← The canon · AItopiaOrAImageddon?

"Artificial Intelligence: A General Survey"

moment · Sir James Lighthill · 1973

Something that happened and changed what people expected next.

Descends from The Dartmouth Summer Research Project on Artificial Intelligence, The Logic Theory Machine. Read on: Moravec's paradox.

Filed as moment, and moment is right — but only on the house definition, and the definition has to be applied carefully, because five entries already in canon/ have been citing this file before it existed and four of them use the same phrase to do it.

The house definition, set by eliza-1966 when it argued its own way out of the kind and used since by clippy-1996, siri-2011 and expert-systems-collapse-1987, is a date on which something visibly happened in public. Those four entries all reach for the identical list — "Dartmouth, Lighthill, Deep Blue, the LLaMA weights" — and dartmouth-1956, when it was written, noted that the canon had been leaning on that list for a day and a half. This is the one entry in the list that had not been written. logic-theorist-1956 calls it "the case where a funder believed the argument"; rosenblatt-perceptron-1958 calls it "the case where a funder believed one"; perceptrons-1969 spends a page of its misuse section saying the first winter is Lighthill's event and not its own, and closes that section with the instruction that if the chain writes one more entry, this is the one. So the kind is settled less by taxonomy than by use: every existing citation treats this as the moment a funder acted, and that is what a moment is.

Two things about the label need saying before anything else, because both are load-bearing.

First, the date in the id is the publication year and the document is older than that. Part I is signed "July 1972." Lighthill submitted it to the Science Research Council on 23 March 1972. Donald Michie had read a leaked March 1972 draft before it was published. It appeared in print in 1973, in a volume called Artificial Intelligence: a Paper Symposium, and it was in 1973 that it became public, was answered in print, was argued on television and started moving money. 1973 is the right year for a moment because a moment is when a thing became visible; a reading quoting the text should date the claims to 1972, which is also how Lighthill dates them internally ("high hopes that are very far from having been realised in 1972").

Second, the tempting re-filing is prediction, and I did not do it — but I have graded it as one anyway. Section 4 of the report is an explicit twenty-five-year forecast with a self-declared horizon: Lighthill says outright that "any speculations beyond the end of this century are regarded as too uncertain to justify mentioning them," which sets a due date of roughly 1997 in his own words. Better still, and rarer than anything else in this canon, the predictor graded himself twice, on the record, and the grades survive — once in private correspondence in 1982 and once in an interview in 1986, where he said the development of AI had been "very much as I forecast." The canon's rule for predictions is that a self-grade and an independent grade must both be recorded when they differ, and here they differ sharply. But re-filing this as prediction would break the five entries that need it as an event, and the forecast is one section of a document whose effect came from the other three. So: moment, with section 3 below carrying the full prediction discipline — claim, date made, date due, what happened, his grade and an outside one. The kind describes what the entry is for; it does not license losing the gradeable content.

On descends_from. Two ids, both in canon/, both genuine ancestors rather than neighbours. dartmouth-1956 supplies the clock and the promises: Lighthill's whole rhetorical structure is a twenty-five-year audit — "Workers entered the field around 1950, and even around 1960, with high hopes" — and the inflated predictions he grades are the ones descended from that proposal's eight-week claim. logic-theorist-1956 supplies the target: the combinatorial-explosion argument is aimed precisely at the general-problem-solving and theorem-proving line that began there, and that entry already anticipates this file by name.

Three more I would want and will not invent. turing-1950 is unwritten, and matters because Lighthill dates the field's start not to Dartmouth but to "Turing's 1947 article Intelligent Machinery" — a different origin claim from the one dartmouth-1956 examines, and worth noting that the report's own genealogy disagrees with the canon's. shannon-chess-1950 is written and is not an ancestor but a sibling: Lighthill devotes two pages to grading computer chess as evidence, and that entry already points here. perceptrons-1969 is written and is emphatically not an ancestor. Lighthill never cites it, and the two documents attack different halves of the field — Minsky and Papert's book bore on connectionism, this report bore on symbolic robotics and general problem solving. They are the two things routinely merged into one cause of one winter, and perceptrons-1969 already does the work of separating them. Where the two files touch, they agree.

