functioning people have reason to value — and establishing that for a given technology is empirical work nobody has done for this one. The honest current position is METHODOLOGY.md's: the interval between a capability being invented and being assumed is itself measurable and appears to be collapsing, and that is the thing to watch. The list can be revised when there is evidence, and the ratchet must be reported alongside the rate, or the measure will eventually publish something false.

Sources

Primary, read directly.

Secondary, used for specific facts and flagged as such.

Fetch failures, recorded per the rule that a failed search is noted and the work continues.