The Doomsday Clock

Refiled from interpretation to idea, on the project's own definitions rather than on taste — the move einstein-1905 made in the other direction, and for the same reason. build.py defines interpretation as "a reading of events that a reading may need to name" and idea as "a framing later work is built out of rather than argued about. Cited when the thing being watched descends from it and the descent explains its shape."

Each January's setting is an interpretation, and proposals.md filing it there is defensible. But the entry the proposal asks for is not about any January. It is about the instrument: one pole, one scalar, one cadence, one board, units that changed twice in seventy-nine years. That is a designed thing, and the idea bucket is where this canon keeps designed things — bombe-bletchley-1940, hollerith-census-1890, google-index-1998, eliza-1966, tcp-ip-1983. The second half of the idea definition settles it: cited when the thing being watched descends from it and the descent explains its shape. The thing being watched, in this one case, is this project's own needle, and the descent explains its shape exactly.

prediction is the filing to argue against explicitly, because it is the one a careless session would reach for. A clock counting down to catastrophe looks like the largest forecast in the canon. It is not one, and the disqualification comes from the predictor: "The Doomsday Clock is not a forecasting tool, and we are not predicting the future. Rather, we study events that have already occurred and existing trends." There is no claim, no due date, and nothing to grade in the kurzweil-2005 sense. Section 3 grades it anyway, against the claim it does make.

This entry is written as a comparison, not as a history, because proposals.md instructed that and the instruction is right. The Doomsday Clock is the closest thing this project has to a methodological ancestor: a single scalar standing for a diffuse civilisational judgment, set by argument rather than measurement, published continuously since before anybody involved was born. Every place the two designs differ is a decision somebody made here, and most of those decisions were made without knowing that the alternative had been running for seventy-nine years.

And that produces a conflict of interest that runs the opposite way from asilomar-1975's, so it is declared here rather than managed. That entry had to avoid flattering a vendor. This one has to avoid flattering the design of the instrument writing it. A comparison authored from inside one of the two instruments will find, if nobody stops it, that the other one's choices were mistakes. So the discipline adopted below: every difference is recorded as a difference and not as an improvement, and where the Clock's choice is better this file says so. Section 2 concedes two — accountability and persistence — and they are the two that matter most.

descends_from is empty and the emptiness is a real finding. Nothing in canon/ precedes this. The honest ancestors are the Franck Report of 11 June 1945, the founding of the Bulletin in December 1945, and the Manhattan Project that produced both; none of the three is in canon/, and a search of proposals.md for trinity, manhattan, hiroshima, oppenheimer, szilard, franck and nuclear returns nothing. There is no id to use and I have not invented one. If the chain ever writes a franck-report-1945, this file's descends_from should gain it. einstein-1905 was considered and declined, and the reason is in section 1: the Bulletin's own masthead names Einstein a founder, but institutional credit is not intellectual descent, and the masthead's founding claim does not match the archival account of who actually founded the magazine.

One small pleasure, against a pattern this canon keeps repeating. asilomar-1975 noted that this canon files events under the year they became famous rather than the year they happened — Dartmouth 1956 for a 1955 proposal, Lighthill 1973 for a report signed in 1972, Asilomar 1975 for a moratorium declared in 1974. This id is exactly right. The Clock appeared on the June 1947 cover and the year in the id is the year of the thing itself.

What it is

The first value was a graphic designer's eye

The Bulletin of the Atomic Scientists began in December 1945 as a newsletter of the Atomic Scientists of Chicago — Manhattan Project physicists from the Metallurgical Laboratory who had already tried, in the Franck Report of June 1945, to argue the bomb should be demonstrated rather than dropped, and had lost. Eugene Rabinowitch, who wrote the bulk of that report, was its editor until he died in 1973.

For the June 1947 cover, the magazine went to Martyl Langsdorf, a landscape painter married to Alexander Langsdorf, one of the Met Lab physicists. She sketched on the back cover of a bound volume of Beethoven sonatas, began with a whole clock face, and stripped it down to the last fifteen minutes of the hour. Then she put the minute hand at seven minutes to midnight, and her account of why is the single most quoted sentence in the instrument's history:

> It seemed the right time on the page … it suited my eye.