What it is

A seventy-odd-page survey of the state of artificial intelligence, written by a man who had never worked in it, at the request of the body that funded it, and published in 1973 by the UK Science Research Council alongside four replies from people who had. Its author was Sir James Lighthill, one of the outstanding applied mathematicians of his generation: a fluid dynamicist, at Trinity College, Cambridge on a scholarship at fifteen alongside Freeman Dyson, wartime supersonic aerodynamics at the National Physical Laboratory, director of the Royal Aircraft Establishment at Farnborough from 1959 where his work on supersonic wing shape fed into Concorde, founder of the field of aeroacoustics, and from 1969 Lucasian Professor of Mathematics at Cambridge — Dirac's chair before him, Hawking's after. He knew nothing about AI and said so in the third paragraph.

How it was commissioned, which is the part that has only recently been documented. In September 1971 Brian Flowers, chairman of the SRC, wrote to Lighthill. Jon Agar found the letter in Lighthill's papers at UCL, and it does not read like a request for an open verdict:

> These subjects, and artificial intelligence itself, are highly complex, very > much the preserve of experts and perhaps sometimes of plausible charlatans. > There is a strong tendency for the experts to pursue their particular > enthusiasms energetically but narrowly and to ignore, or repudiate any other > approach … It is getting increasingly difficult for the SRC to control this mix > of activities and to make properly informed judgments…

Flowers then asked him "to make a considered appreciation of the subject of artificial intelligence, its achievements, its practitioners, its promise and its needs." Agar's phrase for the first paragraph is "this remarkable steer," and his conclusion is that "the Science Research Council did have narrow concerns about the variety and quality of artificial intelligence research, and … certainly gave Lighthill a steer with respect to the target of the review." The word charlatans is in the commissioning document. It is not in the report.

The report describes its own basis with some care, and the description is damning if you read it plainly. Lighthill writes that the SRC's chairman

> decided to commission an independent report by someone outside the AI field but > with substantial general experience of research work in multidisciplinary > fields … I undertook to make such an independent report, on the understanding > that it would simply describe how AI appears to a lay person after two months > spent looking through the literature of the subject and discussing it orally and > by letter with a variety of workers in the field

and closes the section by insisting the opinions "represent merely the broad overall view of the subject which he reached after such limited studies as he was able to make in the course of two months." He declined to write a summary. He then lists some fifty names of people whose work informed him — Minsky, McCarthy, Michie, Nilsson, Winograd, Sutherland, Longuet-Higgins, Needham, Marr, Brenner, Broadbent, Hubel, Kilburn, Strachey, Good, Gregory and others.

The expenses claim survives. Lighthill worked thirty-nine days at £12.60 a day, and charged the Science Research Council £491.40. That is the whole cost of the document that is said to have frozen a national research field.

The ABC. The report's organising device divides AI into three categories by long-term motivation, and everything else follows from it.

The argument is then structurally simple and, as an argument, rather elegant. A and C have "perfectly clear" and quite distinct aims, and if they were the whole field they would "warrant completely separate treatment in respect of research support, departmental organisation, etc." The only thing making AI a single coherent discipline is the claim that B bridges them. So if B fails, AI is not a field; it is two fields with a shared name and a mistaken sense of unity. The report is not primarily an attack on AI's results. It is an attack on AI's existence as a subject.

The verdict. Section 3 is titled "Past disappointments" and contains the sentence everybody quotes:

> Most workers in AI research and in related fields confess to a pronounced > feeling of disappointment in what has been achieved in the past twenty-five > years. Workers entered the field around 1950, and even around 1960, with high > hopes that are very far from having been realised in 1972. In no part of the > field have the discoveries made so far produced the major impact that was then > promised.

He then names one cause, "as a mathematician," and it is the argument the report is remembered for:

> failure to recognise the implications of the combinatorial explosion. This > is a general obstacle to the construction of a self-organising system on a large > knowledge base which results from the explosive growth of any combinatorial > expression, representing numbers of possible ways of grouping elements of the > knowledge base according to particular rules, as the base's size increases.

The consequence he draws is sharper than the premise and is the report's real thesis: because search cannot be brute-forced, programs must use heuristics; heuristics encode human knowledge of the domain; therefore "it depends critically on data derived through the use of human intelligence." The machine contributes speed, not intelligence. He puts this most flatly about chess — "the intelligence contribution is human; what the computer offers is its speed, reliability and biddability" — and generalises it into the finding that carries the whole report:

> Advanced Automation techniques are successful not when they are developed with a > high degree of generality of application, but only when a large quantity of > detailed knowledge about the problem domain is utilised in the program design.

