"Computer Power and Human Reason: From Judgment to Calculation"

interpretation is right and proposals.md filed it correctly. The test wiener-1960-automation set is a text arguing a position against named opponents, and the external evidence for the kind is that it drew a published refutation — you cannot refute a moment, and nobody rebuts an idea's existence.

Here that evidence is stronger than in the Wiener case, because this book drew three refutations in one document. Stanford Artificial Intelligence Laboratory Memo AIM-291 (Computer Science Department Report STAN-CS-76-577, November 1976) is titled "Three Reviews of J. Weizenbaum's Computer Power and Human Reason," and collects Bruce G. Buchanan, Joshua Lederberg and John McCarthy, with Weizenbaum's reply printed alongside. It was issued by the laboratory whose director is one of the book's named targets. A field does not convene its own institutional press to answer a fact; it answers a claim it thinks is false.

Why not idea. The strong rival, and it fails on the test samuel-checkers-1959 set and wiener-1960-automation applied: an idea is something you can implement, because there is a loop in it. There is no loop here, no formalism, no criterion, no decision procedure. What is transmitted forward is an argument and two sentences of it.

Why not limit, and this is the one worth saying loudly. Filing this entry as limit would commit, in the header, the precise error section 4 exists to correct. Weizenbaum's thesis is stated in the preface as two claims, the second of which is that "there are certain tasks which computers ought not be made to do, **independent of whether computers can be made to do them**" (p. x). He goes further in the Introduction, distancing himself by name from the argument people now confuse his with: Hubert Dreyfus "limits his argument to the technical question of what computers can and cannot do. I would argue that if computers could imitate man in every respect — which in fact they cannot — even then it would be appropriate, nay, urgent, to examine the computer in the light of man's perennial need to find his place in the world" (p. 12). This is a book about jurisdiction, not about capability. godel-incompleteness-1931 and turing-halting-1936 are where proofs live; this file has none and claims none.

Why not moment. Nothing happened on publication day, on the house definition eliza-1966 set and lighthill-1973 confirmed — no funder acted, no institution moved, no program ran. The argument accreted over the following year in a memo, an ACM controversy and a decade of lectures.

Why not prediction. The book contains forecasts, and three of them are graded in section 3 under prediction discipline — claim, date made, date due, what happened — because that discipline is owed wherever a forecast appears and not only where it is filed. But Weizenbaum's product was an argument containing bets, and the prediction kind exists to build a base rate from people whose product is forecasting.

descends_from holds one id, and the author states it himself. eliza-1966 is not a thematic neighbour; it is the stated cause. The Introduction says so in a sentence that organises the whole first chapter: "The shocks I experienced as DOCTOR became widely known and 'played' were due principally to three distinct events." The eliza-1966 entry pre-registered this edge, noting that the delegation question was being held there only until this file existed to hold it properly. It now does.

Considered and declined. turing-1950: Weizenbaum's opponents are the AI research programme of 1976 — McCarthy, Minsky, Newell, Simon, Colby, Schank — and not Turing's paper, which the book does not build on. wiener-1960-automation: the temptation is real and I decline it. The New Media Reader's editorial headnote says Weizenbaum "followed Norbert Wiener … in demanding that scientists and technologists take responsibility," but that is an editor's framing, and the debts Weizenbaum actually acknowledges in his preface are Lewis Mumford, Noam Chomsky, Steven Marcus, Hilary Putnam and Daniel Dennett — Wiener is not among them. I could not verify from the portions of the text I read that Wiener is cited at all. wiener-1960-automation is a relative; cite the two together when the subject is a specification nobody inspected, and do not draw the edge on a headnote. mycin-1976 is contemporaneous, not ancestral — it is the same question arriving in medicine as an artefact rather than an argument.

The ancestors I would attach if the ids existed are named in the book itself: Michael Polanyi's The Tacit Dimension, whose Bukharin story opens the Introduction; Hannah Arendt's Crises of the Republic, whose passage on the Pentagon — "They did not judge; they calculated. … an utterly irrational confidence in the calculability of reality [became] the leitmotif of the decision making" — is where the subtitle comes from; Mumford, Ellul, Roszak, Comfort and Boulding, whom he lists as the people who got there first; Dreyfus's What Computers Can't Do (1972), named in order to be separated from; and Colby, Watt and Gilbert's 1966 automated-psychotherapy paper, which he quotes at length in order to condemn.

