← The canon · AItopiaOrAImageddon?

The Shape of Automation for Men and Management

prediction · Herbert A. Simon · 1965 — the book; the sentence is from his 1960 Ford Lectures

A dated claim, carried graded: the claim in its own words, when it was made, when it came due, and what actually happened. Where a predictor graded themselves, their grade and an independent one, saying they differ. The point of the kind is a base rate for how wrong AI forecasting runs.

Descends from The Logic Theory Machine, The Dartmouth Summer Research Project on Artificial Intelligence.

prediction is the right kind and proposals.md filed it correctly. Two other things about this entry are not correct as filed, and both belong in the first paragraph rather than in a sourcing footnote, because the canon has repeated them four times already and this file is the place they get fixed.

First: the sentence is not from 1965. It was published in April 1960, in The New Science of Management Decision, on page 38. The 1965 book reprints it verbatim as Part III. The horizon in the sentence is twenty years, so the claim came due in 1980, not 1985 — and the five-year drift is not Simon's doing, it is what happens to a dated forecast when a publisher reissues it and everyone afterwards cites the reissue.

Second, and larger: the sentence is not a boast. The Shape of Automation is an anti-alarmist book about employment. Its thesis is that automation will not produce mass technological unemployment, and the famous sentence is a premise Simon grants to the other side before taking it away from them: even supposing machines can technologically do everything, the labour market does not empty, because what gets automated is decided by cost and comparative advantage and not by capability. The sentence that follows it in the original text is the economic turn. The most-quoted line in the history of AI over-optimism is the setup for an argument against panic, and it is quoted, routinely and in print, by people making the panic case.

So this entry grades two predictions from one book, made by one man on one page, which is unusual and is most of the entry's value. The technological claim failed and the economic claim held. A base rate built only from the first is the base rate everyone already has. The pair is worth more.

1. What it is

The book

**Herbert A. Simon, The Shape of Automation for Men and Management, Harper & Row (Harper Torchbooks), New York, 1965, 152 pages.** Simon was then professor at the Carnegie Institute of Technology, where he had been since 1949; he would take the 1975 Turing Award jointly with Allen Newell and the 1978 Nobel Memorial Prize in economics, and he is the only person to have both. He was, by 1965, one of the four or five people whose opinion about machine intelligence carried the most weight in the world, and he was also a working economist with a real reputation in a second field. Both halves are load-bearing here.

The book is a collection, and its structure is the first thing a reader needs, because almost nobody who quotes it knows it. From the catalogue record:

> Preface — Introduction: computers and automation — I. The long-range economic > effects of automation — II. Will the corporation be managed by machines? — > III. The new science of management decision [The executive as decision maker / > Traditional decision-making methods / New techniques for programmed decision > making / Heuristic problem solving / Organizational design: man-machine systems > for decision making]

Part II is Simon's 1960 essay "The Corporation: Will It Be Managed by Machines?", written for Melvin Anshen and George Leland Bach (eds.), Management and Corporations, 1985 (McGraw-Hill, 1960) — a symposium held for the tenth anniversary of Carnegie Tech's Graduate School of Industrial Administration, at which fifteen contributors were asked to project twenty-five years forward. That volume's title is where the "1985" in most retellings actually comes from, and it is a different twenty-five-year horizon attached to a different essay.

Part III is The New Science of Management Decision (Harper, 1960) entire — the Ford Distinguished Lectures, volume III, which Simon delivered at New York University in 1960. It runs about fifty pages. Page 38 of it is page 95–96 of the 1965 book.

So the 1965 volume is, in its two most-cited parts, a 1960 book. This is not concealed: the parts carry their original titles and Part III is named after the 1960 book. Whether the Preface dates them is a thing I could not check, and it is the single most useful thing anyone with the physical volume could add to this file.

The sentence

> "Technologically, as I have argued earlier, machines will be capable, within > twenty years, of doing any work that a man can do."

The New Science of Management Decision (Harper, 1960), p. 38, in the chapter "Organizational Design: Man-Machine Systems for Decision Making"; reprinted in The Shape of Automation (1965), pp. 95–96.

Two small textual notes. The original reads "any work that a man can do"; the form that circulates — including in proposals.md, in dartmouth-1956, and in most of the secondary literature — drops the that. It changes nothing and it is a reliable tell that a given citation came from the quotation chain rather than the page. And the opening clause, "as I have argued earlier," matters more than it looks: Simon is not announcing a forecast here. He is recalling one he takes himself to have established in the preceding pages, in order to use it as an input to an economic argument. The most famous prediction in the field's history is, grammatically, a subordinate clause.

What comes immediately after it

This is the half that never gets quoted, and it inverts the meaning of the half that does. Having granted full technological capability, Simon immediately turns to cost:

That list is a sixty-six-year-old forecast of which occupations survive automation, and §5.4 grades it. Note now only that it is specific, which the headline sentence is not, and that specific forecasts are the ones worth having.

The other sentence

Quoted by Hubert Dreyfus in 1967, from the same text:

> "Automation of a flexible central nervous system will be feasible long > before automation of a comparably flexible sensory, manipulative, or > locomotive system."

Some renderings have comparatively for comparably; I could not resolve which the page says, and it does not bear on the meaning.

This is Moravec's paradox, stated as a thesis, in 1960, twenty-eight years before moravecs-paradox-1988. Not noticed in retrospect, not implied by an example list — asserted, prospectively, as an ordering claim about which kind of automation arrives first. And it is right. §5.3 grades it, and it is the best thing in the book.