Category B is then dismissed at length. Eye–hand co-ordination is judged "not of human standard but similar to what an octopus can learn" and its engineering "repellingly formidable"; the chess programs play at "experienced amateur standard characteristic of county club players in England," and "chess masters beat them easily"; "learning programs are not considered applicable to computer chess at present"; "a general-purpose program, the coveted long-term goal of AI activity, seems as remote as ever." A postscript concedes real admiration for Terry Winograd's SHRDLU thesis — Lighthill calls it "an exceptionally good PhD thesis" and praises its author's prose against a literature he finds "depressingly turgid or jargon-dominated" — and then declines to update on it: "one swallow does not make a summer." The section ends: "the attempt to construct a true Bridge between categories A and C is not succeeding."

One passage deserves quoting because it is the strangest thing in the document and it is not an aside. Asking why Building Robots is popular, Lighthill offers the "long-standing captivation of the human imagination by the very concept, as shown by its continual prominence in literature, from medieval fantasies of the Homunculus through Mary Shelley's Frankenstein to modern science fiction," and then this:

> Incidentally, it has sometimes been argued that part of the stimulus to > laborious male activity in creative fields of work, including pure science, is > the urge to compensate for lack of the female capability of giving birth to > children. If this were true, then Building Robots might indeed be seen as the > ideal compensation!

He develops it — most robots inhabit "a world as like as possible to the conventional child's world as seen by a man," playing games, doing puzzles, stacking bricks — and concludes that the relationship between a Robot and its Builder is "pseudo-maternal rather than Pygmalion-like." Sutherland's reply answers it in one line: "a scientist's unconscious motives for undertaking the work he does have nothing whatever to do with the assessment of the scientific value of that work." That the sentence survived into a funding document is the single best piece of evidence for what kind of document it was.

The four replies, published alongside it, which is why the volume is called a symposium.

The irony Agar found in the archives, which is worth more than most of the argument. In 1964 Michie had himself surveyed UK computing research for the SRC — the study was published by the Council in 1965 — and had drawn the field as a dumbbell: two bells, Advanced Computer Science and Automation at one end, Computer-based CNS Research at the other, joined by a connecting piece labelled "or bridge" and marked with a question mark. Lighthill's A, B and C is Michie's own diagram with the question mark answered in the negative. "The two pictures of the field," Agar writes, "Michie's in 1964 and Lighthill's in 1972 are so similar that Lighthill must have, directly or indirectly, drawn his map of artificial intelligence after Michie's own visualisation — a remarkable irony."

The debate. The argument moved to television, filmed at the Royal Institution in London and broadcast by the BBC as part of the Controversy series in 1973: Lighthill against Michie, John McCarthy and the perceptual psychologist Richard Gregory, on the proposition that the general purpose robot is a mirage. The recording survives and is hosted by Edinburgh's Artificial Intelligence Applications Institute. The date is not settled in the sources I could reach: Edinburgh's own page and the Chilton archive both say June 1973, while several secondary accounts and a published transcript give 9 May 1973. The likeliest reconciliation is a recording date and a broadcast date, but I could not confirm that and record the discrepancy rather than pick one. McCarthy's written review followed in Artificial Intelligence in 1974, and his objection was that the scheme has no room for the field's actual subject — the study of "the structure of information and the structure of problem solving processes independently of applications and independently of its realization in animals or humans." His summary judgement: "If we take the [ABC categorisation] seriously, then most AI researchers lose intellectual contact with Lighthill immediately."

What happened next. The SRC reorganised funding along the report's lines, supporting A and C separately and, in Agar's phrase, "leaving B in the lurch." A letter-writing campaign by Michie's international network — Charles Rosen of SRI in June 1973, Nils Nilsson and ten others from SRI, Stanford, MIT and USC in May 1973 — did not change the Council's mind. The blow landed hardest exactly where the commissioning letter had pointed. Edinburgh was already fractured: Gregory had left for Bristol in 1970, Longuet-Higgins and Michie had fallen out, and other departments resented the unit's freedom from teaching. In 1974 a new Department of Artificial Intelligence was established at Edinburgh without Michie, who was left with a small independent Machine Intelligence Research Unit. British AI did not get a major research-council programme again until the Alvey programme's Intelligent Knowledge-Based Systems strand, following a September 1982 review — about a decade.