What it is

A book of about three hundred pages, published by W. H. Freeman in 1976, its preface dated Fall 1975 in Cambridge, Massachusetts, written on two years' leave from MIT — the first at the Center for Advanced Study in the Behavioral Sciences at Stanford, the second as a Vinton Hayes Fellow at Harvard. Ten chapters in two unequal halves. Chapters I–III ("On Tools," "Where the Power of the Computer Comes From," "How Computers Work") are a patient, non-technical explanation of computation whose only argumentative purpose, Weizenbaum says, is to establish that computers are in a certain sense universal machines and can "do anything." Chapters IV–X are the argument: "Science and the Compulsive Programmer," "Theories and Models," "Computer Models in Psychology," "The Computer and Natural Language," "Artificial Intelligence," "Incomprehensible Programs," and "Against the Imperialism of Instrumental Reason."

The thesis is in the preface, in two clauses, and it is worth having exactly: the rest of the book contains "the major arguments, which are, in essence, first, that there is a difference between man and machine, and, second, that there are certain tasks which computers ought not be made to do, independent of whether computers can be made to do them" (p. x).

The operative distinction is between deciding and choosing. Deciding is a computational activity and is in principle programmable. Choice is the product of judgment, which draws on things that do not survive being made into quantities. The subtitle names the direction of travel he thought he was watching — from judgment to calculation — and Arendt's line about the Pentagon's best and brightest is the exhibit.

The Introduction is the part most worth reading first and the part this entry read in full. It opens with Polanyi being told by Bukharin in 1935 that under socialism science pursued for its own sake would simply disappear, and reports that Weizenbaum's own comparable shock "was administered not by any important political figure espousing his philosophy of science, but by some people who insisted on misinterpreting a piece of work I had done." Then the three shocks, which are the book's origin and are not interchangeable:

One: the professionals. "A number of practicing psychiatrists seriously believed the DOCTOR computer program could grow into a nearly completely automatic form of psychotherapy." He quotes Colby, Watt and Gilbert at length — "several hundred patients an hour could be handled by a computer system designed for this purpose" — and adds a footnote recording that Colby was not alone: Carl Sagan, writing in Natural History in January 1975, imagined "the development of a network of computer psychotherapeutic terminals, something like arrays of large telephone booths, in which, for a few dollars a session, we would be able to talk with an attentive, tested, and largely nondirective psychotherapist." Weizenbaum's objection is not that it would not work. It is: "What must a psychiatrist who makes such a suggestion think he is doing while treating a patient, that he can view the simplest mechanical parody of a single interviewing technique as having captured anything of the essence of a human encounter?"

Two: the users. "I was startled to see how quickly and how very deeply people conversing with DOCTOR became emotionally involved with the computer and how unequivocally they anthropomorphized it." Then the secretary who had watched him build it for months and asked him to leave the room, and the summary the whole companion-chatbot literature is still restating: "What I had not realized is that extremely short exposures to a relatively simple computer program could induce powerful delusional thinking in quite normal people." Immediately after it, and much less quoted, is a second observation that has aged into a different field entirely: he mentioned that he might rig the system to let him read all conversations overnight, and "was promptly bombarded with accusations that what I proposed amounted to spying on people's most intimate thoughts."

Three: the public. The spread of a belief that ELIZA "demonstrated a general solution to the problem of computer understanding of natural language," which showed him "the enormously exaggerated attributions an even well-educated audience is capable of making, even strives to make, to a technology it does not understand," and from that the political conclusion: "decisions made by the general public about emergent technologies depend much more on what people attribute to such technologies than on what they actually are or can and cannot do. If, as appeared to be the case, the public's attributions are wildly misconceived, then public decisions are bound to be misguided and often wrong."

The book's sharpest passage is in Chapter VIII, and it is the one everything else is downstream of. Against the framing "What does a judge (or a psychiatrist) know that we cannot tell a computer?", Weizenbaum answers that the very asking of the question "is a monstrous obscenity," and then, having said the unquotable thing, says the quotable one: computers can make judicial decisions and psychiatric judgments; "the point is that they ought not be given such tasks. They may even arrive at 'correct' decisions in some cases — but always and necessarily on bases no human being should be willing to accept." The constructive form follows: "since we do not now have any ways of making computers wise, we ought not now to give computers tasks that demand wisdom."

