← The canon · AItopiaOrAImageddon?

The Hollerith electric tabulating system, and the 1890 census

idea · Herman Hollerith, who built it; John Shaw Billings, who suggested at the Census Office in 1882 that there ought to be a machine and afterwards declined the credit; and the United States Census Office under Superintendent Robert P. Porter, which set the competition, wrote the contract, and owned the population being counted · 1890

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.

Read on: The Turing–Welchman Bombe, and industrial cryptanalysis at Bletchley Park, UNIVAC on election night, CBS, 4 November 1952, The IBM PC — publishing an interface is not granting a right, and it did not matter, The Year 2000 problem — the catastrophe that did not arrive, and why that settles nothing.

Filed under idea in canon/proposals.md, and idea is right. This file confirms the label rather than correcting it, which is the less interesting of the two outcomes, so the argument is kept short and the length goes where it belongs — into the claim the proposal attached to the entry, which is not right as stated and which section 3 grades in pieces.

idea holds because there is a mechanism here you could build. The house test for the kind, set by shannon-chess-1950 and restated in printing-press-1440, is exactly that: something implementable, as against a mood or an event. This is implementable to the point of being patented three times over — US 395,781, 395,782 and 395,783, "Art of Compiling Statistics", filed September 1884 and June 1887, issued 8 January 1889 — and then built, and then run across roughly sixty-three million cards. build.py's gloss for 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." Both clauses land. Nobody argues about the punched card; they build on it. And the descent is not decorative: the German censuses of 1933 and 1939, the American Social Security ledger of 1936, the entire mid-century data processing industry and both of its corporate lineages come out of this machine in a line you can follow year by year.

Why not moment. There is no shortage of dates. The competition ran in the autumn of 1888 and was settled in 1889; census day was 1 June 1890; a preliminary national total was announced on 16 August 1890. llama-weights-2023 fixed the house test for moment as a date on which something visibly happened in public, and this could pass it. It is filed as idea anyway, because the citation occasion this canon actually has is not "on 16 August 1890 a number was announced." It is a standing condition about what a state acquires when it can count its population by machine — a condition that was as true in 1941 and is as true in 2026 as it was in 1890. Filing a standing condition as a moment would put the weight on the wrong end.

Why not interpretation, which is the serious rival and nearly wins. The proposal's payload is a sentence — "the capability to enumerate a population is the capability to select one" — and that sentence is a reading of events rather than a fact about a machine, which is interpretation's definition almost exactly. The reason it is not filed there is that the sentence is a claim about the object and not the object. The object is a card, a press, a mercury cup and a counting dial. So the entry is idea, and the interpretation the proposal hung on it gets graded in section 3 like any other dated claim — which is better treatment than it would get if the two were merged, because merging them would make the claim unfalsifiable by definition.

Why not limit, which is the mistake this entry most invites. "Enumeration is selection" reads like a law, and someone will eventually cite this file as though it established one. It does not. Section 3 produces a case where the mechanism was defeated — Jewish refugees in the occupied Netherlands who avoided registration and survived at a higher rate than registered Dutch citizens — and section 2 lists the safeguards that have actually bitten in practice. A law with documented exceptions and documented countermeasures is not a law. It is a strong tendency, which is a different and more useful thing.

descends_from is [], and the ancestor that looks obvious is not one. babbage-analytical-engine-1837 and lovelace-1843 are both in this canon; both are about a machine controlled by punched cards; and neither is an ancestor of this entry. Two reasons, and the second is the more important.

The first is evidential. Hollerith's own account of where the idea came from names two sources and Babbage is neither. One is John Shaw Billings, at the Census Office in 1882, who told him there ought to be "a machine for doing the purely mechanical work of tabulating population and similar statistics" and suggested the principle of the Jacquard loom. The other is a railway ticket: "I was traveling in the West and I had a ticket with what I think was called a punch photograph … the conductor … punched out a description of the individual, as light hair, dark eyes, large nose, etc." The Analytical Engine was barely known in 1880s America, existed as no working machine, and appears nowhere in Hollerith's telling.