Dreyfus quoted it because he thought it was a concession, and he was correct that it sits badly beside "any work that a man can do." If flexible manipulation and locomotion are much further off than cognition, then the set of work a machine can do at twenty years cannot be any work a man can do, because a great deal of what men did in 1960 was manipulation and locomotion. Simon's own list of surviving occupations, above, says the same thing — its first two items are flexible manual work and operating machinery. The headline claim and the qualifications on the same few pages are not consistent with each other, and the observation is Dreyfus's, not mine. I have not read the intervening pages and so cannot rule out that Simon reconciles them somewhere; on the evidence I have, he does not.

The argument the book is actually making

The context is a specific public panic with a specific document at its centre.

On 22 March 1964, the Ad Hoc Committee on the Triple Revolution sent an open memorandum to President Johnson naming three simultaneous upheavals — cybernation, the nuclear arms race, and human rights. Its cybernation section argued that a combination of computers and automated machinery was breaking the link between work and income permanently, and that the remedy was an unconditional income. Signatories included Linus Pauling, Michael Harrington, Gunnar Myrdal and Tom Hayden. It was a serious document signed by serious people and it set the terms of the automation argument for the rest of the decade.

Simon's book is a reply to that world. His position, stated to The New York Review of Books in May 1966 when Robert Heilbroner pressed him on it:

> "The essence of automation and any kind of technological progress is to permit > fewer to produce more. But on balance, workers have not been displaced from > the labor market."

> "At the end of a half century of rapid technological advance, at nearly the > same pace as at present, unemployment is below 4 per cent and dropping."

> "economic institutions, not technology, determine the level of employment"

> "the world's problems in this generation and the next are problems of scarcity, > not of intolerable abundance."

Heilbroner, reviewing the book in the same magazine on 17 March 1966, drew exactly the line this entry is built on — conceding the technology and refusing the economics:

> "It is difficult for me to judge the validity of Professor Simon's > technological radicalism, although what he writes about the nature of > managerial decision-making and the possibility of duplicating it technically > strikes me as original and important."

His objection was that the Law of Comparative Advantage would not save the labour market once technology reached the service sector, and that Simon had assumed rather than shown that "the market will serve, at the terminal point of that transition, as an adequate mechanism for the distribution of income to the population in general, or of jobs to those who will work."

Heilbroner's phrase "technological radicalism" is, as far as I can trace, the origin of the formula that now circulates as Simon's self-description — a technological radical but an economic conservative. I found the formula attributed to Simon in one search summary and could not verify it at any source. It is an excellent description of his position and I am not able to say it is his sentence, so this file uses it only as a description.

2. On the kind, the year field, and descends_from

prediction is correct and the alternatives are worth naming. jobs/canon.md says the kind exists to assemble a base rate for how wrong AI forecasting runs, and requires four things: the claim in its own words, the date made, the date due, and what actually happened. All four are available and all four are unusually clean here — a printed sentence, a fixed publication date, an arithmetic horizon, and an outcome nobody disputes.

interpretation was the real temptation, because the book's lasting content is an analysis — the cognitive-comparative-advantage account of which tasks go to which kind of agent, which is what economists still read it for. But that analysis is itself expressed as dated claims about 1980, and a dated claim goes to prediction. moment is wrong: the book had no discernible effect on anything, and its most visible contemporary effect was a magazine argument with Heilbroner.

The year field carries a qualifier rather than a bare number, for the same reason minsky-1970 put its attribution dispute in creator: a bare 1965 would smuggle the wrong date into a metadata field where no reader sees it argued. It is written so that the site's four-digit extraction still reads 1965 and the entry still sorts where its id says it should.

descends_from holds two ids, and both are Simon's own.

logic-theorist-1956 is the strong edge and it is evidential rather than rhetorical. The Logic Theorist is why Simon believed this. He and Newell had, within four years of the forecast, built the first program to do something everyone agreed required thinking, and had watched it find a proof of theorem 2.85 shorter than the one in Principia Mathematica. A man who has just seen symbolic search do mathematics has a real empirical warrant for expecting the rest to follow, and the twenty-year figure is an extrapolation from a result he personally produced rather than from a mood. That entry also records the January 1957 lecture where Simon first went public with a ten-year horizon; this one is the same forecaster, three years later, widening the claim from theorems to all work while doubling the horizon.

dartmouth-1956 is the institutional edge and the canon has already drawn it from the other end: that entry names simon-1965, minsky-1970 and kurzweil-2005 as the series it heads, and is explicit that the confident generation-scale forecasts belong to the individuals and not to the 1955 proposal. Simon came to Dartmouth for a few days at the end and described the Logic Theory Machine. The series runs: one summer (1955), ten years (1957), twenty years (1960), three to eight years (1970), twenty-four years (2005).

Two edges I did not draw, and why.

wiener-1960-automation is the same subject in the same year reaching the opposite conclusion, and it is a contrast rather than a descent. Wiener's December 1959 lecture, printed May 1960, is the canonical statement that automation is dangerous because full capability and full obedience do not coexist; Simon's April 1960 lecture is the canonical statement that automation is economically benign. Same season, same technology, opposite verdicts, and the canon now holds both. §4.3 does something with that.

moravecs-paradox-1988 is the other one. The ordering claim in §1 predates it by twenty-eight years, so if anything the descent runs the other way — but Moravec did not cite Simon as far as I know, and I am not going to assert an influence I cannot document. It is prior art, stated in prose, in a management book, and that is all I can establish.