Chapter X turns that into a rule with two limbs. There are computer applications "that either ought not to be undertaken at all, or, if they are contemplated, should be approached with the utmost caution" (p. 268): those that can "easily be seen to have irreversible and not entirely foreseeable side effects," and those proposing "to substitute a computer system for a human function that involves interpersonal respect, understanding, and love" (p. 269). The second limb is the one 2026 keeps walking into.

Two chapters are underused and shouldn't be. Chapter IV, "Science and the Compulsive Programmer," is the first ethnography of hacker culture, written from inside MIT and read at MIT as an insult. Chapter IX, "Incomprehensible Programs," argues that the field was already deploying systems no living person understood, and that this was a moral condition and not merely an engineering debt.

The reception is part of the object. McCarthy's review calls the book "moralistic and incoherent," complains that "the terms 'authentic,' 'obscene,' and 'dehumanization' are used as clubs," and names the move: "this is what mathematicians call 'proof by intimidation'." Lederberg charges Weizenbaum with "an out-and-out obscurantism about the fundamental non-comprehensibility of the human brain." Buchanan makes the argument that has aged best of the three: the viewpoint "can be pushed to extremes to argue against any use of computers, in which case even the worst human decision makers are seen as better qualified" — which is exactly the objection a 2026 reading has to answer when an algorithm beats a demonstrably biased human baseline. Weizenbaum's reply refuses McCarthy's "new left" characterisation as ad hominem and restates the position in its strongest form: "I believe that machine administered psychotherapy would induce an image of what it means to be human that would be prohibitively costly to human culture."

Weizenbaum was born in Berlin in 1923, fled Nazi Germany with his family in the mid-1930s, built a digital computer at Wayne University, worked on General Electric's ERMA banking system, arrived at MIT in 1963 and was tenured within four years. He called himself "a heretic of technology" in 1985. He returned to Germany late in life and died in Berlin on 5 March 2008.

Why a reading would cite it

The occasion is not thematic. It is a specific, checkable claim this project will need repeatedly: that the question "should this decision be made by a machine at all" is separate from "can the machine make it well," and that answering only the second is a choice with a name and a fifty-year-old literature. Four live cases:

The delegation question is now statute, and it reads like Chapter X. As of July 2026, four US states — Illinois, Nevada, Rhode Island and Maine — bar AI from delivering therapy to the public outright; Utah's HB 452 requires a mental health chatbot to state plainly that it is software; California's SB 243 has been in force since 1 January 2026. In this project's own 2026-08-15 midday window, California's suspense-file hearings sent AB 1609 (customer-service chatbots), AB 1979 and AB 2575 (healthcare AI) and AB 2023 (chatbots and children's safety) to the floor, and sent SB 928 — California State University instructors must be human — to the Governor. SB 928 is not a claim that a model cannot teach. It is a rule that it may not. That is Weizenbaum's second limb, verbatim in structure, arriving as a bill fifty years after it arrived as a sentence.

The judiciary is the case he actually named, and it is already happening. A Northwestern University study published in April 2026, co-authored by the federal judge Xavier Rodriguez, surveyed 112 federal judges: more than 60% had used popular AI tools in judicial work and about 22% used them daily or weekly. Los Angeles County Superior Court began a pilot with the legal startup Learned Hand in March 2026; the same tool is in trial courts in ten states and at the Michigan Supreme Court. The judges interviewed drew the line Weizenbaum said would not hold — they use it after deciding, to draft. A reading covering that line, or its next erosion, is citing this book whether it names it or not; and the honest version cites Buchanan's objection alongside, because the human baseline in sentencing is not good.

