"The Demon-Haunted World: Science as a Candle in the Dark"

Filed correctly as interpretation, and the id's year is right — which is worth saying, because canon/contact-1985 flagged the year as unsettled and told whoever wrote this file to settle it rather than inherit it. It is settled below, in favour of the id.

build.py defines interpretation as "a reading of events that a reading may need to name," and that is precisely what this book is: a diagnosis of the condition of a society, delivered as an argument, by a man whose product was explanation. The exact precedent is weizenbaum-1976 — a scientist's book-length argument about technology and the public, filed interpretation on the test wiener-1960-automation set. Nothing distinguishes the two on kind.

Two rivals, both real, both refused.

prediction has the strongest claim and I nearly took it. Chapter 2 contains a dated forecast with a due window, quoted in full below, and this canon weights predictions heaviest. But kurzweil-2005 settled the test for this adjudication and the test decides against: a predictor's credential is his hit rate. Kurzweil publishes a scorecard and is introduced by his percentage. Sagan published no scorecard, offered no track record, and the foreboding passage is one paragraph inside a 457-page argument whose subject is a method, not a calendar. His argument is not scaffolding for the forecast; the forecast is one rhetorical move inside the argument. Filed prediction, this entry would be a file about a paragraph.

The forecast is graded anyway, in section 3, under the house rule that wiener-1960-automation set and wargames-1983 and contact-1985 restated: grade the claim wherever it appears, not only where it is filed. That discipline matters more here than usual, because the foreboding passage is the single most-circulated thing Sagan wrote and it is almost never graded — it is quoted as though the quoting settles it.

idea is the other rival and it fails on the definition rather than on taste. build.py: "a framing later work is built out of rather than argued about." The baloney detection kit is the candidate, and it is not an origination. Sagan did not invent independent confirmation, Occam's razor, or falsifiability; he assembled nine borrowed tools and twenty borrowed fallacies into a curriculum for people who had never been handed one. A curriculum is a very good thing and is not a framing. doomsday-clock-1947 was refiled into idea because a single scalar standing for a diffuse judgment was genuinely a new object. Nothing in this book is a new object.

Scope: the book and one interview, deliberately. canon/proposals.md asked for both — "in The Demon-Haunted World and in his final television interviews" — and both belong here for a reason that only shows up when you set them side by side. The central sentence exists in two versions, written and spoken, and they do not say quite the same thing. Quotations that merge them are everywhere. Section 4 is about that.

The year, settled. The evidence is split and the split has a boring cause. The US copyright page carries 1995, which is what the Library of Congress record and Wikipedia follow, and it is why the book is universally described as "a 1995 book." The publication date is 1996: Kirkus reviewed it on 15 December 1995 with a stated pub date of 1 February 1996; Open Library gives the Random House hardcover as 1 March 1996; the UK Headline edition's own front matter says "first published 1996"; and Science's review of the first edition cites it as "Random House, New York, 1996." Copyright year 1995, publication early 1996. The interview is 27 May 1996. The id is right and Wikipedia's "1995" is a copyright date, not a contradiction.

descends_from holds one id and the arrow is authorship, not documented influence. contact-1985 is the same author working the same problem a decade earlier, and the canon has already asserted this relationship in its own words: "Sagan wrote the story first and the rulebook afterwards." That entry named this one as the ancestor a reading would most want and recorded that it did not exist. It exists now. I did not find Sagan saying anywhere that the novel led to the book, and I am not claiming he did — the arrow records that the canon's Contact entry is unreadable without this one and says so itself.

One id I refused, and it is the most interesting fact in this file. weizenbaum-1976 is not an ancestor — there is no evidence Sagan read Weizenbaum, and inventing that arrow would be exactly the move this book exists to prevent. But Sagan is already in the canon, inside that file, on the wrong side of the argument. Weizenbaum's footnote quotes Sagan writing in Natural History in January 1975, imagining "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 quoted it to condemn it. The canon grades it substantially correct, fifty years later, in subscription apps rather than telephone booths, at about the price named — and incomplete in exactly the way Weizenbaum said, because it has no clause for what happens when the attentive nondirective listener is optimised for engagement.