Why a reading would cite it

The occasion is one this project will meet repeatedly and has a name for: a funder, a government or a board reads the state of a field and withdraws on the strength of a judgement, and the judgement is dressed as a survey. This is the canon's worked case, and it is worked to an unusual depth because the archive is open: we have the commissioning letter, the expenses claim, the leaked draft, the replies, the television debate, the funding consequence, and the author's own grade of himself a decade later. Almost nothing else in the canon lets a reading check all of that against a claim.

Three specific tests it supplies, in order of how often they will be needed.

1. When an assessment is described as independent, ask what the commissioning letter said. Flowers wrote "plausible charlatans" before a word of the report existed, asked an outsider with a documented record of hostility to curiosity-driven programmes, and got back a document that reached the conclusion the letter gestured at. None of that makes the report dishonest, and Agar is explicit that Lighthill's position was a principled one held across his whole career rather than a commission fulfilled. But "independent" is a claim about process, and the process is checkable. When a reading meets an evaluation, a safety audit, a model card, a benchmark result or a government review presented as disinterested, the question is not whether the author was honest. It is who chose the author, what they asked for, and what they already believed.

2. When someone with no standing in a field is given authority over it, notice what that buys and what it costs. The report's outsider status was the point — Flowers wanted someone the incumbents could not capture, and there is a real argument that a field deep in its own enthusiasm cannot audit itself. What it bought was a document that said the unsayable and could not be dismissed as factional. What it cost is visible in Michie's first question and in McCarthy's review: thirty-nine days, no interviews with the field's leaders, a taxonomy lifted unwittingly from the person it was used against, and a category scheme that practitioners found unrecognisable. Both halves are true at once, and a reading that wants only one of them is doing propaganda. 2026 produces this shape constantly — outside evaluators, non-technical legislators, courts, and generalist commissions ruling on systems whose builders insist they cannot be understood from outside.

3. When a pessimistic verdict is graded years later, grade the whole of it and not the quotable part. This is the use I expect to matter most, and section 3 does it in full. The one-line version of this report — arrogant outsider says AI is hopeless, is refuted by history — is wrong in a specific and checkable way: some of his forecasts were right, including one that is almost never credited to him, and the ones that were wrong were wrong for a reason that is still live.

Honest note on the occasion. I read the three digests on disk (2026-08-14, 2026-08-15-12, 2026-08-16-00) looking for a live instance and did not find the report's own shape. Nobody commissioned a survey to settle a field this week. The 16 August window is dominated by AI financing structure — Nvidia cutting its planned backstop for the OpenAI Ohio project from up to $250bn to under $120bn, Bloomberg pricing roughly $70bn of off-balance-sheet credit support, an 80.3% concentration in Nvidia's disclosed equity book. That is a funder reducing exposure, which is the same family as this entry but not the same event: a private guarantor trimming a commitment on its own initiative is not a public body withdrawing from a research programme on an outside judgement. This is a standing citation occasion, not a this-week one, and I would rather say so than dress a generic claim as a current one. The nearest genuinely live shapes I found are outside the digests and I did not read the underlying material: reported pressure on US civilian research funding against rising defence spending, and the fact that the SRC's own institutional descendant now runs a £1.6bn UK AI strategy — an irony available to a reading that wants it, but a search-result summary and not a finding.

What it got right, and what it got wrong

Not required for moment. Required here, because the report carries a dated forecast with a self-declared due date and an author who graded himself, and because the received one-line verdict on it is lazy in both directions.

The claim, in its own words. Section 4, "The next twenty-five years," forecasts fission: that research areas within A and C will make substantial progress by forging "far stronger links to the immediate field of application than to the supposed bridge activity B"; that continued failure in B "may cause progressive loss of prestige, from which a diminution of funding will ultimately follow"; and that "in due course the overriding significance of the links between each research area and its field of application will rupture the always fragile unity of the general concept of AI research."

Date made: July 1972 (submitted March 1972; published 1973). Date due: by his own stipulation, the end of the twentieth century — "any speculations beyond the end of this century are regarded as too uncertain to justify mentioning them."

Right, and almost never credited: the twentieth-century call

> Students of all this work have generally concluded that it is unrealistic to > expect highly generalised systems that can handle a large knowledge base > effectively in a learning or self-organising mode to be developed in the 20th > century.

Made 1972. Due 2000. Outcome: correct. No such system existed in 2000. This is the sharpest, most falsifiable sentence in the report, it was made against a correspondent who had told him machine intelligence might exceed human intelligence by that same year 2000, and it is right. A reading that mocks Lighthill without noticing that his one crisply dated prediction landed is repeating the thing this canon exists to prevent.