3. Three dates are in circulation, and none of the three is wrong the same way

This section exists because the canon's own entries repeat one of them and this file is where it gets corrected.

1956. Steve Lagerfeld, "Lessons of the First Automation Crisis," The American Interest, 15 February 2020, gives the sentence as Simon's, "declared in 1956." That is four years early and there is no 1956 source for it. It is probably a collision with Simon's 1956–57 chess forecast, which is a different claim with a different horizon.

1960. The scholarly date, and the correct one. Quote Investigator traces the sentence to The New Science of Management Decision (Harper, 1960), p. 38, from lectures delivered at NYU in April 1960. Melanie Mitchell's Why AI is Harder Than We Think (2021) cites it to 1960. It is the date every careful source uses.

1965. The popular date, and the one this entry's id carries. It comes from the reprint, and it entered general circulation through two specific channels: Hubert Dreyfus cited the 1965 book in The Review of Metaphysics in 1967, and Ray Kurzweil cited it in American Scientist in 1985. Brynjolfsson and McAfee's The Second Machine Age (2014, p. 141) carries it the same way. When a 1960 claim is famous because two influential critics quoted the 1965 printing, the 1965 printing becomes the date.

What follows from this, and it is not pedantry. The horizon is twenty years from the claim, and the claim is 1960. Every citation to "Simon, 1965" grants the forecast five extra years it was never given. Nobody did this on purpose; there is no manoeuvre here and no dishonesty. But the effect is identical to the manoeuvre amodei-2024-loving-grace records — the due date moves forward while the claim stays fixed — and it is worth knowing that a forecast's horizon can drift by five years through nothing more than a reprint and a citation habit. A reading grading a 2026 claim should write down the date of the first publication, not the date of the printing it happens to be holding.

This file therefore grades against 1980 as the true due date and reports 1985 as the popular one. Both have passed. Nothing in the grade turns on the choice, which is why it is safe to be exact about it.

And the canon should be corrected rather than quietly patched. dartmouth-1956, minsky-1970, musk-robotaxi-2019, dijkstra-1959 and amodei-2024-loving-grace all refer to this claim as "Simon in 1965" or "Simon's twenty years." They are repeating the standard citation, which is what they should have done before this entry existed. The id stays simon-1965 because an id is a handle and changing it would break every reference; the date is fixed here, in prose, where a reader will meet it.

4. Why a reading would cite it

4.1 The two-track grade, which is the thing this entry has and nothing else does

Here is the case, stated as plainly as it can be:

A senior figure made two dated predictions on the same page. The one about what machines would be able to do was badly wrong. The one about what that would do to employment was substantially right. He is remembered exclusively for the wrong one.

That is a citation occasion that recurs in this project almost weekly, because the readings' Capabilities lens and their Work and the economy lens routinely disagree, and the disagreement is usually treated as a contradiction to be resolved. It is not. It is the normal state of affairs, it has a sixty-six-year-old worked instance, and the instance comes with both halves graded.

The concrete form: when a reading records a large capability claim and then records that no displacement data appeared in the same window, the two are not in tension and Simon is why. Capability and deployment are separated by price, by comparative advantage, and by the fact that automating a task is not automating a job. frey-osborne-2013 is the modern claim that put a number on the same question and it is the one payroll data grades; this entry is the sixty-year-older statement of the mechanism that makes the number hard to hit.

4.2 The live occasion, on the day this file was written

The 17 August 2026 midday reading supplies both halves inside one digest, which is why this entry got researched now.

Under Work and the economy, the reading recorded, in full:

> Nothing notable found since the last reading. The labour-adjacent items in > Monday's coverage are older than the window — Oracle's planned August cuts were > reported on 12 August, OpenAI's enterprise-token report on 13 August — and no > hiring, displacement or productivity data landed inside it.

Under Compute and infrastructure, the same reading recorded roughly $3 trillion of off-balance-sheet AI commitments read out of nine companies' filing footnotes, and a signed 20-year lease on a 10GW campus with about $105bn of financing backstopped by the chip supplier.

Three trillion dollars of commitments and a nil return on displacement, in the same document, on the same day. A reading that meets that pattern and reaches for "the capability must not be real" or "the displacement must be coming and just hasn't been measured" is reaching past the available explanation. Simon's is that capability and employment are connected by an economic mechanism with slack in it, and that the slack is where sixty years of automation went. He was right about that in 1960 and the reading of 17 August 2026 is another data point for him.

The same reading supplies the sharper item for the other half. Under Robotics and embodiment: SoftBank putting $200m into Gravis Robotics, an ETH Zurich spinout that retrofits existing excavators for autonomous digging and loading, flagged honestly as capital rather than a capability demonstration. In August 2026, the frontier of machines doing physical work in unstructured environments is bolting autonomy onto construction equipment somebody else built. Simon in 1960 named flexible manual work and operating machinery as the two things machines would get last. Sixty-six years later the news is a company whose entire proposition is that machines still cannot do the first, so it is selling the second.

And under The public square, the 3M expert-witness item — a paid engineering expert who generated 85–90% of his report with ChatGPT from prompts including "show how 3M is 0% at fault," in a wrongful-death case, discovered through a citation artefact rather than disclosure. That is not a capability story either. It is nonprogrammed problem-solving — the third item on Simon's list of work humans would keep — being partially delegated to a machine by a human who remained the one answerable for it. Simon's framework says the task moved and the job did not. The jury assigned 3M 30% of the fault and awarded $61m, which is the institution doing what Simon's argument assumes institutions do.