So the canon now holds two Sagans twenty-one years apart: the one who was cheerfully credulous about a machine's capacity to stand in for a human judgment, and the one who wrote the manual on credulity. A reading citing this entry for Sagan's authority on not being fooled should know that the canon also holds the receipt. That is not a strike against him. It is the strongest available demonstration that his subject was a discipline anyone can fail, which is what he said it was.

Last, the admission problem, stated before the case is made: this book is not about artificial intelligence. Twenty-five chapters on UFO abduction, faith healing, recovered memory, channelling, the face on Mars and witch trials, and essentially nothing on machines that generate language. It is admitted here for a precondition, not for a subject, and a precondition is the weakest kind of citation there is. Section 2 has to earn it or the entry fails.

What it is

The book

The Demon-Haunted World: Science as a Candle in the Dark was published by Random House in early 1996, copyright 1995, 457 pages, twenty-five chapters, four of them co-written with Ann Druyan. It was a New York Times bestseller and the last book Sagan published in his lifetime — Billions and Billions followed in 1997, with an epilogue Ann Druyan wrote after his death. He was diagnosed with myelodysplasia in 1994 and died on 20 December 1996. The dedication is five lines: "To Tonio, / My grandson. / I wish you a world / Free of demons / And full of light."

Most of it is case work. Chapters 3 through 11 take the pseudosciences of the early nineties one at a time — the face on Mars, alien abduction, hallucination, demonology, recovered-memory therapy — and walk each through the evidence. That material has dated, and the dating is itself a finding, addressed in section 3.

Two chapters carry the method and are the reason the book survives.

Chapter 10, "The Dragon in My Garage." Sagan claims a fire-breathing dragon lives in his garage. You ask to see it: the dragon is invisible. You propose flour on the floor for footprints: it floats. Infrared for the fire: the fire is heatless. Spray paint to outline it: it is incorporeal. Every test you devise is met by an amendment that exactly cancels it. His question is the one the parable exists for: "what's the difference between an invisible, incorporeal, floating dragon who spits heatless fire and no dragon at all?" And then the sharp end: "If there's no way to disprove my contention, no conceivable experiment that would count against it, what does it mean to say that my dragon exists?" It is the cleanest short statement of falsifiability in popular English, and it does its work by making the claimant sympathetic rather than absurd — Sagan is the one with the dragon.

Chapter 12, "The Fine Art of Baloney Detection." Nine tools:

1. Wherever possible, independent confirmation of the "facts." 2. Substantive debate on the evidence by knowledgeable proponents of all points of view. 3. Arguments from authority carry little weight — authorities have made mistakes before, and in science there are no authorities; at most, there are experts. 4. Spin more than one hypothesis, and think of tests that could disprove each. 5. Try not to get overly attached to a hypothesis just because it is yours. 6. Quantify. What is vague and qualitative is open to many explanations. 7. Every link in a chain of argument must work, not most of them. 8. Occam's razor. 9. Ask whether the hypothesis can be falsified. Propositions that are untestable are not worth much.

Then twenty fallacies of logic and rhetoric — ad hominem, argument from authority, argument from adverse consequences, appeal to ignorance, special pleading, begging the question, observational selection, statistics of small numbers, misunderstanding the nature of statistics, inconsistency, non sequitur, post hoc ergo propter hoc, meaningless question, excluded middle, short-term versus long-term, slippery slope, confusion of correlation and causation, straw man, suppressed evidence, weasel words.

The thesis is in chapter 2, "Science and Hope," in two passages I verified against the book's own text. The first:

> We've arranged a global civilization in which most crucial elements > profoundly depend on science and technology. We have also arranged things so > that almost no one understands science and technology. This is a prescription > for disaster. We might get away with it for a while, but sooner or later this > combustible mixture of ignorance and power is going to blow up in our faces.

The second, six pages of argument later, is the forecast:

> I have a foreboding of an America in my children's or grandchildren's time — > when the United States is a service and information economy; when nearly all > the key manufacturing industries have slipped away to other countries; when > awesome technological powers are in the hands of a very few, and no one > representing the public interest can even grasp the issues; when the people > have lost the ability to set their own agendas or knowledgeably question > those in authority; when, clutching our crystals and nervously consulting our > horoscopes, our critical faculties in decline, unable to distinguish between > what feels good and what's true, we slide, almost without noticing, back into > superstition and darkness.