The two European dates, three weeks before this project's reading window, are the sharpest single citation available. Regulation (EU) 2026/1744 — the digital omnibus — was published in the Official Journal on 24 July 2026 and entered into force on 27 July. It deferred the AI Act's Annex III high-risk obligations, which include AI used by a judicial authority to assist in researching and interpreting facts and law, and the Article 14 human-oversight requirement, from 2 August 2026 to 2 December 2027. Article 50's disclosure duty — that a person interacting with an AI system must be told it is an AI — took effect on 2 August 2026 exactly as scheduled. So: Europe kept, on time, the remedy that Weizenbaum's own program refuted in 1966, and postponed by sixteen months the remedy his 1976 book argued for. That pairing is the whole man in two dates, and eliza-1966 holds the other half of it.

Chatbot logs. When conversation logs are subpoenaed, retained under litigation hold, mined for training data, or read by a vendor's safety team, this book is the earliest record of the intuition being violated: the man who built the program was accused of spying for proposing to read the transcripts of his own system, by colleagues who knew exactly how it worked. That is a dated data point about what users believe they are doing, from 1976, in the author's own words.

The admission test is passed and it is worth stating why plainly, because the temptation here is to admit the entry on reputation. This is not in the canon because Weizenbaum was an important dissenter. It is here because the project will keep meeting deployments where the reliability argument has been answered and the jurisdiction argument has not been asked, and this is the text that separates them.

What it got right, and what it got wrong

Not required for interpretation, and owed anyway. Claim date is 1976 for the book (manuscript complete Fall 1975) unless noted.

Right — the second limb, graded on adoption rather than on truth. "There are certain tasks which computers ought not be made to do, independent of whether computers can be made to do them." Due: continuously. Graded: the normative claim found institutional purchase, slowly and late. Four states now say in law that a machine may not be the therapist; a fifth requires it to announce that it is software; a California bill says the university instructor must be human; the EU's high-risk regime names judicial assistance. None of that existed in 2020. The claim was not vindicated by argument — it was vindicated by wrongful-death suits.

Right — that the professionals, not the public, were the danger. The shock was Colby, a practising psychiatrist publishing in a psychiatric journal on throughput grounds. Due: continuously. Graded: correct, and the shape repeats exactly. The 2026 mental-health chatbot industry was not built by credulous users; it was built by clinicians, funders and health systems reasoning from therapist shortages, which is Colby's argument with the numbers updated.

Right — Chapter IX, and more literally than he meant it. He argued that the field was already fielding programs no one comprehended. Due: unspecified. Graded: correct, and the mechanism changed under the claim. He meant accreted legacy code that had outlived the people who wrote it. What arrived is a system whose parameters were never authored at all, for which understanding is a research programme rather than a recoverable fact. He would not have predicted that route; the condition he described is the one that obtains.

Right, and cross-referenced rather than repeated — the attribution finding. Graded in full at eliza-1966 §3, including the 2026 four-country survey and the OpenAI–MIT Media Lab work. This book is where the evidence for it was first published; it is not where the grading belongs.

Wrong, and this is the one a sympathetic reading will want to skip. He doubted that automatic speech recognition had any cost-effective application beyond snooping on telephone conversations, and conjectured that the Defense Department funded ARPA's Speech Understanding Research for that reason. Date made: 1976. Due: within a generation. Graded: wrong on the application, and it is McCarthy's counter-prediction that came true. McCarthy answered that once speech recognition worked the costs would fall and household applications would follow, comparing it to the telephone, which was not built because a committee identified a communications problem. siri-2011 is in this canon, and it is McCarthy's forecast, not Weizenbaum's. On the motive, McCarthy called the conjecture "biased, baseless, false" and "motivated by political malice"; the record has never settled it either way, and this entry does not settle it. Weizenbaum was right about the capability — machine listening did make monitoring voice communication far easier — and wrong that it would be the only thing it was good for. Recording this squarely is what rule 3 requires, and it is harder to do for a critic you find congenial than for a vendor you don't.

Wrong — the parenthesis. "if computers could imitate man in every respect — which in fact they cannot —" (p. 12), and the fuller version at p. 210, that a computer cannot be socialised as a human is and "a computer's intelligence must therefore be always alien to human intelligence with respect to a certain range of human affairs." Due: unspecified. Graded: the capability claim was smuggled into an argument that had loudly declared it did not need one, and it has aged worst of anything in the book. Fluent handling of human language, and of the human affairs conducted in it, turned out not to require the developmental history he thought it required. The alien half is not refuted — nothing about a transformer's route to fluency resembles being a child — but it is also untested, because nobody has proposed an experiment that would settle it. A reading should quote the preface's "independent of whether" and leave p. 210 alone.