Right: fission actually happened, and held for about his stated span

For roughly the quarter-century he was forecasting, the field did fragment exactly as described. "AI" became an unfashionable word; the parts ran under their own names — machine learning, computer vision, natural language processing, robotics, information retrieval — each bound tightly to its own application and its own conference, each doing better for the separation. The general-purpose programme had no institutional home. If you had graded this forecast in 1997, at the end of his own window, you would have scored it a hit.

Right: the combinatorial explosion was real, and microworlds were a dead end

Nothing has refuted the mathematics. Search spaces are still exponential. And his specific judgement on category B as it then existed — blocks worlds, table-top universes of discourse, SHRDLU-style microworlds — is the consensus view now. He said the methods would not extend past "a very limited material" and would be "opposed violently by the combinatorial explosion" on a wider universe of discourse, and the microworld programme did in fact collapse for that reason. His "one swallow does not make a summer" on Winograd was, on the narrow question of whether SHRDLU's approach would scale, correct.

Right for about fifteen years, then exactly backwards: knowledge beats generality

> Advanced Automation techniques are successful not when they are developed with a > high degree of generality of application, but only when a large quantity of > detailed knowledge about the problem domain is utilised in the program design.

This was the best-supported empirical generalisation in the report, and the decade that followed vindicated it hard — the expert-systems boom is this sentence turned into an industry, and this canon carries both ends of it in mycin-1976 and expert-systems-collapse-1987. Then it inverted. The claim is the precise negation of what is now the field's central methodological finding: that general methods which scale with computation beat methods built on hand-supplied domain knowledge. Lighthill did not merely guess wrong about which would win; he named the mechanism that would decide it and got the sign backwards. He wrote that theorem proving was "particularly an area where hopes have been disappointed through the power of the combinatorial explosion in rapidly cancelling out any advantages from increase in computer power." Increase in computer power is what won. See alexnet-2012 for the demonstration, transformer-2017 for the architecture and scaling-laws-2020 for the point at which the field began treating it as a measurable regularity.

Wrong: the central thesis, and the whole basis of the entry

The bridge got built, and it is the entire subject in 2026. General-purpose systems that mimic a wide range of human functions — the literal definition of category B — are what the field now is, what the money is in, and what this project takes readings of twice a day. "Such a general-purpose program, the coveted long-term goal of AI activity, seems as remote as ever" (1972) is the sentence the record has most comprehensively destroyed. So is the fission forecast, past his window: the parts he expected to separate permanently re-merged after about 2012 into a single stack of methods, and the unity he said was a mirage is now the ordinary condition of the field.

Note carefully when it failed. Through 1997 — his own horizon — he was defensible on nearly all of it. The refutation came from the two decades after the due date. An honest grade is therefore: right on his stated timescale, wrong about the shape of the world beyond it, and wrong in a way that his own argument ruled out rather than merely failed to anticipate.

Wrong, in detail, on his own horizon

The self-grade and the independent grade differ, and both are on the record

His own grade is high. In November 1982, writing privately to John Kingman at the SERC — and marking the letter PERSONAL & CONFIDENTIAL because he intended to name people — Lighthill said progress had come precisely from abandoning what he called the "nineteen-sixties view that progress in AI would come out of viewing it as a seamless robe stretching all the way from advanced automation to neurobiology," a view he called "just wrong-headed." He claimed direct credit for the outcome at Edinburgh: "the change of direction and of organisational structure that followed my Report greatly helped AI work at Edinburgh to flourish," adding that the work there had been "seriously held back" by Michie's approach and that "his personal characteristics tended to foment additional, very serious difficulties." In 1986 he put it in one sentence to Machine Intelligence News: "The development of AI has been very much as I forecast."

Independent grades ran lower even before the record turned against him. The 1982 SERC review by Needham — who had supported him in 1973 — and Peter Swinnerton-Dyer concluded: "It must be recognised that the Lighthill Report was right for its day, but that its day is past. AI is a different subject from what it was ten years ago, and it is set in a different context. Then it was the wild blue yonder. Now it is the open frontier of IT." James Fleck's sociological account (1982) reads the report as conservative — an "affirmation of the status quo" that "affirmed the value of currently existing areas" with strong links to established disciplines, operating a prestige hierarchy. Agar (2020), the most thorough treatment, accepts the causal claim that most concerns a reading — "Lighthill's report did indeed cause AI research to stall — the criticisms resonated beyond the UK's shores" — while arguing that the motivation has been misread, and that it followed from a career-long principle rather than hostility to AI as such.