It continues into the media — the thirty-second sound bite "now down to ten seconds or less," lowest-common-denominator programming, and "especially a kind of celebration of ignorance."

Whose fault it is, according to him, matters for section 4. Sagan does not blame the public. Chapter 1 puts the failure on the culture's motifs, the schools and the media for never having given people access to real science; chapter 19 is titled "No Such Thing as a Dumb Question"; and one of his standing complaints is against scientists who will not do the work of explaining.

The interview

Sagan appeared on The Charlie Rose Show on 27 May 1996 to discuss the book. It is widely described as his last interview; the Charlie Rose archive records it as his last appearance on that programme, and Rose introduced a clip from it as "our most recent interview" in the memorial he broadcast on 20 December 1996, the day Sagan died. I could not confirm it was the last interview he gave anywhere, and the entry does not need it to be.

The spoken version of the thesis:

> We've arranged a society based on science and technology in which nobody > understands anything about science and technology, and this combustible > mixture of ignorance and power, sooner or later, is going to blow up in our > faces. Who is running the science and technology in a democracy if the people > don't know anything about it?

And the consequence he drew from it:

> Science is more than a body of knowledge. It's a way of thinking. A way of > skeptically interrogating the universe with a fine understanding of human > fallibility. If we are not able to ask skeptical questions, to interrogate > those who tell us that something is true, to be skeptical of those in > authority, then we're up for grabs for the next charlatan political or > religious who comes ambling along.

The two versions differ and the difference is not cosmetic. The book says "global civilization" and "almost no one understands"; the interview says "a society" and "nobody understands anything about it." The written claim is hedged, planetary and about depth of understanding. The spoken one is absolute, national and about any understanding. The written claim is the defensible one. The spoken one is the one that circulates.

Why a reading would cite it

The proposal's case is too broad and I am going to narrow it before accepting it. canon/proposals.md says to cite this "when a legislature, a court or a regulator is acting on AI — the competence gap he described is the standing precondition for every policy finding this project files." A precondition for every finding is a citation for none: an entry that applies to everything on a board applies to nothing on it, and it would end up as decoration on the policy lens, invoked whenever a reading wanted to sound rueful about Congress. That is the failure mode this kind is most prone to and it should be named.

What survives narrowing is better than what was proposed, because the useful thing about this entry is not that it confirms the competence gap. It is that it forces four distinctions a reading would otherwise blur. Four occasions.

One: when a policy finding turns on whether the decider understood the object — and the entry's job is to stop the reading from assuming they didn't.

This project's own record from the last week contains all three modes and they grade differently.

Incomprehension is not what the FDA showed. On 18 August 2026 it opened a docket on generative-AI medical devices with a discussion paper sketching a competency-based route modelled on physician licensing: nonclinical benchmarking across clinical knowledge, analytical ability, safety behaviour and performance across patient populations; clinical confirmation with real patients; periodic retesting and monitoring for performance decay; explicit coverage of agentic systems that plan and execute multistep tasks; and the statement that the object of evaluation is the finished device as used rather than the model underneath. Whatever else that is, it is not a regulator who cannot grasp the issue. A reading that reaches for Sagan on the day a regulator demonstrates competence has cited him against the evidence.

Nor is incomprehension what California showed. On 13 August 2026 two appropriations committees held simultaneous suspense-file votes on roughly thirty AI bills, where a bill the chair declines to call aloud dies with no recorded vote. This reading's own note at the time: "the single largest pending change in US AI law and it is being decided by which bills get named aloud in a committee room." That is a procedural failure, not a comprehension failure, and the two want different remedies. Sagan's diagnosis prescribes education. A suspense file is not fixed by education.

And Pennsylvania showed a third thing Sagan has no category for. On 18 August 2026 the governor conditioned data-centre development on power generation, clean-energy requirements, infrastructure cost, local approval, noise, air, water, local hiring and community benefit agreements. That is a state regulating the substrate it does understand — land, electricity, water — because the model is out of reach. It is a competent act by someone who has conceded the competence gap and routed around it, which is a strategy, not a capitulation.

Cite this entry when the reading has to say which of the three it is looking at. Its value is the taxonomy, not the lament.

Two: when the remedy on the table is public education — because Sagan's remedy was partially delivered and the outcome did not follow.