The second is conceptual, and it is the whole invention. The Jacquard card and the Babbage card are instructions. They tell a machine what to do next; the holes are a program. Hollerith's card is a record. The holes are a person — age, sex, race, nativity, occupation, literacy — rendered machine-readable so that a machine can count them without understanding any of it. That inversion, from card-as-order to card-as-datum, is what later work is built out of, and Babbage did not perform it. Treating the two as one lineage because both involve perforated pasteboard is the genealogical equivalent of grouping the abacus with the calculator because both have beads.

The real ancestors — Joseph-Marie Jacquard's loom of 1801, Billings' suggestion, the conductor's punch photograph — are not in canon/, and this job does not invent ids. The descent that matters runs forward, and it runs to files this job does not write in: codd-relational-1970, where the record-about-a-person finally gets a formal theory; google-index-1998 and mass-collection-2013, where the scale stops being national; national-content-filtering and social-credit-narrative, where the administrative apparatus is the subject under dispute; and nineteen-eighty-four-1949, which is the fiction this entry is the machinery of, written fifty-nine years later by a man who had watched the machinery work.

On year: 1890. Defensible without argument. The patents issued in January 1889 and the mechanism existed before that, but the idea's public proof — the thing that made every subsequent state want one — is the census of 1890. build.py reads the first four-digit number and will mark the canon's chart at 1890, which is where a reader should file it.

---

What it is

Herman Hollerith was born in Buffalo in 1860 to German immigrant parents, took a degree at the Columbia School of Mines, and joined the Census Office in 1879 as a nineteen-year-old engineer working on the census of 1880. He therefore met the problem from the inside, as a young man doing the arithmetic by hand, which is the ordinary way important tools get invented and is worth stating because the mythology prefers a lone genius arriving from outside.

The problem was constitutional, not clerical. The census is not an optional statistical exercise; Article I of the United States Constitution makes the apportionment of the House of Representatives depend on a decennial enumeration. The 1880 census took somewhere between seven and eight years to tabulate by hand — the published volumes were not complete until 1888 — and the country was growing at about twenty-five per cent per decade. The arithmetic was blunt and everybody in Washington could do it: the 1890 count would not be finished before the 1900 count began, and at that point apportionment would be running on data a decade stale, permanently. The state paid for Hollerith's machine because the alternative was a slow constitutional failure.

The mechanism

One card per person. Hollerith sized the card to the dimensions of 1890 United States currency so that existing Treasury drawers, boxes and filing furniture would hold the sixty-odd million of them without anyone having to build new cabinets — a detail that reads as trivia and is actually the tell of an engineer who had done the job by hand and knew where the labour went. The census card carried twenty-two columns of eight punch positions. It was not a character encoding; each position was assigned a specific meaning, so that a hole in one place was "male" and a hole in another was "born in Ireland."

Punching. A clerk read the enumerator's paper schedule and transferred it to a card using a pantograph punch — a template with a stylus at one end and the punch head at the other, so the clerk aimed at a large labelled board rather than at the card itself. Output was on the order of five hundred cards a day per clerk. This step is entirely human, it is where the errors enter, and no part of it is automated in 1890.

Reading. The card went into a press. A bed of spring-loaded pins descended; where the card was solid the pin stopped; where a hole had been punched the pin passed through and dipped into a small cup of mercury beneath. Contact closed an electrical circuit, the circuit energised an electromagnet, and the magnet advanced a counting dial by one. The 1890 machine carried forty such dials. The operator's whole job was to lay a card in, pull the press down, lift it, and take the card out; an experienced clerk managed about eighty cards a minute, and the working figure was roughly seven thousand cards a day — around ten times what hand tabulation achieved.

Sorting. Beside each tabulator stood a sorting box: a cabinet of compartments, wired so that the lid of exactly one compartment sprang open as each card was read. The clerk dropped the card into whatever had opened. This is the part that makes the system a system rather than a counter. A single pass both accumulated totals and physically sorted the deck for the next question, so the office could ask for married farmers over forty and get both the number and the pile.

What the card did not carry was a name, and this fact does more work in this file than any other. The reasoning: 176 punch positions, each assigned a fixed meaning rather than functioning as a character code, cannot represent an arbitrary string; the machinery had no alphabetic capability whatsoever and none existed anywhere until around 1931; and every published description of the 1890 card enumerates its fields as characteristics — household relation, sex, age, race, nativity, occupation, literacy, veteran status — with no name field among them. The name stayed on the enumerator's paper schedule, which is where the legal record lived. A tabulator counts. It does not list. Section 3 turns on this.