The founding number of the world's best-known risk indicator was a composition decision. It expressed no judgment about 1947, was derived from no deliberation, and was not intended to be moved. What it did carry was two metaphors chosen with real care, and the Bulletin still explains them in the same terms in 2026: "the imagery of apocalypse (midnight) and the contemporary idiom of nuclear explosion (countdown to zero)."

It became a judgment retroactively. In 1949, after the first Soviet atomic test, Rabinowitch moved the hand to three minutes — and from that moment an aesthetic choice was the baseline against which every subsequent assessment of civilisational risk would be expressed. The anchor was arbitrary and the scale built on it was not. Anyone designing an instrument of this kind should sit with that for a moment before deciding it is only an anecdote.

Who sets it, and how

Until 1973 one man set it. Rabinowitch decided whether the hand moved, drawing on his own expertise and his correspondence with scientists worldwide. That is the fact most likely to surprise someone who assumes the Clock has always been a committee product: for the first twenty-six years of its seventy-nine — including 1949, 1953, 1960, 1963, 1968, 1969 and 1972, which is most of the Cold War's actual movement — it was one editor's judgment, published as a picture.

After his death the Science and Security Board took over. It meets twice a year, deliberations culminating each November, and the setting is announced in late January. The 2026 statement describes the arrangement in one sentence: the Clock "is set every year by the Bulletin's Science and Security Board in consultation with its Board of Sponsors, which includes eight Nobel laureates."

That laureate count is worth recording because it has changed and the change is never remarked on. The 2020 statement says thirteen. A 2018 Bulletin piece on the process says fifteen. The 2026 statement says eight. Three primary statements, three numbers, no explanation offered in any of them.

The 2026 board, read from the statement's own biographies section, is seventeen people. Alexandra Bell (president and CEO, ex officio), Edmund G Brown Jr (executive chair, and a former four-term Governor of California), Steve Fetter, Inez Fung, Asha M. George, Alexander Glaser, Daniel Holz (chair, and an astrophysicist), Jill Hruby, David Kuhlman (ex officio, and the statement notes explicitly that he "does not set the Doomsday Clock"), Robert Latiff, Melanie Mitchell, David Relman, Scott Sagan, Ambuj Sagar, Manpreet Sethi, Robert Socolow, Jon Wolfsthal. Nuclear physicists, arms-control officials, climate scientists, a microbiologist, a public-health security specialist, a retired Air Force major general.

Exactly one of the seventeen is an AI researcher: Melanie Mitchell, of the Santa Fe Institute, author of Artificial Intelligence: A Guide for Thinking Humans. Hold that against section 1's next fact — that the 2026 statement mentions artificial intelligence more often than it mentions any other technology except nuclear weapons — and you have the shape of the problem this canon exists to notice. It is not a criticism of Mitchell or of the board. It is the ordinary lag of an instrument built for one hazard being asked to carry another.

The scale, and what it is not

Midnight is the only pole. There is no defined good end. The Clock records improvement only as more distance from disaster, and the best it has ever managed — seventeen minutes, in 1991 — is not a name for anything. It is just further away. This is the property the project's own prior-art work found to be near-universal among civilisational scales: a bad pole and no good pole.

The proposal's phrasing needs one correction. proposals.md calls the Clock one-directional, glossed as "there is no way to record that things got better except as moving back from disaster." The gloss is right; the word is not. The hands move both ways and have moved away from midnight eight times. What the Clock lacks is not a direction, it is a pole — a named state at the other end that a reading could be about. The distinction matters when copying: a two-poled scale is not the same design with an extra arrow, it is a scale that has to argue both cases every time it is set, which is rule 3 here and is genuinely more work.

The units changed twice, both times for emphasis. Minutes until 2017, when the board first used a fraction — 2.5 minutes. Seconds from 2020, at 100 seconds, and the Bulletin's then-president Rachel Bronson explained the switch as a communicative act: "We are now expressing how close the world is to catastrophe in seconds – not hours, or even minutes." I looked in the 2020 statement for a methodological justification and did not find one. Then in 2025, having moved by a single second, the board had to explain what a second means:

> Because the world is already perilously close to the precipice, a move of even > a single second should be taken as an indication of extreme danger and an > unmistakable warning that every second of delay in reversing course increases > the probability of global disaster.

That is an institution conceding, in its own statement, that the increment is rhetoric. It is also the clearest external argument for this project's integers-only rule, which bin/ethos-check.py enforces by refusing to distribute a non-integer needle.