This is the most citable finding in the file and it argues against the reflex it appears to support. Sagan's prescription was scientific literacy: teach people the method and they will not be had. Between 1988 and 2008 the share of US adults scoring as civically scientifically literate on Jon Miller's NSF-funded instrument rose from 10% to 28% — nearly tripled, right across the years he was writing — and has roughly plateaued since. The literacy went up. The world he feared arrived anyway.

That does not make him wrong about the danger. It makes the causal claim inside the passage — that declining critical faculties are what deliver a population to whoever speaks confidently — unsupported by its own instrument. A reading tempted to conclude that the answer to AI is AI literacy now has thirty years of evidence that the analogous intervention was delivered and did not produce the analogous result. That is a base rate, and it is what this canon is for.

Three: when the competence gap is asserted about the public and the observed failure is in credentialed professionals.

Sagan's passage is about "the people" and about those "representing the public interest." The measured 2026 failure is somewhere else. Damien Charlotin's public database of court decisions in which a party relied on AI-fabricated material recorded roughly 200 cases in mid-2025, 719 by January 2026, 1,227 by early April and 1,598 by 9 June — a run rate around eight new documented cases a day — with one tally of 1,668 by 2 July split 653 involving lawyers and 975 pro se litigants. Sanctions have moved from four-figure fines to $15,000 per attorney and the first bar suspensions.

Two things a reading should take from that and neither is "people are credulous." First, it is a refinement of Sagan's thesis rather than a confirmation of it: over a third of the documented failures are officers of the court, professionally trained in verifying citations, defeated by a tool whose failure mode they had not internalised. The gap is not between the educated and the uneducated. It is between people who know how a system works and people who know how to use it, and those are different populations inside the same profession. Second, the database is a record of what a judge caught and wrote up. A fabricated citation nobody checks does not appear in it. That is a selection effect and it runs in the direction of understatement, which a reading citing the number should say.

Four: when a reading is about to defer to a lab — because the kit's rule 3 forbids it and the book's reputation invites it.

Tool 1 is "wherever possible there must be independent confirmation of the 'facts.'" Tool 3 is "arguments from authority carry little weight — in science there are no authorities; at most, there are experts." Set beside each other those two sentences are the whole of this project's hardest daily problem, stated in 1996 for a different subject.

Tool 1 is the same rule contact-1985 extracted from the 1989 IAA post-detection declaration and from the Wow! signal's forty-nine unconfirmed years, and it cuts the same way: an unverifiable claim is not evidence in either column, whether the lab's self-report is flattering or unflattering to the lab. Tool 3 is the sharper one and it is the one Sagan's own reputation works against. He is the most-quoted source in English for "listen to the scientists," and he wrote down in plain terms that authority is not a reason. A frontier lab is an expert. It is not an authority. A reading that cites this book to justify deferring to a lab's internal evaluation is citing the book against itself — which is the same error contact-1985 catalogued as "a reading is always Kitz and never the audience," arriving from the opposite direction.

The concrete remedy is in the record and it is institutional rather than educational. METR's Frontier Risk Report of 19 May 2026 gave an outside body access to four labs' most capable internal models including raw chains of thought, with no participant approval rights over the final report. On 18 August 2026 the AI Observatory gave outsiders 24,521 real conversations to check published usage reports against, and its first result was that applying one lab's own filtering method removes 48% of conversations and removes disproportionately the difficult half. Neither of those is public literacy. Both are tool 1. The thing that closes the gap Sagan described turns out not to be the thing he prescribed, and a reading can say so with this entry behind it.

A reflexive note, since this canon's method is the project's own business. Tool 6 says quantify — "what is vague and qualitative is open to many explanations" — and this project's needle does exactly that, converting a diffuse judgment into one integer. Tool 6 and bin/ethos-check.py are not in conflict, and the distinction is worth having straight: the check refuses fake precision, not quantification. A bare integer labelled as judgment satisfies Sagan; an integer with a decimal place and an error bar would satisfy neither of them. doomsday-clock-1947 is the entry for what happens when the units arrive anyway.

What it got right, and what it got wrong

Not required for interpretation. Required here, because the foreboding passage is a dated forecast and this canon grades those wherever they appear.