4.3 When a horizon is asserted for full automation

proposals.md filed this entry with its occasion already written — "Cite when: a twenty-year horizon is asserted for full automation" — and it is worth being exact about what the citation should and should not do.

It should not be swung as a refutation. "Simon said twenty years in 1960 and he was wrong, therefore the current claim is wrong" is bad reasoning, and this canon has said so before in minsky-1970 and it is still true. The field of 1960 had a theorem-prover and a checkers player; the field of 2026 ships systems that a practitioner runs from a 17GB file on a laptop. A base rate is not a refutation.

What it should do is force the two questions Simon's own page forces. First: what is the argument for the number, as against the direction? Simon had one — an extrapolation from the Logic Theorist and from hardware trends — and it was wrong by a factor of at least three and counting. Second, and this is the part that is peculiar to this entry: does the claimant also hold a view about what the capability implies economically, and are the two views consistent? Simon held both and his are the pair the canon can grade. Most 2026 principals assert the capability horizon and leave the economics implicit, which makes the implicit half unfalsifiable, which is the thing LENSES.md built its rule against:

> "A person's shipped work deposits; a person's forecast does not."

4.4 Simon and Wiener, four months apart

Worth a short paragraph because the canon now holds both and a reading will eventually need to cite them together.

December 1959: Norbert Wiener delivers "Some Moral and Technical Consequences of Automation" (wiener-1960-automation), printed the following May. The argument is that a machine capable enough to be useful is capable enough to pursue the objective you literally gave it rather than the one you meant, and that you may not be able to intervene in time. It is the founding statement of the control problem.

April 1960: Simon delivers the Ford Lectures at NYU. The argument is that machines will be able to do everything within twenty years and that this is economically fine.

Same technology, four months apart, from two men of comparable stature. Wiener was right about the shape of the risk and wrong about nothing in particular because he refused to date anything. Simon was wrong about his date and right about the economics. The two files together are the canon's cleanest demonstration that "how fast," "how dangerous" and "how disruptive" are three independent questions, that a person can be excellent on one and poor on another, and that a reading which treats an authority's answer on one as evidence about the others is doing something the record does not support.

There is one more thing in that comparison and it is uncomfortable for Simon. Wiener's forecast cost him nothing because he attributed it to others, gave it a range, and declined to endorse it. Simon's cost him sixty-six years of being the standing example of AI over-optimism, and he paid that price for one subordinate clause in a book whose actual thesis he was right about.

5. What it got right, and what it got wrong

5.1 The claim, the dates, the outcome

The claim, in its own words: "Technologically, as I have argued earlier, machines will be capable, within twenty years, of doing any work that a man can do."

Date made: April 1960, published in The New Science of Management Decision (Harper, 1960), p. 38. Reprinted without change in The Shape of Automation for Men and Management (Harper & Row, 1965), pp. 95–96.

Date due: 1980 from the claim. 1985 from the reprint, which is the date the secondary literature grades against.

What happened by 1980. Nothing resembling it. The British funding consequences of lighthill-1973 had run most of their course; ARPA had moved from open-ended to mission-directed funding from 1974 and cancelled the speech-understanding programme in 1976. The commercial AI industry that would briefly exist had not yet started — Symbolics was founded in 1980.

What happened by 1985. This is the interesting one, because 1985 is not a trough. It is close to the peak of the first commercial AI boom, and the prediction still fails without argument. expert-systems-collapse-1987 records what the money bought: corporate spending worldwide "over a billion dollars" by 1985, Lisp machines at $36,000 and up per seat, and the flagship production system — DEC's XCON — configuring computer orders from a rule base, real and paying for itself and bounded. A machine capable of doing any work that a man can do would, in 1985, have been able to configure a VAX, and could not drive, lift, converse, negotiate, teach, nurse, or hold a screwdriver. Two years later the industry that built the configurator collapsed.

And the claim was graded as failed in print in its own due year, by the person who would later become this canon's standing example of a generous self-grader. Ray Kurzweil, "What Is Artificial Intelligence Anyway?", American Scientist 73(3), May–June 1985, pp. 258–264, quotes Simon's prediction at p. 261 among AI's unfulfilled promises. kurzweil-2005 records Kurzweil scoring his own hit rate at 86% against outside recounts of 7–42%. In 1985 he was scoring someone else's, accurately and on time. Nobody is uniformly generous or uniformly strict; people are generous about themselves. That is the finding, it is small, and it is the sort of thing that only shows up if you check who did the grading.

What happened by August 2026. Not met. The clearest available statement is the one the 17 August reading makes without meaning to: the frontier of machines doing physical work in unstructured environments is a company retrofitting excavators. On the language and reasoning side the picture is genuinely different — see §5.3 — but "any work that a man can do" is not a claim about language and reasoning, it is a claim about any work, and the set of work machines cannot do in 2026 is still enormous and still contains most of what most people are paid for.

Grade: a miss, unambiguous, at both due dates, still open. Sixty-six years elapsed against a twenty-year horizon — at least 3.3× and not closed. For calibration inside the canon: minsky-1970 is worse (three to eight years, 56 years elapsed, roughly 7×) and remains the largest single miss here by proportion; turing-1950 is comparable in kind, wrong by roughly a generation with the direction right; the four Simon–Newell 1958 predictions graded in deep-blue-1997 were wrong by a factor of three to six, every one, in the same direction. This forecaster's errors are consistent in sign and roughly consistent in magnitude across three decades, which is more informative than any single one of them.