The gap between the grades is the whole point of recording both, and it is narrow and specific. Lighthill and his critics broadly agree on the 1970s facts. They disagree about category B, and only about category B: he graded its abandonment a success, and the field it later became is the reason anybody reads his report at all. His 1986 self-assessment was made twenty-six years before alexnet-2012. A reading that quotes "very much as I forecast" without the date attached is quoting a man who could not yet have known.

One more grade belongs in the record because of how it aged. Lighthill died in 1998, swimming around the island of Sark, a nine-mile circuit he had completed several times before. His Guardian obituary said that he had "infuriated the computing lobby by advising against investment in artificial intelligence because he believed the concept difficult, costly, and probably ill-founded," and added: "Many now believe him to have been right." That sentence was published in July 1998. Agar notes drily that the obituary "is riddled with errors, not least consistently calling Lighthill 'Lightman'."

Commonly misused as

Not required for moment, and the largest section here anyway, because this is a document almost nobody quoting it has read and the errors are load-bearing.

"It called for a halt to AI research in Britain"

The standard version, from Daniel Crevier's history: "his 1973 report called for a virtual halt to all AI research in Britain." It is false on the face of the document, twice over.

First, Lighthill supported categories A and C explicitly and at length, expected "substantial advances" in C over the following twenty-five years, and wanted A funded closer to industry. His target was one third of a taxonomy of his own construction. Second — and this is the part that surprises people — the report makes no funding recommendations at all. He says so himself: "this report's contribution consists not of any detailed costed recommendations, but of certain general considerations." There is no sum of money in it, no programme named for cancellation, no list of grants. It is an argument that a field is not a field. The SRC drew the funding conclusions; the report handed them a reason.

That distinction is the reusable part. A document that recommends nothing can still end a programme, because what it withdraws is not money but the standing to ask for it. When a reading meets an evaluation or a review that carefully makes no recommendation, this entry is the reason not to treat that restraint as neutrality.

"An ignorant outsider wrecked the field by accident"

The most popular version and the least useful, because it treats the report as an error rather than a position. Three corrections.

He was an outsider by design, not by oversight — Flowers wanted one, and said why. His hostility was not to AI specifically: he had form. In 1966 he sat on a working party reviewing UK fusion research at Culham, judged the majority report "half-hearted," broke ranks and filed a minority report with F. E. Jones arguing that no work on fusion power was justifiable, and proposed redirecting Culham's 120 scientific and 130 other staff to raising the efficiency of British industry. A man who had already tried to close down British fusion research was not an unlikely choice for this commission; he was a predictable one. And the ABC taxonomy that practitioners found so alien was, as above, Michie's own 1964 diagram.

What survives of the "ignorance" charge is narrower and stronger than the popular version: thirty-nine days, no interviews with the field's international leaders, and fifty names in the acknowledgements who mostly had not been asked. That is Michie's charge, it is specific, and it is the one that stands.

"It was an independent report"

The word appears twice in Lighthill's own first section and it was true in the sense he meant — nobody dictated his conclusions, and Agar's reading is that they followed from convictions he had held for decades. But the framing was set before he began. The commissioning letter contained "plausible charlatans" and a complaint that the SRC could not "control this mix of activities." A small group was convened afterwards to compose an "office view" that would "introduce and steer discussion of the report" when it reached the Council. And the format of the published symposium, which looks generous, was in Agar's word invidious: it gave the targets a platform on which to attack each other, and they took it — Sutherland used his reply to criticise Edinburgh's teaching, Longuet-Higgins distanced his own work from the category under attack by relocating it to the safe one, and Needham sided with Lighthill outright. Four replies from four institutions, disagreeing with each other in print, is not a united field; it is what a steered process produces, and the Council read all of it.

"It caused the AI winter"

Half right, and the half that is wrong is a geography error the canon should be strict about. This report is the British cause, and the funding consequence in Britain is documented and roughly decade-long. The American first winter has its own causes, running on a different clock and hitting through different mechanisms: the Mansfield Amendment, ARPA's retreat from open-ended funding from 1974, the cancellation of the speech-understanding programme in 1976. perceptrons-1969 sets those out in detail and also records that some historians dispute whether a first winter is visible in the funding record at all. Agar accepts that Lighthill's criticisms "resonated beyond the UK's shores," so the international effect is real — but resonated is not caused, and a reading that makes one man's £491 report the cause of a decade on two continents has substituted a citable event for a diffuse process. That is the exact error perceptrons-1969 is about, run on a different document.