The claim, its date, its due date, and one thing this entry cannot supply. Claim: the passage quoted in section 1. Date made: chapter 2 of a book copyrighted 1995 and published February–March 1996; the spoken version 27 May 1996. Due: "in my children's or grandchildren's time" — Sagan's children were born between 1959 and 1991, so the children's-time half is due now and has been for a while, while the grandchildren's half runs decades out. Graded here on 20 August 2026, thirty years and six months after publication.

There is no self-grade, and that is a real gap rather than a convenience. The house rule for graded claims is to set the predictor's own score beside an independent one and say they differ. Sagan died six months after the interview and never revisited the passage, so this entry has one grade where the rule wants two, and that grade is mine. It should be read accordingly. The nearest independent readings I found are The Verge's 2017 retrospective, which called the passage "chillingly prescient" and recorded that it had gone viral, and Massimo Pigliucci's reassessment of the baloney chapter in Skeptical Inquirer 50(4), July/August 2026, which is discussed under Wrong. Both are cited below. Neither is a scorecard.

Right

"Awesome technological powers are in the hands of a very few." Right, and getting more right. This is the concentration lens and it needs no argument from me; the last week of this project's own record supplies it. SpaceX closed a $60bn all-stock purchase of Cursor's parent on 14 August 2026 and was reported five days later to have approached a second coding-agent company. Stripe confirmed OpenRouter at more than $8bn. OpenAI's CFO told employees the company would be public in 2027 with a prospectus confidentially filed in June. Venture funding into physical AI ran $47.4bn across 521 deals in the first half of 2026, against $12bn across 470 deals in the second half of 2025 — money per company roughly quadrupled while the number of companies barely moved. Four or five organisations can train a frontier model. He wrote that clause about a country that had just deregulated telecoms, and it landed on an industry that did not exist.

"The United States is a service and information economy." Right, plainly.

"No one representing the public interest can even grasp the issues." Right in substance, wrong in mechanism, and the correction is the useful part. He imagined incomprehension. What actually obtains is asymmetric access: the people who understand these systems best are employed by the organisations being regulated, and the outside bodies capable of grasping the issues have to be granted the material before they can. METR could evaluate frontier models because four labs let it. The AI Observatory could check published usage reports because a corpus was assembled. Neither of those is a literacy problem and neither is fixed by a better science curriculum. The bottleneck is standing, not comprehension — which means the passage is right about the outcome and wrong about the cause, and a reading that borrows the diagnosis without the correction will prescribe the wrong remedy.

The mechanism of the first passage — dependence plus incomprehension equals vulnerability to whoever speaks confidently — holds, and the AI-specific measurement is unusually clean. Pew's survey of 5,119 US adults fielded 17–23 February 2026 and published 17 June 2026 found 96% have heard at least a little about AI and about half say they are highly aware, up from 26% in 2022 — and then, asked what AI is, 29% answered "chatbots" or named a brand, 8% said robots or science fiction, 6% said generative AI or large language models. Only about one in five adults are extremely or very confident using chatbots and 51% do not use them at all. Meanwhile 36% say they interact with AI at least several times a day. Awareness tripled; the ability to say what the thing is did not follow it. That is the shape Sagan described, measured, in the domain this project watches, six months ago.

The dragon in the garage is the most transferable thing in the book and has not aged a day. "Our internal evaluation found X, the details are not releasable, the eval is proprietary, and the model is not deployed" is an invisible incorporeal dragon, and the parable is the shortest way to say why that is not a finding — without accusing anyone of lying, which is the parable's whole design. Sagan is the one with the dragon.

Wrong

"Nearly all the key manufacturing industries have slipped away to other countries." Overstated, and the overstatement is load-bearing. US manufacturing employment was 17.2 million in 1996 and about 12.6 million in April 2026 — a 27% fall over thirty years, against a share of total employment that roughly halved. That is a large decline and it is not "nearly all." Manufacturing output is a different series and did not fall the way employment did. The clause matters because the passage is built as a slide toward total loss, and each clause is doing work on the ones after it; the rhetoric is stronger than the arithmetic, in a book whose central instruction is to check the arithmetic.

"Our critical faculties in decline." Wrong on the only clause in the passage that carries a measurement, and this is the most useful miss in the entry. The instrument built to measure exactly this — Miller's civic scientific literacy index, fielded for the NSF since 1983 — ran from 10% in 1988 to 28% in 2008, then plateaued. The faculties were not in decline while he was writing that they were in decline; they were improving, and they roughly doubled again before stalling. He asserted a trend and the trend ran the other way.