5.2 The claim is under-specified, and the under-specification is the lesson

"Any work that a man can do" has no rubric. It cannot be checked against a task list, it does not say whether "work" means an economic occupation or a physical act, and it does not say what standard of performance counts. Sixty-six years on, a determined defender could argue the language half is met and a determined prosecutor could point at plumbing, and neither would be refutable.

Compare it to the two forecasts on either side of it in the canon, both by Simon or about him. The 1958 Simon–Newell predictions graded in deep-blue-1997 name four specific achievements with dates — chess champion, new theorem, music critics accept, psychology theories as programs — and every one of them is checkable, which is why deep-blue-1997 can grade them item by item and extract a factor. minsky-1970's quoted list names five concrete tasks — read Shakespeare, grease a car, play office politics, tell a joke, have a fight — and that entry grades them two-for-five and gets a real finding out of the split.

A forecast with a rubric can be graded and a forecast without one cannot, and the one without is the one that got famous. This is not an accident. The unrubricked version is more quotable, survives compression better, and never has to be wrong on a particular day. When a 2026 reading meets a capability forecast, the useful move is the one minsky-1970 arrived at from the other direction: write down the task list verbatim if there is one, and note its absence if there is not, because the absence is itself information about the claim.

5.3 What he got right, and it is bigger than what he got wrong

The ordering claim. Correct, prospective, and twenty-eight years early.

> "Automation of a flexible central nervous system will be feasible long before > automation of a comparably flexible sensory, manipulative, or locomotive > system."

Graded against August 2026: right, and by a margin that is still growing. The cognitive half has substantially arrived — text, reasoning, code, translation, analysis, image understanding — and it arrived, as alexnet-2012, transformer-2017 and scaling-laws-2020 record, by a route Simon would not have recognised. The sensory-manipulative-locomotive half has not. The practical frontier in 2026 is teleoperated demonstrations, scripted warehouse motion, and retrofitting autonomy onto machines built for humans to sit in. Dexterous manipulation and cross-environment generalisation are named by practitioners as the open problems.

moravecs-paradox-1988 records the canonical version of this observation, made retrospectively and explained evolutionarily: what evolution spent a billion years on is expensive to automate, what is evolutionarily recent is cheap. Simon had the ordering in 1960, forward-looking, with no explanation attached. minsky-1970 found the same ordering sitting accidentally inside Minsky's 1970 task list — the recent symbolic items fell, the ancient embodied ones did not. Three independent arrivals at the same result, in 1960, 1970 and 1988, and the earliest is the only one that was stated as a claim about the future.

That is worth a reading's citation on its own terms, and here is the sharpest form of it: the same page contains the field's most famous wrong forecast and one of its most durable right ones, and the wrong one is wrong partly because the right one is right. If manipulation is far behind cognition, "any work" cannot land on the cognitive schedule. Simon wrote down the reason his own headline claim would fail, in the same lecture, and did not follow it.

The economic claim. Substantially right, and it is the book's actual thesis.

Simon's position — that automation does not empty the labour market, that comparative advantage and cost govern what is automated, and that "economic institutions, not technology, determine the level of employment" — was the minority position in 1964 against the Triple Revolution memorandum. Graded:

Richard N. Langlois, who studied the 1960 essay closely — "Cognitive Comparative Advantage and the Organization of Work: Lessons from Herbert Simon's Vision of the Future," Journal of Economic Psychology 24(2), 2003, pp. 167–187 — states the split directly in his abstract: "Perhaps unlike Simon's better-known predictions about progress in artificial intelligence research, the predictions of this 1960 article hold up remarkably well and continue to offer important insights." That is an independent grader, in a peer-reviewed journal, drawing the same two-track distinction this entry is built on, twenty-four years ago.

A caution the entry has to state, or it is doing propaganda. "Simon was right about employment" is a claim about aggregates over sixty-six years, and it is not a claim that displacement does not hurt people or that it did not happen. It did: US manufacturing employment peaked in 1979 and the communities that depended on it were not made whole by the aggregate holding. Heilbroner's 1966 objection — that the market may not distribute the income or the jobs even if it clears — was not answered by the aggregate, it was sidestepped by it. The correct citation is that the mass-unemployment prediction failed, not that the transition was benign. A reading that flattens the second into the first is using this entry to say something Simon's own critic already showed it cannot say.

5.4 The occupation list, graded

Simon named four kinds of work humans would retain. This is his only rubric, so it should be graded like one.

"Flexible manual work" — retained. Emphatically. Plumbing, electrical work, construction, care work, repair. Sixty-six years and the mechanic's job is untouched, which is the same grade minsky-1970 gives to "grease a car."

"Operating machinery" — retained, and this is the interesting one. Simon expected humans to keep operating machines rather than be replaced by them, and in 2026 the leading edge of construction robotics is a company whose whole proposition is retrofitting autonomy into machines designed for a human operator. Partly met: the task is being automated at the margins while the fleet is still built around a seat.

"Nonprogrammed problem-solving" — retained, but this is where the pressure is now. This is Simon's own term for the work he thought machines were furthest from and it is precisely the ground 2026 systems are contesting. The 3M expert who generated most of a report with a chatbot, the Wiz agent that found and exploited a command-injection flaw with no human in the loop — both are nonprogrammed problem-solving being partly done by a machine. Simon's category is holding at the level of the job and eroding at the level of the task, which is exactly the distinction the Bowen Commission drew in 1966 and exactly the distinction a 2026 reading needs when it meets a capability demonstration and has to decide whether it is a labour finding.