The record, counted

From the Bulletin's published timeline: twenty-seven resets in seventy-nine years, on top of the 1947 setting. 1949 (3 min), 1953 (2), 1960 (7), 1963 (12), 1968 (7), 1969 (10), 1972 (12), 1974 (9), 1980 (7), 1981 (4), 1984 (3), 1988 (6), 1990 (10), 1991 (17), 1995 (14), 1998 (9), 2002 (7), 2007 (5), 2010 (6), 2012 (5), 2015 (3), 2017 (2.5), 2018 (2), 2020 (100 sec), 2023 (90), 2025 (89), 2026 (85). Years not listed are years the board looked and left it alone.

Eight of those twenty-seven moved away from midnight: 1960, 1963, 1969, 1972, 1988, 1990, 1991, 2010. That confirms, against the primary timeline, the count the project's own methodology notes had asserted. The follow-on fact is the one to carry: the last move away from midnight was in 2010. Since then the board has reset the Clock eight times — 2012, 2015, 2017, 2018, 2020, 2023, 2025, 2026 — and every one of the eight moved it closer. Sixteen years, one direction.

The full published range runs from 17 minutes (1,020 seconds) in 1991 to 85 seconds in 2026. A factor of twelve, across the entire life of the instrument.

The scope kept widening

Nuclear weapons only, at first. Climate change entered the deliberations in 2007. "Disruptive technologies" — biotechnology, cyber, and eventually artificial intelligence — followed. By 2026 the masthead calls the Clock "a universally recognized indicator of the world's vulnerability to global catastrophe caused by man-made technologies", which is a much larger claim than a nuclear countdown.

The board says what the Clock is not, and says it plainly: "The Doomsday Clock is not a forecasting tool, and we are not predicting the future. Rather, we study events that have already occurred and existing trends." A 2018 Bulletin piece on the process compares it to reading lab tests and x-rays. Diagnosis, not prognosis. Every fair criticism of the Clock has to start by accepting that disclaimer, and most do not.

2026: 85 seconds, and what the board said about AI

Announced 27 January 2026 by Alexandra Bell; statement edited by John Mecklin; attributed to the Science and Security Board. "Because of this failure of leadership, the Bulletin of the Atomic Scientists Science and Security Board today sets the Doomsday Clock at 85 seconds to midnight, the closest it has ever been to catastrophe."

The reasoning names a nuclear arms race with New START expiring, a US "Golden Dome" missile-defence programme extending competition into space, climate policy going "from wholly insufficient to profoundly destructive", the laboratory synthesis of "mirror life" flagged by scientists from nine countries in December 2024, degraded US public-health infrastructure, and the rise of what it calls nationalistic autocracy. This file records those as the board's claims and does not adopt, grade or endorse any of them. They are contested political judgments about live actors, and a canon entry that quietly ratified them would be doing the thing rule 7 exists to prevent.

What is worth quoting at length, because it is the part a reading here would actually reach for, is the AI content. From the Disruptive technologies section:

> Artificial intelligence continues to be a significant and disruptive > technology. … The sophistication of large language models (LLMs) and their > applications in critical processes, coupled with lingering concerns about > their accuracy and tendency to "hallucinate," have generated significant > public debate about potential risks.

> The head of the US Strategic Command recently stated that, while a human will > always make the final decision on the use of nuclear weapons, it is > conceivable that AI will be embedded in decision-support systems used for > nuclear weapons. Even if a human is always in the decision loop, a too-heavy > dependence on black-box systems could present a serious danger.

> At the same time it is rushing to apply AI technology, the current US > administration revoked a previous executive order on AI safety, reflecting a > prioritization of AI innovation over safety and risk management that is > mirrored in other major powers.

And from the biological-threats section, the passage closest to this project's standing subject matter:

> Applications of large language models to biology and biological design tools > grew exponentially in power and capability over the past year. Developments > included more powerful AI tools that enable the design of new proteins—both > beneficial and harmful—and genome language models that can design novel > functional viruses. Because there are few safeguards against misuse of these > tools and because the element of surprise strongly favors malicious actors, > the risk of harm from misuse of AI tools has increased over the past year. > Unfortunately, there are few if any incentives for the private sector to > prioritize guardrails and AI safety measures, especially under current US > political leadership.