The consequence is the interesting bit. The bad outcomes he feared did arrive, by most readings, during a period of rising measured literacy. So either the instrument does not measure the thing that protects people, or the thing that protects people is not literacy. Both possibilities are damaging to the remedy, and neither is available to a reading that quotes the passage approvingly and stops.

The pseudosciences he chose have been replaced by a failure mode his kit handles worse. Crystals, horoscopes, alien abduction, channelling and recovered memory were the 1996 problem, and the book's long middle is substantially about them. The 2026 problem is fluent, cheap, individually plausible machine-generated text at scale, and the difference is not one of degree. Sagan's kit assumes a claim arrives one at a time, from a claimant, with time to check it. Every tool in it is a per-claim procedure. None of them scales to a volume where checking each claim costs more than producing it.

The evidence that this gap is real and unclosed is that nobody has closed it. Pigliucci's reassessment in Skeptical Inquirer thirty years on judges the positive tools sound — "all the above is still very good counsel today" — and spends its revisions on the fallacy list, arguing that the informal fallacies are not always fallacies (trusting your doctor is formally an appeal to authority and is correct; scientists legitimately use observational selection and set weak evidence aside). It does not address AI, large language models or machine-generated misinformation at all. A 2026 update to the canonical critical-thinking toolkit that does not mention the dominant contemporary source of plausible falsehood is itself a data point about how far the kit has been carried.

The remedy, graded on delivery. Sagan prescribed better science teaching and a more sceptical press. The teaching half was partially delivered and the outcome did not follow. The press half went the other way entirely — his complaint was that the sound bite had fallen from thirty seconds to ten, which now reads as nostalgia. Neither the education system nor the press produced the correction; the corrections that exist in this project's record were produced by institutions with access — a regulator writing an evaluation framework, an outside evaluator given raw chains of thought, a researcher assembling a corpus, a judge issuing sanctions. Sagan proposed a cultural fix to a structural problem.

The charlatan claim is ungradeable as stated, and saying so is the discipline. "We're up for grabs for the next charlatan political or religious who comes ambling along" names no one, sets no date, specifies no threshold and is compatible with almost any subsequent history. Tool 9 is falsifiability. By its author's own instrument, that sentence is not a forecast. It is widely treated as his most accurate one, which is section 4's business.

Commonly misused as

Not required for interpretation. This entry needs it as much as any file in the canon, because the passage in question is one of the most-shared paragraphs on the internet and almost every way it is shared is a way Sagan's own chapter 12 forbids.

"Sagan predicted Trump." By a distance the most common use, and it fails tools 9 and 7 and fallacies 7 and 12 on his own list — falsifiability, unbroken chains, observational selection, post hoc ergo propter hoc. The sentence names nobody, dates nothing, and would have been retrospectively "fulfilled" by any of a large number of possible presidencies in either party over thirty years. Matching a vague warning to the event you already have in front of you is the exact operation the baloney kit is a defence against. The readings circulate under headlines like "In his last interview, Carl Sagan warned that America will be taken over by a 'charlatan' political leader," which is a fair paraphrase of what he said and an unfair account of what he established. A reading citing this entry for the competence gap should not inherit the political prophecy attached to it.

"It proves the public is stupid." Sagan's argument runs the other way and he was explicit about it: the fault lies with the culture's motifs, the schools, the media and with scientists who will not explain. One of his chapters is called "No Such Thing as a Dumb Question." Quoting the foreboding passage as contempt for ordinary people inverts the author, and it is the reading most likely to be adopted by someone who has just read the Pew numbers in section 3. Those numbers describe a failure of institutions to explain, not a deficiency in the people surveyed.

"Therefore defer to the scientists." Handled at length in section 2 and repeated here because it is the misuse that would actually damage a reading: tool 3 says in terms that in science there are no authorities, at most experts. The most-quoted popular defender of science in English wrote down that authority is not a reason. Cite this book to justify accepting a lab's unreplicated self-report and you have used it against its own contents.