"Face-to-face service" — retained, and the honest grade is complicated. Nursing, teaching, hairdressing, hospitality — all still human at scale. But the canon's her-2013 occasion is that a substantial category of human-facing interaction has moved to machines in a way Simon's framework has no slot for, because he was thinking about services that require a body in a room and not about companionship, counselling or customer support at the end of a network. Call it right about the occupations he had in mind and blind to a category that did not exist.

Three of four holding, one partly eroded, and the erosion is on the category he named as furthest away. For a sixty-six-year-old list of which jobs survive automation, made by a man in a management lecture, that is a much better record than the sentence he is remembered for.

5.5 The self-grade and the independent grade, and they differ

jobs/canon.md requires this section and this entry has an unusual answer to it.

Simon never graded this claim. I searched for a retrospective on the twenty-year forecast and found none. What exists is three things, and together they are a distinct manoeuvre the canon has not recorded before.

First — on the sibling forecast, he conceded the miss and reframed the enterprise. George W. Baylor (1997–98): "He was stuck with his prediction that a computer would beat the world chess champion within ten years — it took 40." Hans Berliner (2001), recalling conversations with him: Simon "always said that it was important for science to make such predictions as they galvanized the field toward certain important goals," and would point to Leonardo's drawings of human flight. Sourcing caveat, and it matters: both are other people's recollections, and Berliner's is of undated private conversation. It is the weakest evidence in this file and it is doing real work, so it should be treated as reported rather than established.

Second — in October 2000, four months before his death, he restated the substantive claim with the date removed. At Carnegie Mellon's Earthware Symposium, Simon said computers "will be doing most of the things intellectually, that humans do in a relatively short time" — the same claim as 1960, forty years later, with "twenty years" replaced by "a relatively short time" — and asked the audience "how far will and should we humans adopt the role of spectators in a world where much perhaps all of the world's necessary work can be done by computers?" In the same talk he declined the whole activity: "I don't believe that predicting the future is really what we're about… our task is not really to predict it. Our task is to design it, to design a sustainable, and acceptable world."

Third — that is the manoeuvre, and it is a fourth one. minsky-1970 sets out three ordinary, non-dishonest ways a forecaster can relate to a failed forecast, each of which destroys the record differently:

Simon's is the fourth and it is the most gracious of them:

None of the four produces a scorable data point from the forecaster's own account. jobs/canon.md says a base rate assembled from self-assessments would be worthless; four worked cases now say why, and the fourth is the one that most resembles what a thoughtful person actually does. "Our task is not to predict the future but to design it" is a good sentence. It is arguably the correct attitude for a scientist. And it removes the last dated claim from the record.

The independent grades, and they agree with each other and not with him.

They differ from his, and the difference is not about the facts. Every independent grader scores it a miss. Simon never contested that and never conceded it either; he moved, over forty years, from a dated claim to an undated one to a rejection of dating. A reading citing this entry may say the forecast failed; it may not say Simon admitted it, because he did not, and it should not say he denied it, because he did not do that either.

5.6 One thing the shape of the claim gets right about forecasting itself

Simon's horizon was twenty years. Armstrong and Sotala's 2012 analysis of 95 AI timeline predictions found that they cluster at 15 to 25 years out from whenever they are made, regardless of when that is. (That paper's other headline — that experts and non-experts predict alike — was subsequently retracted after AI Impacts found a column misread; the clustering result is not what the error touched and stands. minsky-1970 sets out the correction in full and a reading should not repeat the retracted claim.)

Simon's is a textbook instance of the clustering; Minsky's three-to-eight is not. The canon now holds one forecast from the fat middle of the distribution and one from its short tail, both failed. That is worth noticing before treating "fifteen to twenty-five years" as a signal that a claim has been thought about: it is the modal shape of an AI forecast, which means it is the shape of the ones that were wrong too.

6. Commonly misused as

Not required for prediction. Included because the misuses of this particular sentence are unusually clean, unusually consequential, and one of them is committed in the canon's own files.

"Simon predicted mass unemployment from automation."

This is the exact inversion of what the book says, and it is in print. Steve Lagerfeld's "Lessons of the First Automation Crisis" (The American Interest, 15 February 2020) presents the quotation as exemplifying "the pessimistic strand of mid-century automation anxiety," alongside Wiener. Simon was the most prominent academic opponent of that anxiety; he wrote a book against it, argued with Heilbroner in public over it, and was vindicated within two years by a federal commission.

The mechanism of the error is worth naming because it generalises: a sentence that grants the strongest version of an opponent's premise, in order to defeat the conclusion, reads as an endorsement of the premise once it is separated from the argument. Steelmanning does not survive quotation. Any reading that quotes a single sentence from a longer argument in this project is exposed to the same failure, and the defence is to check what the sentence is doing in its paragraph before deciding what it says.

"Simon said it in 1965."

Common, harmless-looking, and it silently grants the forecast five years. §3. The canon's own entries do this and this file is the correction.

"The early AI researchers were recklessly overconfident, and today's are the same."

The first clause needs qualification and the second does not follow.

Simon's forecast was wrong and the confidence was real, but the warrant was not nothing: he had personally built the first program to prove theorems and had watched it produce a proof shorter than Principia's. That is a genuine empirical basis for extrapolation, and it is a different thing from bravado. logic-theorist-1956 is the reason his 1960 confidence was reasonable and the reason it was wrong — he generalised from a result whose method did not generalise.