The board's list of actions that would pull the world back includes one on AI specifically: that the United States, Russia and China "can engage in bilateral and multilateral dialogue on meaningful guidelines regarding the incorporation of artificial intelligence in their militaries, particularly in nuclear command and control systems."

Why a reading would cite it

Start with the honest part: no reading has cited it yet. A search of every digest in digests/ and every event in history/events.jsonl for doomsday returns nothing. This entry therefore does not have asilomar-1975's warrant, which was three consecutive windows of live citation occasions. It has a different one, and the difference should be stated rather than papered over: the occasions below are two prospective and one that is already running — and the one already running is not in the readings at all. It is in the project's own design work.

Occasion one: the comparison, which the design documents already need

METHODOLOGY.md reaches for the Doomsday Clock seven times and ARCHITECTURE.md once, in every case as the standing negative example — the ratchet, the fake resolution, the missing good pole, the unanswerable "did AI move it". Those references were made from search summaries during a long design conversation. This file is the checked version, and three of the assertions come out confirmed, one comes out sharpened, and one comes out needing correction.

| | Doomsday Clock | this project | |---|---|---| | poles | one: midnight. No named good end | two, both named; the good pole appears to be genuinely unusual | | set by | 17-person board, 8 Nobel-laureate sponsors, deliberating twice a year | one headless model run | | cadence | annual, announced each January | twice daily, against a twelve-hour window | | unit | minutes, then seconds — changed twice for emphasis | a bare integer, decimals refused in code | | record | 27 resets in 79 years; no move away from midnight since 2010 | 9 readings in 6 days; all between -2 and -3 | | cause legible? | no — AI is named repeatedly and attributed nothing | per-lens digests, so in principle yes |

Four of those rows cut this project's way and two cut the Clock's, and the two are the important ones.

Accountability. Seventeen named people publish their biographies alongside the number, and a former Governor of California signs off on it. A model run at midnight publishes a number under a project name. The Clock's arrangement is straightforwardly more accountable, and no amount of append-only logging closes that gap. The most this project can claim is that its reasoning is published in full every time and the log cannot be edited — which is a real property, and a different one.

Persistence. The Clock has been published continuously for seventy-nine years through changes of editor, board, funding, subject matter and unit. This project is six days old and has nine readings. The single most valuable property of the instrument being compared against is the one this project has not yet earned at all, and any comparison that forgets that is a comparison written by the six-day-old.

Where the Clock's design fails, and the failure is structural rather than a lapse of judgment: it cannot tell you what moved it. The 2026 statement names AI more than a dozen times across four seconds of movement and attributes none of them. That is not the board being evasive; a single scalar has nowhere to put the attribution. The project's per-lens structure has somewhere to put it. Cite this when the question "how much of that was AI" comes up, because the Clock is the eighty-year demonstration that a lone number can never answer it.

Occasion two: the ratchet, aimed at this project's own log

Nine readings, all between -2 and -3. The Clock's record is what that looks like after seventy-nine years: sixteen years without a single move away from midnight, in a period that includes the end of one arms-control era and the beginning of a second Cold War and also includes the largest fall in global extreme poverty in recorded history.

The useful form of this citation is not "the Clock is biased, so we might be." It is narrower and harder to wriggle out of: an instrument shaped like this one has drifted one way for sixteen years, and no mechanism inside it noticed. That is a fact about instrument design, dated and sourced, and it belongs in front of any reading about to file another negative call. It argues for nothing about the world. It argues that a scale with one pole, a short window and no scoring channel will keep producing the same sign, and that the producer will experience each individual reading as obviously correct.

Occasion three: January, and the standing scalar

Every late January the Clock is announced and is covered everywhere for forty-eight hours. That is a reading window this project will meet at least annually for as long as it runs, and the entry gives it three things to do with it: report the setting as another instrument's judgment, note what the statement says about AI specifically, and refuse to treat any of it as evidence.

The same applies whenever a single scalar is offered as a summary of diffuse risk — a risk index, a p(doom) figure, a threat level, a safety rating on a lab's own model. The Clock is the case with the longest record and the fullest published reasoning, and the fair use of it is neither as a warning nor as a model but as the base rate for what these instruments do over decades.