The competition, and the numbers

The Census Office ran a trial in 1888–89 on 1880 data from four districts of St Louis. Three systems competed. Charles F. Pidgin's transcribed the data onto coloured slips; William C. Hunt's onto coloured cards; Hollerith's onto punched cards read electrically. The published figures are consistent across the Census Bureau's own accounts and the USPTO's: data capture took 144.5 hours and 100.5 hours for the two rivals against Hollerith's 72.5 — a respectable but not decisive win — and then tabulation took 55.5 and 44.5 hours against Hollerith's 5.5, which is not a margin, it is a different category of thing. (Sources agree on the numbers and are inconsistent about which figure belongs to Pidgin and which to Hunt; that assignment is not asserted here.) He won the contract and leased the Census Office something over fifty machines.

1890

Census day was 1 June 1890. On 16 August 1890 the Census Office announced a preliminary national total of 62,622,250. The final official figure was 62,979,766, a 25.5 per cent increase over 1880. Full tabulation of all the census's inquiries was complete in under three years against the previous decade's seven-plus, and Superintendent Robert P. Porter had official population figures in front of Congress by December 1890, which no census superintendent had managed before.

A note on the widely repeated line that the count came in "six weeks." Census day was 1 June and the announcement was 16 August, which is ten and a half weeks; the six-week figure is presumably elapsed tabulating time once cards began arriving rather than time from census day. That reconciliation is this file's, not a source's, and a later reading with access to the Census Office's own bulletins should check it.

The result was disbelieved, and the disbelief was partly correct. Public expectation had run to seventy-five million. New York City's Mayor Hugh J. Grant did not accept the federal figure and ordered the Police Department to recount the city between 29 September and 14 October 1890; the police census returned about two hundred thousand more people than the federal one. Later statisticians have generally concluded the 1890 census undercounted by somewhere between one and several million, with the shortfall falling hardest on African Americans, working-class households and young children.

This is the single most transferable fact in the section and it is almost never told alongside the triumph. The first machine-tabulated census in history was also the first to be publicly accused of being wrong, and the accusation had merit. The machine made the count fast. It did not make it true, because everything that determined truth — whether an enumerator knocked, whether a household answered, whether a clerk read the schedule correctly — happened upstream of the first punched hole. The tabulator processed the misses at eighty cards a minute along with everything else.

The business, and the fork

Hollerith founded the Tabulating Machine Company in 1896 (some accounts say 1895) to sell the system to railroads, insurers and other governments. He leased the machines rather than selling them, and sold the cards. The recurring revenue was in the consumable and in the meter, not in the hardware, and the customer never owned the capability. That is the ancestral form of every metered compute business on this project's board.

In 1911 the financier Charles Flint merged the Tabulating Machine Company with several other firms into the Computing-Tabulating-Recording Company; Hollerith took about $1.2 million and a consulting engineer's role, and retired in 1921. In 1924 Thomas J. Watson Sr. renamed C-T-R International Business Machines.

Meanwhile the customer revolted. S. N. D. North, Director of the Census, would not pay Hollerith's rental increase, declined to renew the contract, and set up the Census Machine Shop to build the government's own machines. In 1907 the Bureau hired James Legrand Powers as a technician; Powers' mechanical punched-card system took most of the 1910 census work; he formed his own company in 1911, renamed it the Powers Accounting Machine Company to reach a commercial market, and in 1927 it merged into Remington Rand — which became Sperry Rand, which built UNIVAC. Two industrial lineages came out of one census, they spent the next sixty years suing each other, and both are still legible on the board today in every argument about whether a state should rent a capability or build it.

The record's fate

On 10 January 1921, at about half past five in the evening, a fire started in the basement of the Commerce Building in Washington. The 1890 population schedules — the paper with the names on it — were stacked on pine shelves in an unlocked basement file room rather than in the fireproof vault where earlier censuses were kept, and were, in the phrase from the National Archives' own account, "first in the path of the firemen." Roughly a quarter was destroyed outright and half of the remainder damaged by water and smoke. Congress authorised destruction of the survivors on 21 February 1933 and the remnant was destroyed in 1934 or 1935.