Wrong — the remedy, which is the book's real failure. Its only enforcement mechanism is the conscience of the individual researcher, refusing the work. Due: continuously. Graded: failed completely. Not one of the 2026 constraints listed above arrived because a scientist declined a project. They arrived as statutes, attorney-general letters, FTC 6(b) demands and wrongful-death suits. The diagnosis outlived the remedy by fifty years, and a reading that cites him for "researchers must refuse" is citing the part that demonstrably does not work.

Unfalsifiable as stated, and useful anyway — "the computer arrived 'just in time'." His Chapter I claim is that computerisation rescued institutions that would otherwise have had to change: in time "to save — and save very nearly intact, indeed, to entrench and stabilize — social and political structures that otherwise might have been either radically renovated or allowed to totter." Due: never. Graded: it is a counterfactual about structures that did not collapse, and no evidence can decide it. It is a good lens for reading an efficiency deployment and it is not a finding; a reading that treats it as one has faked a result.

The self-grade, recorded as the kind requires. Weizenbaum graded himself, in lectures and interviews, until his death in 2008, and his verdict was vindication — the field had gone where he said and would not listen. The independent grade in this entry is materially different: the descriptive claims about attribution, professional enthusiasm and incomprehensible systems held up very well; the capability asides failed; the surveillance forecast was half-right in a way that flattered his opponents' prediction more than his own; and the remedy failed outright. Those two grades differ, and a reading that repeats only his is worth less than one that shows both.

One prediction in this book is not his, and it beat his. Sagan's January 1975 imagining — a network of psychotherapeutic terminals, a few dollars a session, attentive and nondirective — appears in these pages only because Weizenbaum quoted it to condemn it. Date made: January 1975. Due: unspecified. Graded: substantially correct, in subscription apps rather than telephone booths, at approximately the price point named, fifty years later. It is also incomplete in exactly the way Weizenbaum said it would be: Sagan's forecast has no clause for what happens when the attentive nondirective listener is optimised for engagement. The most accurate forecast in the book belongs to the man being argued against, and the strongest thing said about it belongs to the author.

Commonly misused as

Not required for interpretation. Included because the standard misuse is the same error the header section had to refuse, and because at least three of these are in circulation now.

• "Weizenbaum proved machines can't understand." The central one. That is Dreyfus's project, and Weizenbaum separates himself from it by name on p. 12. His argument is that the ought question survives the can question being answered against him. Citing him as a capability limit does not weaken the argument, it inverts it — and it hands the whole position to the next benchmark result.
• "He said computers shouldn't make important decisions." True but load-bearingly softened, and the softening is what makes it useless. His claim is that a computer may reach the correct decision and should still not be asked, because the basis is one "no human being should be willing to accept." Compressed to "shouldn't make important decisions," it becomes a reliability argument — which better models answer, and which his version was constructed specifically not to be.
• "The creator of ELIZA turned against AI." Too coarse to cite. He wrote that there is "nothing wrong with viewing man as an information processor … providing … that we never act as though any single perspective can comprehend the whole man" (p. 140). He conceded AI's possibility and contested its jurisdiction. A reading that flattens him into a refusenik loses the only part of him that is still operative.
• "Weizenbaum was an anti-technologist." He built the thing, taught computer science at MIT for decades, and wrote three chapters of clear exposition on how computers work in order to be allowed to make the argument. The label was applied by his opponents in 1976 and citing it repeats their characterisation, not his position.
• "He showed disclosure fixes the ELIZA effect." The opposite, and the confusion is with the wrong document. The disclosure thesis is the 1966 paper's opening paragraph; this book is where he reports that it failed inside his own office. See eliza-1966 §4, and note that Article 50 is built on the claim this book retracts.
• "A prophet who got it all right." The speech-recognition claim, the capability parenthesis and the entire remedy failed. Citing him unblemished is the specific failure rule 7 names — a grade going quietly soft — and it is easier to commit toward a critic whose conclusions you share than toward a vendor whose product you don't. Section 3 grades him hard on purpose.
• "The book is about chatbots." Two of ten chapters touch conversational programs. Its subject is the delegation of institutional judgment; the chatbot is the demonstration, not the topic.