The boundary, stated because it is easy to cross

The Clock's settings are judgments published by somebody else, and this project has a rule for that shape of thing: comments, not deposits. Concretely, and this is the part a later session might get wrong in good faith:

• No Clock value ever enters history/events.jsonl. It is not an event this project observed and not a reading this project took.
• No Clock value ever becomes a landmark in SCALE.md. Landmarks are needle values, a needle value is the owner's to set, and importing 1962's seven minutes or 1991's seventeen as a position on this scale would be a model supplying a needle by the back door. Rule 5 and jobs/canon.md both forbid it; this sentence exists so that nobody has to rediscover why.
• A reading may cite the Clock — the setting, the date, the board's stated reasoning — as context for something it found elsewhere. It may not treat the setting as evidence that risk rose or fell, because the setting is an opinion and this project already has one of those.

A note on where the occasions above come from. Section 2's ratchet argument refers to this project's own event log, and section 1's contrast refers to its own design documents. Both are the project's own output, both are named here solely to establish the citation occasion in the manner asilomar-1975 and dartmouth-1956 established theirs, and neither is evidence — here or anywhere.

What it got right, and what it got wrong

An idea does not require this section. It gets one because the whole value of the entry is the seventy-nine-year operating record, and a record carried ungraded is decoration.

Grade it against its own claim, not against the one critics assign it. The Clock does not claim to forecast. It claims to be "an indicator of the world's vulnerability to global catastrophe caused by man-made technologies" derived from "events that have already occurred and existing trends." So the question is not "did midnight arrive on schedule" — nothing about the design ever promised that. The question is whether the number tracked vulnerability.

Right, and the strongest case for the instrument: 1953 and 1991. Two minutes in 1953, after the American and Soviet hydrogen-bomb tests, is a defensible nomination for the most dangerous technical moment of the century — a thousandfold increase in yield in eighteen months, with no arms-control machinery of any kind in existence. Seventeen minutes in 1991, after START and the deep arsenal cuts, is the safest reading it ever gave and looks correct thirty-five years later. Both calls were made contemporaneously, in public, with reasons attached, by people who could have hedged and did not. That is the property this project is trying to copy and it should be credited plainly.

Wrong, and it is the failure that matters most: October 1962. The Cuban Missile Crisis is the closest approach to nuclear war anyone can document. The Clock stood at seven minutes to midnight throughout, and did not move. The Bulletin's own explanation, in its FAQ: "They were not moved during the 10-day crisis because too little was known at the time about the circumstances of the standoff or what the outcome would be." A second Bulletin account adds that the near-catastrophe had given Washington and Moscow fresh incentive toward arms control, and that the Clock captures trends rather than moments.

Both explanations are honest and both are beside the point, which is that the instrument's cadence made the miss inevitable. An annual number set by a board that meets twice a year cannot register a thirteen-day event, and the relevant issue of the magazine appeared after the danger had passed. Whatever the Clock measures, it demonstrably does not measure "how close we came", because the one time the world is known to have come very close, the hands did not move at all.

For this project the transfer is direct and cuts against complacency in both directions. A twelve-hour cadence would have caught October 1962. It would also have caught, and scored, forty other weeks that turned out to be nothing — which is the flow-versus-level problem, and is why this entry is a comparison rather than a verdict. Neither cadence is right. They fail differently, and knowing how each fails is the whole benefit of having an ancestor.

Wrong, structurally: the ratchet. Sixteen years and eight consecutive resets without a single move away from midnight. The generous reading is that the world genuinely got worse for sixteen straight years. The ungenerous one is that an instrument with one pole, a board that convenes to consider dangers, and an annual news cycle to feed has no mechanism that could ever produce a sustained retreat — the Clock has no scoring channel, nothing ever tells it that a past setting was wrong, and moving away from midnight is a story with no headline in it. Nothing available to me distinguishes those two readings, and the entry should not pretend otherwise. What can be said is that the design admits the second one and contains nothing that would catch it.

The self-grade, and it is not on accuracy. The Bulletin's own summary of the instrument is that it "has become a universally recognized indicator" — a claim about recognition. Defending the Clock against Steven Pinker in 2018, Rachel Bronson pointed to "the painstaking and serious process, culminating each November, that decides the time", to the Clock having "identified positive and negative developments in the evolution of existential threats", and to its having moved away from midnight in 2010 after New START. That is a defence on process, on balance, and on reach. It is not a defence on correspondence to risk, and as far as I can find the institution has never offered one.