The most technologically advanced census of its century is therefore the one American census whose individual records essentially do not survive. The cards went first, as waste stock; the paper went in the fire; and what remains is the aggregates. There is a lesson in there about which artefact a data system actually preserves, and it is not the one about the person.

---

Why a reading would cite it

The lenses this touches are Policy and regulation, The public square, Concentration, and the explicitly named harmful branch of AI as accelerant — "warfare, campaigning, mass manipulation, surveillance." Six concrete occasions, then the admission test answered honestly.

1. A reading meets a population-scale inference system being stood up by a government. A national identity register linked to biometrics; a residency or risk score computed over everyone; a policing platform that ingests checkpoints, cameras and utility records and emits flags. The current instance of record is Xinjiang's Integrated Joint Operations Platform, which Human Rights Watch reverse-engineered between January 2018 and February 2019 by taking apart the mobile app police use to reach it, publishing the forensic artefacts. What this entry supplies is a dated, non-speculative precedent for the specific mechanism rather than an atmosphere: a system built to administer a population acquires, with no change of intent and no change of design, the capacity to be pointed at part of it. Cite this instead of reaching for Orwell, which is fiction, which this canon already carries as fiction under nineteen-eighty-four-1949, and which lets a reader dismiss the point as literary.

2. A reading meets the claim that a data system is safe because its purpose is benign. This is the strongest thing in the file and it comes with numbers. The Dutch comprehensive population registration system was completed before the Nazi occupation, for administrative and statistical purposes. In 1938 H. W. Methorst, director-general of the Dutch Central Bureau of Statistics and formerly head of population registration, reported on its progress in exactly the language a modern civic-technology programme would use: a system following each person "from cradle to grave," opening "wide perspectives for simplification of municipal administration and at the same time social research." By early 1941 Methorst's successor, J. L. Lentz, had adapted that general system into special registers covering the Jewish and Roma populations, with linked identity cards, and later produced dot maps of Jewish residence by district.

Dutch Jews had the highest death rate of Jews in any occupied western European country: 73 per cent, against 40 per cent in Belgium and 25 per cent in France.

And then the near-control, which is the reason this paragraph is worth a reading's word count. Jewish refugees from Germany and elsewhere who were living in the Netherlands under the same occupation, in the same cities, subject to the same police, experienced a lower death rate than Dutch Jews. Seltzer and Anderson's explanation is that these refugees, unlike Dutch citizens, had avoided registration. That is as close to a natural experiment as this territory ever produces: same regime, same intent, same geography, different exposure to one administrative system, different survival.

3. A reading asks who is building this and why. The doom framing supplies a villain and the industry framing supplies a naïf, and the historical finding is neither. Seltzer and Anderson, having reviewed the twentieth-century cases, report that "ideology, patriotism, and fear seemed less decisive in determining complicity than bureaucratic opportunism and professional zeal." The people in their list — J. C. Capt and Calvert Dedrick in the United States, René Carmille and Henri Buhle in France, Gunnar Jahn in Norway, Richard Korherr in Germany, Lentz in the Netherlands — were, in their own words, "well-regarded heads of national statistical agencies" and "highly regarded senior technicians with extensive experience." Several took the initiative: in occupied and Vichy-influenced states the statisticians proposed new data-gathering efforts of their own accord, seeing an opening for work they had wanted to do anyway. This transfers directly and uncomfortably to the question of who is currently building population-scale machine learning infrastructure, and it is the finding a reading should reach for when the discourse wants a monster.

4. A reading needs the safeguards half, which the catastrophe framing reliably drops. These are real, they are specific, and they are the model for any argument about data minimisation:

5. A reading meets the argument that the law forbids this use. The counter-instance is domestic and unambiguous. United States census confidentiality provisions were set aside by the War Powers Acts in both world wars — and Seltzer and Anderson record that even before the Second War Powers Act was introduced in Congress in early 1942, the Director of the Census had already indicated a willingness to ignore the existing law in order to give defence authorities a way of checking confidential census information on individual Japanese Americans. Legal safeguards are the ones that fail first, and they fail at exactly the moment they are needed, because the perceived emergency that motivates the misuse is the same emergency that motivates suspending the protection.