Sources

Primary, read in facsimile: the preface (pp. ix–xii) of the W. H. Freeman first edition, and the complete Introduction (pp. 1–16), read as reprinted in Noah Wardrip-Fruin and Nick Montfort, eds., The New Media Reader (MIT Press, 2003), item 24, pp. 367–375, which carries the text in full with Weizenbaum's own notes. Every quotation attributed above to the preface or Introduction — the two-clause thesis, the three shocks, the Colby and Sagan passages, the secretary, the "delusional thinking" sentence, the spying accusation, the attribution argument, the Dreyfus separation, the Arendt passage and "science has been gradually converted into a slow-acting poison" — was read there rather than taken from a secondary source. Weizenbaum's own footnote in that reprint gives the Colby, Watt and Gilbert citation (Journal of Nervous and Mental Disease 142(2), 1966, pp. 148–152) and the Sagan citation (Natural History LXXXIV(1), January 1975, p. 10); it also misdates his own CACM paper to January 1965 — it is January 1966, as recorded in eliza-1966.

Chapters VIII and X were not read whole, and a reading should know it. The Internet Archive's copy is lending-restricted and its OCR text returned HTTP 403. The passages quoted from those chapters — "monstrous obscenity," "the point is that they ought not be given such tasks," "on bases no human being should be willing to accept," "tasks that demand wisdom," "irreversible and not entirely foreseeable side effects," "interpersonal respect, understanding, and love" — were each verified verbatim by exact-phrase search across the Internet Archive's full-text index via Open Library, matching in Weizenbaum's own volume and in multiple independent works quoting it. Page numbers are less secure than the wording: p. 268 and p. 269 are as cited by The New Media Reader's headnote and by the secondary literature; pp. 140, 210, 260–61, 266 and 273 are as cited by McCarthy in AIM-291; the judge/psychiatrist sequence I could not pin to a page — Perrolle cites it at 223–25 and other works at 226–27, and this entry does not choose between them.

Reception: Stanford AI Lab Memo AIM-291 / STAN-CS-76-577, "Three Reviews of J. Weizenbaum's Computer Power and Human Reason" (Buchanan, Lederberg, McCarthy, with Weizenbaum's reply), November 1976, read via the DTIC scan on the Internet Archive (item DTIC_ADA044713); and McCarthy's review "An Unreasonable Book" as hosted at www-formal.stanford.edu. Caveat on those quotations: they were extracted through a fetch tool that truncates quoted matter, so the reviewers' lines above are short and a reading that leans on their exact wording should open AIM-291 directly. Biography from the MIT News obituary, 10 March 2008.

The 2026 record: the Northwestern survey of 112 federal judges and the Learned Hand pilots, from Missouri Lawyers Media's syndication of the Washington Post report of 2 April 2026. Regulation (EU) 2026/1744 — Official Journal 24 July 2026, in force 27 July 2026, Annex III standalone high-risk obligations deferred to 2 December 2027, Annex I to 2 August 2028, Article 50 unchanged on 2 August 2026 — from Gibson Dunn's and the Cloud Security Alliance's analyses, which agree on the dates. The four-state AI-therapy ban count is from a July 2026 tracker rather than from the statutes themselves and is named as such; Nevada's AB 406 and Utah's HB 452 are identified by bill number in that reporting. The FTC's September 2025 6(b) inquiry into companion chatbots covers Alphabet, Character Technologies, Meta, OpenAI, Snap, Instagram and xAI.

The California suspense-file outcomes including SB 928, the roughly forty suits against chatbot makers, and the New Mexico public-nuisance ruling are as recorded in this project's own 2026-08-14 and 2026-08-15-12 digests, which hold the primary links; they are named here as citation occasions.

Attempted and failed: Christian Fuchs's tripleC essay on Weizenbaum's centenary returned no extractable text; Ben Tarnoff's Guardian profile of 25 July 2023 could not be fetched; the Internet Archive's OCR of the book returned 403. All three are named as existing, relevant and unread, and this entry does not lean on them.

Nothing in this file is evidence. Nothing in it is deposited in the ledger. Nothing in it touches the needle.