And the field of 1960 was not uniform. Dreyfus said so publicly in 1965, was correct, and was treated as an enemy for it; John McCarthy's own attested range was "five to 500 years"; Richard Bellman published a critical response to the Simon–Newell paper in Operations Research itself. The recklessness story is assembled by quoting the loud and dropping the record of the people who objected at the time, which is a thing that will also be true of 2026 when it is history.

The transfer to today does not go through automatically in either direction. scaling-laws-2020 gave the current generation something the 1960 generation did not have — a measured regularity rather than an intuition — and whether it extrapolates as far as claimed is precisely the open question. "They were wrong then" is not evidence about it.

"The overclaim caused the AI winter."

No, and this canon is consistent about it. lighthill-1973 traces the British funding retreat to the Lighthill report and the Science Research Council's response; the American retreat ran through the Mansfield Amendment, ARPA's move to mission-directed funding from 1974, and the cancellation of the speech-understanding programme in 1976; expert-systems-collapse-1987 traces the second collapse to hardware economics — general-purpose workstations getting cheap enough to run Lisp — rather than to disappointment. A management book is not one of those mechanisms, and I found no evidence in any source reached that The Shape of Automation influenced a single funding decision. Its documented effect is an argument with Robert Heilbroner in a literary review.

"Comparative advantage guarantees there will always be jobs."

The strongest version of Simon's economic argument, and it is stronger than its opponents usually allow and weaker than its proponents claim. Heilbroner's 1966 objection stands unanswered on its own terms: comparative advantage explains why humans keep some work at some price, and says nothing about whether that price supports a life or whether the market distributes the resulting income to the people displaced. The record grades the aggregate claim and does not grade the welfare claim, because the aggregate is measured and the counterfactual is not. frey-osborne-2013 is where this canon holds the modern numerical version of the question, and a reading should cite these two together rather than either alone.

"Simon was a hype man."

He was a man who made one badly wrong forecast, one durably right economic argument, and one correct ordering claim, on adjacent pages, and then declined for forty years to grade any of them. That is a more interesting and more common kind of person than a hype man, and the canon is worth more for holding him accurately.

7. Sources

Primary — the artefact. Herbert A. Simon, The Shape of Automation for Men and Management (Harper & Row / Harper Torchbooks, New York, 1965), 152 pp., and The New Science of Management Decision (Harper, New York, 1960, Ford Distinguished Lectures vol. III, delivered at New York University), c. 50 pp. I read neither. Open Philanthropy's scan at openphilanthropy.org/files/Focus_Areas/AI/Simon_(1965).pdf now returns a 301 to the root of coefficientgiving.org and the file is gone; the Internet Archive's copy (archive.org/details/shapeofautomatio00simo) is access-restricted and its full-text search endpoints were unreachable from here. Every quotation from Simon's books in this file is second-hand, and that is this entry's largest gap. Anyone with the physical volume should check three things: whether the Preface dates the reprinted parts, what the paragraphs around p. 38 (1960) / pp. 95–96 (1965) actually say, and the comparably / comparatively variant in the nervous-system sentence.

The quotation's provenance. Quote Investigator, quoteinvestigator.com/2020/11/11/ai-can-do/, read via fetch. It is the source for the 1960 attribution and page 38, the April 1960 lecture date, the 1965 pp. 95–96 reprint location and chapter, the $10,000 rental and ten-middle-managers passage at p. 33, the four occupation categories, and the Dreyfus 1967 and Kurzweil 1985 citations. QI states it verified the wording; I could not check it against a scan.

The book's structure. The Internet Archive catalogue record for shapeofautomatio00simo, for the table of contents, publisher, pagination and the note that the volume revises the 1960 edition. WorldCat and Open Library for corroboration of the 1965 Harper & Row imprint.

The 1960 essay in Part II. Melvin Anshen and George Leland Bach (eds.), Management and Corporations, 1985 (McGraw-Hill, 1960) — Simon's contribution "The Corporation: Will It Be Managed by Machines?" at pp. 17–55, from a symposium for the tenth anniversary of Carnegie Tech's GSIA. Bibliographic details from search results and the Internet Archive record for managementcorpor0000ansh; the essay itself was not read.

The NYRB exchange. Robert L. Heilbroner, "Where Do We Go From Here?", The New York Review of Books, 17 March 1966 (nybooks.com/articles/1966/03/17/where-do-we-go-from-here/), and Herbert A. Simon, "Automation", NYRB, 26 May 1966 (nybooks.com/articles/1966/05/26/automation-3/). Both read through a fetch-and-summarise tool rather than by eye; the sentences quoted above are reported by that tool as verbatim and I could not re-verify them character by character. Robert Theobald's reply in the same issue was not read. Daniel Bell's "The Bogey of Automation" (NYRB, 26 August 1965) was read and does not mention Simon, despite a CMU archival record (digitalcollections.library.cmu.edu/node/48889, which would not load — TLS error) describing a review of the book by "Bell and Heilbroner." That record suggests a Bell review I did not find, and it is a live lead.

Dreyfus. Alchemy and Artificial Intelligence, RAND P-3244, December 1965 — Allen Newell's own copy is scanned at iiif.library.cmu.edu/file/Newell_box00093_fld06448_doc0002/, and no text could be extracted from it here. "Why Computers Must Have Bodies in Order to Be Intelligent," The Review of Metaphysics 21(1), 1967, p. 15 — not read; the Simon quotation it contains reaches this file via Quote Investigator. The reception, Simon's "that garbage" reaction, Papert's "gossip," the lunch remark, Crevier's assessment and Dreyfus's 2007 "I figure I won" are from Wikipedia's Hubert Dreyfus's views on artificial intelligence, read via fetch; none is traced to a primary here and the Simon quotation in particular deserves one before a reading leans on it.