6. Concentration. Hollerith leased; he did not sell. Dehomag leased; it did not sell. The lease is why the corporate relationship with IBM was continuous rather than a one-off sale in 1933, and it is therefore why there is an ownership question to argue about at all. When a reading is looking at who holds a capability and on what terms — weights against API, rented inference against owned silicon — this is the same question in its 1890 form, and the answer then was that the party who rents the machine does not control what it is pointed at but does keep collecting the rent.

The admission test, answered plainly. This entry earns its place on points 2 and 3 specifically, because those are quantitative, dated and transferable: a 73/40/25 spread with a quasi-control attached, and a documented finding about motive that contradicts the intuitive story. It would fail the test if the citation occasion were only "computers were once used for something terrible," which is a mood rather than a citation and which a reading should refuse to deposit. If a reading finds itself citing this entry without reaching for either the Dutch numbers or the professional-zeal finding, it is probably citing it decoratively and should stop.

---

What it got right, and what it got wrong

Not required for idea. Done anyway, because the proposal attached a dated claim to this entry, and a dated claim gets graded rather than repeated.

The claim, in its own words, from canon/proposals.md, proposed by the owner on 2026-08-17:

> Cite when: a data system built to administer a population is later used to > identify part of it — the punched-card census is the canonical case, in the > 1890s and again, catastrophically, in 1930s and 1940s Europe. The lesson is > not that the machine was misused; it is that the capability to enumerate a > population is the capability to select one.

Four separable assertions. Two hold, one holds strongly, one does not.

Right, and strongly: "the lesson is not that the machine was misused"

This is the best sentence in the proposal and it survives every check. Nothing that happened in 1933, 1939 or 1941 required the machines to be operated in a way their designers had not envisaged. Counting people by characteristic, and physically sorting them into bins by characteristic, is the specification. The 1890 sorting box opened the lid of the compartment matching the card; the Dehomag D11 sorter did the same thing faster. There is no misuse event to point at, no moment where someone repurposed the tool, and that absence is the whole finding. A technology whose harm requires no misuse is a different kind of policy problem from one whose harm requires an abuse, and readings that reach for "dual use" are reaching for the wrong frame here.

Right, on the general pattern: built for one thing, turned on a subgroup

Overwhelmingly borne out. Seltzer and Anderson's Table 1 lists ten documented cases of population data systems implicated in major human rights abuses: Germany 1933–1945; Poland 1939–1943; France 1940–1944; the Netherlands 1940–1944; Norway 1942–1944; Romania 1941–1943; the United States and Native Americans through the nineteenth century; the United States and Japanese Americans 1941–1945; the Soviet Union 1919–1939; and Rwanda in 1994, where a registration system built under Belgian colonial administration in the 1930s to fix the "Hutu" and "Tutsi" categories was still running, still producing monthly ethnicity-classified reports up the administrative chain, at the start of the genocide in April 1994.

Ten cases. Not one of them involved a system built for the purpose it was eventually used for. That is as clean a base rate as this canon carries for any proposition.

Right about the descent, and the descent is documented

Dehomag — Deutsche Hollerith-Maschinen Gesellschaft, the name says it — was founded in 1910 as the German licensee of Hollerith's patents and was taken over outright by C-T-R in 1923, the year before C-T-R became IBM. Dehomag machines tabulated the German censuses of 1933 and 1939; the United States Holocaust Memorial Museum holds a Dehomag D11 tabulator, a D11 sorter and a facsimile census punch card in its collection, catalogued as being of the type used for exactly that. The 1939 census carried a supplementary card asking after the religion of a respondent's grandparents — a question with no statistical function whatsoever and only a selective one, put to a whole population by a census whose one novel capability was that it could process the answers quickly. Straight-line descent from the machine described in section 1, through a corporate structure that never stopped being the same corporate structure.

Wrong, on the identity claim, and the correction is technical

The failing word is is. "The capability to enumerate a population is the capability to select one" asserts an identity, and enumeration and selection are not identical capabilities. They are separable, they were separated for forty-six years, and every safeguard that has ever worked attacks the gap between them.

The 1890 machine could not select. It carried no names, for the reasons given in section 1. You could ask the Census Office how many Chinese-born married farmers over forty lived in Alameda County and get a number back in an afternoon. You could not ask for their addresses, because that information was never punched, and no machine in the world could read a name off a card until roughly 1931.