The independent grades, and they differ from the self-grade on exactly that axis. Steven Pinker, who called it a "propaganda stunt" in Enlightenment Now and a "fear-mongering gimmick" on Twitter, makes three arguments: that a clock implies a finite time after which catastrophe is guaranteed, which misdescribes a non-zero annual probability; that the setting "doesn't track indicators of the actual risk" and instead reflects how much the Bulletin wants to alarm people; and that the 2007 admission of climate change turned a nuclear indicator into a general dread-generating device. Alex Wellerstein, a historian of nuclear weapons, makes the narrower and in some ways more damaging version: that the Clock is a subjective metaphor for current risk that was never meant to be a scientific instrument, and that "midnight" implies a hard stop that fits nuclear war and fits nothing else the Clock now covers.

Where the two grades actually conflict is on what the instrument is for. The institution grades itself on recognition and process; the critics grade it on correspondence to risk. Both are coherent and they do not meet. And the substitution the institution has made — success measured as reach — is precisely the one rule 7 was written here to forbid, in the sentence that says every shortcut which would spread the judgment faster works by making it worth less. The Doomsday Clock is the eighty-year worked example of an instrument that took the other option and is, by its own account, better known than ever.

Two things complicate the critics' case and belong in an honest entry. Pinker declared complete solidarity with the Bulletin's anti-nuclear purpose and framed his objection as tactical — how to motivate rather than paralyse — and proposed collaborating with psychologists. And the Bulletin published his critique itself, in its own pages, and Bronson's reply ended "Come join us, Steven Pinker. We welcome you to the discussion." On 17 December 2025 the Bulletin announced that Wellerstein had joined as a Senior Fellow. An institution that publishes its sharpest critics and then hires one of them is not behaving like a propaganda operation, and the entry should say so even though it slightly weakens the tidier story.

A negative result, recorded because absences are findings. I looked for any retrospective scoring of the Clock's settings against what followed and found none in seventy-nine years — no institutional review, no external audit, no published attempt to ask whether a 3-minute year differed from a 9-minute year in any measurable way. The one quantitative treatment I found is Sinclair and Silbersweig, "Apocalypse now? Mortality and mental health correlates of the Doomsday Clock", published on 12 March 2025 — in the Bulletin, the Clock's own publisher — correlating Clock settings against US mortality and mental-health series. I could not read the body (premium) and cannot assess it. What can be said is the structural fact: the only quantitative study of the instrument that I could locate appeared in the instrument's own magazine, and nobody has ever graded the settings against outcomes.

Right, and the thing this project cannot yet claim: it lasted. Seventy-nine years, twenty-seven resets, through the death of the editor who set it alone, the addition of two entire hazard classes, the collapse and reconstitution of its subject matter, and a change of unit. Every January it is read out and argued with. The reasoning is published in full each time. Whatever the number is worth, the practice of publishing a dated judgment with its reasoning attached, without interruption, for eight decades, is the achievement here, and it is the one worth copying most.

Commonly misused as

The limit kind requires this section and an idea does not. It is here for dartmouth-1956's reason: this entry will be misused more often than it is used, and the compressed version in circulation reverses its meaning.

"Scientists say we are 85 seconds from nuclear war." The most common form and the one the Bulletin disowns in its own FAQ. It is not a countdown, not a probability, not an estimate of time remaining, and not a forecast. It is a board's summary judgment of current vulnerability, expressed in a metaphor chosen in 1947 because a countdown was the idiom of the day. The seconds are not seconds of anything.

"It is the closest ever, so risk is at an all-time high." The comparison is not available, and this is the misuse most likely to be committed by this project — because it is exactly the move a chart invites. The 2026 setting and the 1953 setting are not two values of one variable. Between them the scope changed twice (climate in 2007, disruptive technologies after), the unit changed twice, the setter changed from one editor to a seventeen-person board, and the board's membership turned over entirely. The scale was re-anchored repeatedly and never re-based, which is the Kardashev failure in another costume: Earth was Type I in 1964 by the definition of the man who coined it, and "85 seconds is the worst ever" is the same sentence. What is comparable is a setting against its immediate neighbours. What is not comparable is 1953 against 2026.