Simon's later statements. The October 2000 Earthware Symposium talk at Carnegie Mellon, transcript and video at the Simon Institute for Longterm Governance (simoninstitute.ch/blog/post/how-computers-will-continue-to-shape-the-world-herbert-a-simon-video-transcript), read via fetch — the source for "a relatively short time," "our task is not really to predict it," and the spectators question. Simon died 9 February 2001. The chess-prediction retrospectives are from the Chessprogramming wiki's Herbert Simon page (chessprogramming.org/Herbert_Simon), which quotes George W. Baylor (1997–98) and Hans Berliner (2001). Berliner's "galvanized the field" is a recollection of undated private conversation and is the weakest evidence used in this file. Simon and Schaeffer's "The Game of Chess" was retrieved and could not be read (PDF text extraction failed).

The economics. Richard N. Langlois, "Cognitive Comparative Advantage and the Organization of Work: Lessons from Herbert Simon's Vision of the Future," University of Connecticut working paper 2002-20, published in Journal of Economic Psychology 24(2), 2003, pp. 167–187. Abstract read in full via RePEc; the paper itself could not be read — the UConn PDF downloaded but no text extractor was available in this environment, and SSRN returned 403. The abstract is the source for the "hold up remarkably well" grading and the cognitive-comparative-advantage framing. This is the most valuable unread source in the file, because it is the only scholarly treatment found that grades the 1960 economic predictions systematically.

The automation panic and its official answer. The Triple Revolution memorandum, 22 March 1964, via Wikipedia. The National Commission on Technology, Automation, and Economic Progress (PL 88-444; chair Howard R. Bowen), Technology and the American Economy, February 1966 — the report was not read; the "technology eliminates jobs, not work" formulation and the commission's recommendations come from search summaries of ERIC records and secondary accounts. Steve Lagerfeld, "Lessons of the First Automation Crisis," The American Interest, 15 February 2020, read via fetch, for the sixteen-of-thirty-six industries figure, the 3.4% wartime unemployment figure, and as the documented instance of the misuse graded in §6.

The employment record. BLS The Employment Situation, June 1985, via FRASER (St. Louis Fed) for 7.2% unemployment and 97.5 million nonfarm wage and salary jobs; the c. 54 million figure for 1960 is the standard payroll series and is stated here from general knowledge rather than from a retrieved table, so treat the doubling as approximate and directional.

Grading and base rates. Ray Kurzweil, "What Is Artificial Intelligence Anyway?", American Scientist 73(3), May–June 1985, pp. 258–264 — pagination confirmed via ERIC EJ321513; the article was not read, and the p. 261 location of the Simon citation comes from Quote Investigator. Melanie Mitchell, "Why AI is Harder Than We Think" (arXiv:2104.12871), read via ar5iv, for the 1960 dating and the four fallacies. AI Impacts' MIRI AI Predictions Dataset pages for the pre-1980 clustering and the 1.6% figure, via search summaries rather than in full. Erik Brynjolfsson and Andrew McAfee, The Second Machine Age (2014), p. 141, via Quote Investigator; not read.

Reached and not used. AIWS.net's history entry on the quotation gets the book right and the date only as "the 60s," and contains unrelated errors. Luke Muehlhauser's "What should we learn from past AI forecasts?" returned HTTP 403 at coefficientgiving.org, exactly as minsky-1970 reported; it is still the obvious next source for anyone revising either file. Richard Bellman's published response to Simon and Newell in Operations Research was located by title only — the Semantic Scholar record would not render and the INFORMS page returned 403 — so §6 names its existence and quotes nothing from it.

8. On the boundary

This file cites logic-theorist-1956, dartmouth-1956, minsky-1970, kurzweil-2005, amodei-2024-loving-grace, deep-blue-1997, turing-1950, wiener-1960-automation, moravecs-paradox-1988, lighthill-1973, expert-systems-collapse-1987, frey-osborne-2013, alexnet-2012, transformer-2017, scaling-laws-2020 and her-2013 as related entries already in canon/. It invents no ids. It quotes LENSES.md once, in §4.3, to name the rule this forecast is an ancestor of; it reads that file and does not write it.

It refers to the project's own reading of 17 August 2026 three times, in §4.2 — the nil return on the Work-and-economy lens, the Gravis Robotics item, and the 3M expert-witness item — to name a live citation occasion. Nothing is taken from it as evidence, and the reading is cited as a thing this project wrote, not as a source about the world.

It corrects the date used by five entries already written. It does not edit them; dartmouth-1956, minsky-1970, musk-robotaxi-2019, dijkstra-1959 and amodei-2024-loving-grace say "Simon in 1965" and will go on saying it until someone revising them reaches this file, which is the correct order of operations in a record that is not supposed to be rewritten behind the reader's back.

It deposits nothing in the evidence ledger, grades no model and no vendor, places no needle and no landmark. The judgements it does make are the grades in §5, and they are stated as judgement: that the technological claim failed at both due dates and remains open; that the ordering claim was correct; that the economic claim substantially held in aggregate and that the aggregate does not settle Heilbroner's objection. Where an independent grader and the forecaster differ, §5.5 says so and says which is which, and where the evidence is a recollection rather than a document, it says that too.