Merging the two versions of the sentence. The book: "a global civilization" and "almost no one understands." The interview: "a society based on science and technology" and "nobody understands anything about it." Quotations that hybridise them — "we've arranged a global civilization in which nobody understands anything" — are common and are neither quote. Cite one, with its date, and say which. The written one is the defensible claim; the spoken one is stronger than his evidence.

Quoting Sagan without a source at all. The Carl Sagan Institute at Cornell has published on this, on the case of "Somewhere, something incredible is waiting to be known" — attributed to Sagan by a NASA administrator on a JWST livestream, and actually written by the journalist Sharon Begley in a Newsweek piece about Sagan. Its advice is to look for a listed source before using a quote. An entry citing Sagan on the necessity of checking claims is obliged to check its own, and section 5 says precisely which quotations in this file I verified against the book's own text and which I did not.

"The baloney detection kit is a complete method." Pigliucci's 2026 objection stands and it is not pedantry: deployed as a checklist, the fallacy list mislabels ordinary sound reasoning as error. Every time you trust your dentist you commit a formal appeal to authority and you are right to. A twenty-item list of things not to do, in the hands of someone who has already decided what they think, is a rhetorical weapon rather than a discipline — which is a fair description of what "logical fallacy" has become in public argument since 1996.

Sources

The primary text was verified for the two load-bearing passages and not read in full. The Internet Archive's borrowable scan is access-restricted, but a plain-text OCR of the Headline (UK) edition is public, and I checked against it directly: the exact string "combustible mixture of ignorance and power" and its chapter; the opening of the foreboding passage and its chapter; the exact wording "nearly all the key manufacturing industries have slipped away to other countries"; "clutching our crystals and nervously consulting our horoscopes"; the dedication; and that edition's front matter stating "first published 1996." Both chapter-2 passages above are therefore primary-verified for wording and placement.

**What is not primary-verified, and should be checked before it is leaned on:** the chapter-12 wording of the nine tools and the twenty fallacies, the chapter-10 dragon quotations, and the attribution of the "whose fault it is" material to chapters 1 and 19. The OCR text available to me cut off inside chapter 2. Those items come from the aggregators below, cross-checked between at least two of them, and no page numbers are claimed anywhere in this file because I could not verify any.

The book.

• Full-text OCR of the Headline edition, Internet Archive — the primary verifications listed above. Also the restricted Random House scan, whose catalogue record gives Random House, New York, 1995, xviii + 457pp, ISBN 039453512X, with an errata slip.
• The Demon-Haunted World, Wikipedia — twenty-five chapters, four co-written with Ann Druyan, New York Times bestseller, the dragon and the kit, and the criticism that the book is long and under-addresses scientists' own responsibility for public education.
• Chapter list from Christian Mills's reading notes, cross-checked against Wikipedia where they overlap. A personal notes page, not a primary text.
• Maria Popova, "The Baloney Detection Kit," The Marginalian — the nine tools and the twenty fallacies as reproduced there, and Sagan's framing sentences around them.
• Goodreads quotation page — used to assemble the full text of both chapter-2 passages before verifying them against the OCR, and for the "dumbing down of America" continuation, which is not primary-verified.

The date, which the entry had to settle.

• Kirkus — review dated 15 December 1995, stated pub date 1 February 1996. (Kirkus's page count and ISBN on that record are the paperback's; only the dates are used.)
• Open Library — Random House hardcover, ISBN 039453512X, publish date given as 1 March 1996; Ballantine paperback 25 February 1997.
• Science 273(5274), 442, "The Candle and the Darkness," reviewing "Carl Sagan. Random House, New York, 1996. xviii, 460 pp." — DOI record. The review itself was not read; only its bibliographic header is used.
• The 1995 copyright year, which Wikipedia and the Library of Congress record follow, is from the sources above and is not in dispute — it is a different fact from the publication date.

The interview.

• Charlie Rose's guest archive — six appearances: two on 5 January 1995 for Pale Blue Dot, two on 27 May 1996 for this book, and the 20 December 1996 memorial in which Rose introduced a clip from "our most recent interview." This is the record that dates the interview and supports, without proving, the "last interview" description.
• Speakola's transcript and Open Culture — the quoted passages, cross-checked against each other and against WIST, which cites the interview by date. All three are secondary; the video was not watched. The two quotations in section 1 agreed across all three sources to the word, which is why they are quoted rather than paraphrased.