Seltzer and Anderson's own taxonomy is the refutation, and it comes from the same paper that supplies the entry's strongest supporting evidence, which is what makes it worth trusting. They distinguish four things a data system can provide: macrodata (aggregates for large units), mesodata (aggregates for units small enough to be operational — a tract, a ward, a city block), microdata ("information, such as identification number or name and address, along with related information that identifies an individual as a member of a vulnerable group"), and staff and expertise. Their assessment of the first category is explicit: macrodata tabulations, "while important for policymaking, propaganda, and general administrative purposes, usually have limited value in planning or carrying out the operational aspects of a major human rights abuse." Three separable capabilities, three different thresholds, three different remedies. An identity claim erases all of that.

The chronology of the gap, and who closed it. Alphabetic punched-card equipment — machines that could carry and print letters rather than accumulate counts — arrived around 1931. In 1936 the Social Security Board took over six floors of the Candler Building in Baltimore and opened ledger accounts for something over twenty-six million people on IBM equipment, running six hundred thousand cards a day with over two thousand machine operators and checkers. That is the first American name-linked, machine-readable, population-scale register, it was built for a benign purpose by a reforming government, and it is forty-six years after 1890. The increment that turns enumeration into selection is name-linkage; it was added by Hollerith's own commercial descendants; and it took nearly half a century.

The honest formulation is the weaker one, and it is more useful precisely because it is weaker:

> The capability to enumerate a population is the precondition for the > capability to select one, and the two are separated by an increment — > name-linkage at population scale — that is cheap to add, invisible when added, > and almost always gets added.

That version has a policy consequence: every safeguard in section 2 attacks the increment rather than the count. "Enumerate is select" has no policy consequence at all, because if the two are the same thing the only available remedy is to stop counting, which no state will do and which section 4 explains is not even desirable.

And the empirical case runs against the identity claim too. If enumeration were selection, refugees in the occupied Netherlands avoiding registration would not have helped them, and it did.

Wrong, or at least imprecise, about the 1890s — and the real 1890s case is a better one

The proposal says the pattern shows up "in the 1890s and again, catastrophically, in 1930s and 1940s Europe." There was no 1890s American selection operation run off the tabulator, for the reason just given: it could not produce a list.

But there is a genuine 1890s harm, it is directly downstream of Hollerith's tabulations, and it is more interesting than the one the proposal reached for because it runs through the macro channel that the identity claim's framing misses in the other direction.

In August 1896 Frederick L. Hoffman, statistician to the Prudential Insurance Company of America, published Race Traits and Tendencies of the American Negro — 330 pages, under the auspices of the American Economic Association, built substantially on 1890 census data. Its argument was that Black Americans were biologically and morally degenerating, headed for extinction, and commercially uninsurable. It had been solicited in response to a wave of state legislation prohibiting racial discrimination by insurers, and it was used to justify continuing that discrimination in the face of those laws. Kelly Miller published A Review of Hoffman's Race Traits and Tendencies of the American Negro through the American Negro Academy in 1897 and took the statistics apart at the time — attributing the mortality differentials to social conditions rather than to race, which is the correct answer and was available immediately. Hoffman won anyway, in the sense that mattered, which was underwriting practice.

The chain here is careful and should be cited carefully: the tabulator produced the aggregates; the aggregates were published; a corporate statistician built a racial argument on them; an industry acted on the argument. No individual was ever identified by a machine. The machine's contribution was to make a detailed, racially disaggregated national picture available at a level of granularity and speed that had not existed before — the 1890 schedule's race question ran to white, black, mulatto, quadroon, octoroon, Chinese, Japanese and Indian, and a machine that sorts on that field will produce exactly the table Hoffman needed. That is the macrodata channel doing real harm at scale without touching a single name, and it is the case a reading should reach for when someone argues that aggregate statistics are inherently safe.

---

Commonly misused as

Not required for idea. Included because this entry sits on the most abused piece of computing history there is, and the abuse runs in both directions, which means a reading can go wrong here while trying to be careful.

1. "IBM did the Holocaust." The popular version descends from Edwin Black's IBM and the Holocaust (2001). Sorting it requires separating what is not in dispute from what is.