"AI moved the Clock four seconds." Not knowable, by anyone, including the board. The 2026 statement raises AI throughout and assigns it no share of the movement, alongside nuclear arms racing, New START's expiry, Golden Dome, climate policy, mirror life, degraded public health and nationalistic autocracy. Any sentence attributing a portion of the four seconds to any one of those is invented. This is the structural point from section 2 arriving as a concrete error, and a reading that repeats it has fabricated an attribution that the source deliberately does not make.

Using the settings as a measured variable. The Clock is a sequence of numbers with dates, which makes it regressable, and regressing it is not the same as it measuring anything. The authors of the 2025 correlational study frame their variable as a socio-political indicator, which is a defensible thing to do; the misuse is downstream, when the correlation gets read as evidence that the Clock tracks something real. A board's rhetorical choice of increment — one second in 2025, four in 2026 — is not a quantity, and treating a judgment as an instrument reading is the failure model-laundering-2008 is proposed to cover and goodharts-law-1975 already covers from the other side.

"Asilomar for AI" energy applied to the Clock: 'we need a Doomsday Clock for AI'." Several exist and none has the property that makes this one work. What the Clock has is not a metaphor; it is seventy-nine years of continuous publication by an institution that survives its individual members, with the reasoning printed in full every time. That is an endurance property, not a design property, and it cannot be launched. A new clock on day one has the arbitrary anchor and none of the record.

And the inverse misuse, which this canon should be equally hard on: "it is a propaganda stunt, therefore worthless." The strong critical reading is stated above at full strength. It also has to survive the fact that the Clock publishes a dated judgment with its full reasoning attached every single year, names the people responsible, prints their biographies, has moved away from midnight eight times when it could have ratcheted uninterrupted, publishes its critics in its own pages, and hired one of them. Almost no forecaster, index or commentator offers a comparable record to argue with. The honest position holds both halves: the number is arbitrary and the practice is exemplary, and this project is trying to inherit the second without the first.

Sources

Read directly, as rendered pages.

• **"It is now 85 seconds to midnight," 2026 Doomsday Clock Statement, Science and Security Board, Bulletin of the Atomic Scientists, Editor John Mecklin, 27 January 2026** — the PDF at thebulletin.org/wp-content/uploads/2026/01/2026-Doomsday-Clock-Statement.pdf, read as page images (pages 1–4 and 10–16). Source for every block quotation of the 2026 statement above: the masthead founding line and the "universally recognized indicator" sentence; "eight Nobel laureates"; the 85-seconds setting and "the closest it has ever been to catastrophe"; the entire Disruptive technologies section quoted in section 1; the biological-design-tools passage; the AI item in the list of actions; "Our current trajectory is unsustainable"; the mirror-life and Golden Dome material; and the full Science and Security Board roster with biographies, from which the seventeen names, the two ex officio members, the note that Kuhlman "does not set the Doomsday Clock", and Melanie Mitchell's presence as the board's one AI researcher are taken. Limits: I read pages 1–4 and 10–16 of a longer document. The nuclear and climate sections (roughly pages 5–9) I did not read, so the characterisation of the board's nuclear and climate reasoning above rests on pages 2–4 and on the press release, not on those sections.
• PRESS RELEASE: "It is 85 seconds to midnight," 27 January 2026 (thebulletin.org/2026/01/press-release-it-is-85-seconds-to-midnight) — Alexandra Bell as announcer, Daniel Holz's quotation on autocracies, and the AI-guidelines ask.

Fetched and used, primary but through a summarising fetch. For each of these I can vouch for the quoted sentences and not for having seen every sentence in the document.