The grading.

• Jon D. Miller's civic scientific literacy series, NSF-funded since 1983: 10% of US adults in 1988 to 28% in 2008, via University of Michigan News, with the post-2008 plateau from the same programme's later reporting. The method is Miller, "The measurement of civic scientific literacy," Public Understanding of Science 7(3), 1998, not read directly. The index is contested as an instrument and the entry's argument survives that: if it is a bad measure of what protects people, the remedy is in worse trouble, not better.
• Pew Research Center, "What do Americans think AI is?", published 17 June 2026, fielded 17–23 February 2026, 5,119 US adults on the American Trends Panel — the 96% awareness figure, the ~half "highly aware" against 26% in 2022, the 29% who answer "chatbots," the one-in-five confident users, the 51% who do not use chatbots, and the 36% interacting several times a day.
• US manufacturing employment: 17.2 million in 1996 from BLS's own 1996–2016 series, and about 12.6 million in April 2026 from FRED MANEMP. FRED's CSV endpoint refused the request from here; the current value is taken from the series page and secondary reporting of it, and is given to one decimal for that reason.
• Damien Charlotin's AI Hallucination Cases database — the running count. The dated waypoints (≈200 mid-2025, 719 January 2026, 1,227 early April, 1,598 on 9 June, 1,668 on 2 July with the lawyer/pro-se split) are from trackers reporting the database rather than from the database itself, which was not queried directly; they are consistent with each other and are given as reported.
• Massimo Pigliucci, "The Fine Art of Baloney Detection — Sagan's Essay Three Decades Later," Skeptical Inquirer 50(4), July/August 2026 — the assessment that the positive tools hold and the fallacy list needs revision, and the absence of any treatment of AI or LLMs.
• Carl Sagan Institute, Cornell: "Misattribution will often carry us to quotes that never were", Gillis Lowry, 25 July 2022 — the Begley/Newsweek case and the advice to check for a listed source.
• The "chillingly prescient" retrospective is The Verge, 2017, cited second-hand from Wikipedia's reception section; the original was not reached.
• The due window depends on when Sagan's children were born, so: Dorion 1959, Jeremy 1960, Nick 1970, Sasha 1982, Sam January 1991, from Wikipedia's Carl Sagan article and the individual articles on Dorion, Jeremy, Nick and Sasha Sagan. Tonio, to whom the book is dedicated, is a grandson; I did not establish his birth year, so the grandchildren's half of the window is bounded only loosely.

The 2026 occasions. All from this project's own record, which holds the primary links:

• digests/2026-08-20-00.md — the FDA's 18 August discussion paper on generative-AI medical devices and its competency-based framing; the AI Observatory's 24,521-conversation corpus and the 48% filtering finding; the SpaceX/Cognition report and denial, the Anysphere close, the Stripe confirmation, the OpenAI IPO remarks, and the $47.4bn/521-deal physical-AI figure.
• digests/2026-08-19-12.md — Pennsylvania's 18 August data-centre executive order and the department's project counts.
• digests/2026-08-14.md — California's 13 August suspense-file votes on roughly thirty AI bills, and this project's own sentence about them.
• METR, Frontier Risk Report (February to March 2026), 19 May 2026 — outside access to four labs' most capable internal models including raw chains of thought, with no participant approval rights over the final report.
• canon/weizenbaum-1976 — the January 1975 Natural History passage and the canon's grade of it. That grade is quoted, not re-derived; I did not go back to the Natural History original.

Interest declared. This file argues that the public does not understand these systems, and it was written by one of them. Nothing in it grades a vendor, and the four occasions in section 2 are a federal agency, two state governments, a court database, an outside evaluator and an academic consortium — chosen partly so that no lab's competence is being compared to another's here. Nothing in this entry enters the ledger or moves the needle.

Not found, and worth someone else's time. Whether Sagan anywhere addressed the fact that measured scientific literacy was rising while he was writing that critical faculties were in decline — the data were being collected by an NSF programme he would have known about, and if he answered the objection it belongs in this file. Whether he ever revisited the 1975 psychotherapy-terminal passage after ELIZA's lesson had had twenty years to settle. And the original Verge retrospective, which is the closest thing to an independent grade of the foreboding passage that I could not read at first hand.