Not in dispute: IBM owned Dehomag outright from 1923. The machines were leased, not sold, so the commercial relationship was continuous rather than a single pre-war transaction. Dehomag equipment tabulated the 1933 and 1939 censuses. Dehomag's director Willy Heidinger was a public enthusiast for the regime, describing his machines to Nazi officials in the language of a physician diagnosing the national body and identifying foreign matter within it. IBM's own 2001 statement acknowledges that equipment supplied by Dehomag was used by the Nazi government, while noting that Dehomag came under the control of Nazi authorities during the war and that most records were lost or destroyed.

Disputed by historians: whether the technology was decisive. Michael Thad Allen's review in Technology and Culture 43:1 (January 2002), 150–154 — titled, unkindly and memorably, "Stranger than Science Fiction" — contests Black's causal linkages and argues the book exaggerates deliberate intent over what was in fact bureaucratic opportunism. Richard Bernstein, reviewing in the New York Times, accepted the documentation while questioning whether IBM bore unique or decisive responsibility. Black himself concedes the point that matters most: the Holocaust would have proceeded without IBM, "with simple bullets, death marches, and massacres based on pen and paper persecution."

The framing a reading should actually use is Seltzer and Anderson's, because it is the only one that yields a measurable quantity. The machines were an efficiency multiplier on a process whose binding constraints were registration and political will. The attributable harm is therefore not "the Holocaust" — it is the difference made by the data system, and the Dutch/Belgian/French spread is the closest thing to a measurement of it that exists.

2. The mirror misuse: "so the technology was irrelevant." More comfortable, equally wrong, and increasingly the reflexive position of people who have learned that Black overstates. The 73/40/25 spread is the answer, and the refugee comparison inside the Dutch case removes the obvious confound. Seltzer and Anderson state the general principle without hedging: even where a safeguard does not prevent the underlying abuse, "without the assistance of the data sought, the efficiency of the perpetrators is likely to be reduced. As a result, lives may be saved and the extent and duration of other harms reduced." A reading that cites this entry to argue that infrastructure is neutral has inverted it.

3. "The census is inherently a weapon; therefore stop counting." The anti-enumeration conclusion, and the one this entry is most likely to be recruited for by a reading in a dark mood. The record refutes it in two places. Seltzer and Anderson close their own paper with the reminder that population data systems "have also been important sources of evidence in the prosecution of human rights abuses." And they note that in both the United States and the United Kingdom, advocates for some vulnerable groups have been "among the strongest supporters of continued or more detailed coverage," because detailed identification is the statistical basis on which redress programmes are built and on which a group's presence in the body politic is recognised at all. The trade-off is genuine and unresolved; an entry deployed to argue "therefore do not count" is being used to close a question the sources hold open.

4. "Hollerith built the first computer." He did not, and the canon should not let this pass. The tabulator has no stored program, no conditional branch, no general computation and no memory in any useful sense; it is a counting machine with a sorting attachment and it is exactly as intelligent as a turnstile. Its descent into computing runs through the card format and through the company, not through the architecture. The distinction matters whenever someone draws a smooth 1890-to-2026 line implying a continuous ramp of machine intelligence: there are at least two genuine discontinuities in between, and this canon carries one of them at turing-halting-1936.

5. "The machines made the census more accurate." They made it faster. The 1890 undercount is the standing rebuttal — a count publicly disputed, recounted by the New York police to a difference of two hundred thousand in one city, and retrospectively judged short by one to several million nationally, with the shortfall concentrated on the people least able to object. Everything that determined accuracy happened before the first hole was punched. Mistaking throughput for accuracy is the specific error this entry exists to inoculate a reading against, and it is the reason a nineteenth-century tabulating machine belongs in a canon about AI at all: it is the same error, in its earliest clean instance, with a hundred and thirty-six years of hindsight attached.

6. Attribution of the idea, in both directions. "Billings invented it and Hollerith commercialised it" and "Hollerith invented it alone" both circulate. The defensible version is that Billings suggested a machine and by most accounts declined the credit afterwards, and Hollerith designed, patented and built it. The primary texts of that credit dispute were not reached in this run; see Sources.

---

Sources

Primary where reachable; every fetch failure and every unverified figure recorded.