"The combinatorial explosion argument was refuted"

It was not. It was routed around, which is a different thing and the distinction is live. Nobody has shown that exhaustive search over a large knowledge base is tractable; the exponentials are where he left them. What happened is that the field stopped making search the primary mechanism and made learned function approximation the primary mechanism, so that the relevant question became how well a model generalises rather than how much of a space it can enumerate. Lighthill's error was not the mathematics. It was the inference from it: that because search must be pruned, the pruning must encode human domain knowledge, and therefore "the intelligence contribution is human." The premise held and the conclusion failed, because a third option existed that he did not consider — that the pruning could be learned from data rather than supplied by an expert. When a reading meets a 2026 impossibility argument built on a complexity result, this is the template: the theorem is probably fine, and the load-bearing step is the sentence after it that says what therefore must be true of any system that beats it.

The inverse misuse: "Lighthill was right all along"

The over-correction, and it has a documented peak — the 1998 obituary's "many now believe him to have been right," published fourteen years before alexnet-2012. The claim recurs whenever capability progress stalls or a funding cycle turns, and it will recur in this project's own subject matter. The correct response is not to deny that he got things right, because he did, and section 3 lists them. It is to insist on the dates. He was right about the 1970s, largely right through the 1990s, and wrong about the thing that decides whether the entry matters. A verdict that was accurate for twenty-five years and then inverted is the most dangerous kind, because for twenty-five years it accumulated evidence.

The template, which is the reason to keep the file

Underneath all of these is one reusable shape, and it is the sibling of the one perceptrons-1969 names. That entry is about crediting a limit result with the death of the approach it limited. This one is about the step before: an assessment commissioned by a funder, from an outsider, framed by the funder's prior, arriving at a moment when the field's internal disputes make it easy to believe — and then doing its work not by recommending anything but by removing the standing to ask.

The three questions a reading should actually take from this file, in order:

1. Who commissioned it, and what did they say when they did? The prior is usually in writing and is usually retrievable. 2. What does the document actually recommend? If the answer is nothing, ask what it withdrew instead. 3. What was the state of the field's internal quarrels when it landed? In 1972 Edinburgh had already lost Gregory to Bristol, Longuet-Higgins and Michie had fallen out, and neighbouring departments resented the unit. An outside verdict arriving on top of an existing fracture will be credited with the break.

The last word should go to the man the report cost most, because he identified the category error better than anyone and did it in the author's own field. Calling the bridge between AI's applications and AI's science "Building Robots," Michie wrote, is like calling the bridge between aero-engineering and the study of bird flight "Building Wind-tunnels":

> It is surely more fruitful, if one seeks inter-disciplinary connections, to > choose a common body of theory rather than to seize on a piece of laboratory > equipment. One feels that Sir James would be among the first to agree that to > speak of Building Wind-tunnels as the Bridge between aero-engineering and the > study of bird flight, would direct attention away from the true bridge, namely > the science of aerodynamics. The equivalent science in the case of AI is at a > primitive stage. It is the hope of every AI professional to contribute in some > way to bringing the required theory into being.

Fifty-three years later the equivalent science is still at a primitive stage, the machines work anyway, and that combination is the standing condition this project is reporting on.

Sources

Read directly, in full.

Read, but through a summarising fetch rather than in full.

Consulted and not relied on.

On the boundary. This file cites the canon's own dartmouth-1956, logic-theorist-1956, perceptrons-1969, rosenblatt-perceptron-1958, shannon-chess-1950, eliza-1966, clippy-1996, siri-2011, mycin-1976, expert-systems-collapse-1987, deep-blue-1997, alexnet-2012, transformer-2017 and scaling-laws-2020 as related entries, and names turing-1950 as proposed-but-unwritten. It refers to the project's own digests of 14–16 August 2026 only to record, in section 2, that none of them contains an instance of this entry's occasion, and to describe that window's financing items as a related but distinct shape. It deposits nothing in the evidence ledger, grades no vendor, places no needle and no landmark, and makes no claim about the present state of AI beyond what entries already in canon/ establish.