• 2026 Doomsday Clock Statement, web version (thebulletin.org/doomsday-clock/2026-statement) and its Disruptive technologies section — cross-checked against the PDF, which is why the quotations above are taken from the PDF.
• 2025 Doomsday Clock Statement, "Closer than ever: It is now 89 seconds to midnight," 28 January 2025 (thebulletin.org/doomsday-clock/2025-statement) — source for the "a move of even a single second" passage and for the 2025 AI sentences on bioweapon design, AI in military targeting in Ukraine and the Middle East, and AI-assisted disinformation.
• Doomsday Clock Timeline (thebulletin.org/doomsday-clock/timeline) — every setting and one-line reason listed in section 1, Martyl Langsdorf's authorship and her "it suited my eye", and her death in 2013. The counts in section 1 — twenty-seven resets, eight moves away from midnight, no move away since 2010 — are mine, derived from this page, not figures the Bulletin publishes. They confirm the count already asserted in the project's own methodology notes, which had it from search summaries.
• Doomsday Clock FAQ (thebulletin.org/doomsday-clock/faq) — "not a forecasting tool"; the Science and Security Board meeting twice a year; climate entering the deliberations in 2007; and the Cuban Missile Crisis answer quoted in section 3.
• "Doomsday Clockwork," January 2018 (thebulletin.org/2018/01/doomsday-clockwork) — Rabinowitch deciding alone until his death in 1973, the board taking over, deliberations culminating each November, the lab-tests-and-x-rays comparison, and the "fifteen Nobel laureates" figure.
• 2020 Doomsday Clock Statement (thebulletin.org/doomsday-clock/2020-doomsday-clock-statement) — the "13 Nobel laureates" figure and the 100-seconds setting. My finding that it contains no methodological justification for the switch from minutes to seconds is a negative result from a targeted query against a summarising fetch, which is weaker than a full read; Bronson's "in seconds – not hours, or even minutes" comes from the 2020 press-release coverage rather than from the statement.
• 2018 Doomsday Clock Statement (thebulletin.org/doomsday-clock/2018-doomsday-clock-statement) — autonomous weaponry making "kill" decisions without human supervision, and the emerging-technologies framing. Used to establish that the disruptive-technology material predates the recent AI passages; I did not establish which statement first used the words "artificial intelligence", and the entry makes no claim about that.
• Steven Pinker, "Real risks, undeniable progress," 11 April 2018 (thebulletin.org/2018/04/steven-pinker-real-risks-undeniable-progress) and Rachel Bronson, "Welcome to the discussion, Professor Pinker," April 2018 (thebulletin.org/2018/04/welcome-to-the-discussion-professor-pinker) — the exchange in section 3. Pinker's "propaganda stunt" (from Enlightenment Now) and "fear-mongering gimmick" (from Twitter) reach this file through the Bulletin's page and a search summary, not from the book or the post, and should be cited to those originals rather than to this entry.
• "Alex Wellerstein joins the Bulletin," 17 December 2025 (thebulletin.org/2025/12/alex-wellerstein-joins-the-bulletin) — his appointment as Senior Fellow and its date.
• "How the Bulletin of the Atomic Scientists got its start" (thebulletin.org/virtual-tour/how-the-bulletin-of-the-atomic-scientists-got-its-start) — the December 1945 founding by "a group of Chicago scientists who helped build the world's first atomic bomb". Note the discrepancy recorded in the header section: this page names no founders at all, while the 2026 statement's masthead credits "Albert Einstein, J. Robert Oppenheimer, and University of Chicago scientists". Two of the Bulletin's own surfaces, two different founding accounts.

Relied on at one remove, and not read. Anyone extending this entry should start here.

• Alex Wellerstein, "Outdated Doomsday Clock struggles with new scary threats," Press of Atlantic City opinion, reported as March 2023. The URL redirects to a paywall proxy and I could not open it. His two positions attributed above — that "midnight" implies a finality that fits nuclear war and fits nothing else the Clock now covers, and that the Clock is a subjective metaphor never meant to be a scientific instrument — are from search summaries and are not verified against the article. He is the most qualified living critic of this instrument and this file is thinner for not having him in full.
• **Samuel Justin Sinclair and David A. Silbersweig, "Apocalypse now? Mortality and mental health correlates of the Doomsday Clock," Bulletin of the Atomic Scientists, 12 March 2025 (thebulletin.org/premium/2025-03/apocalypse-now-mortality-and-mental-health-correlates-of-the-doomsday-clock). Authors, venue and date verified; the body is premium and unread.** The method described in section 4 (Spearman rank-order correlations with a Bonferroni correction, against US mortality and mental-health series) is from a press summary. My characterisation of the authors' framing is correspondingly provisional and should be checked before it is repeated.
• The archival account of the Bulletin's founding — Rabinowitch, Hyman Goldsmith and John Simpson Jr at the University of Chicago, Simpson as first chairman, Rabinowitch as editor until 1973 — is from search summaries of the University of Chicago Library's finding aids and the Nuclear Museum, not from the archives themselves. It is used above only to establish that the Bulletin's own masthead line and its own history page disagree, which is visible from the two Bulletin pages alone.
• The 1962, 1984 and 1991 statements themselves. Everything above about those years comes from the timeline's one-line summaries and the FAQ, not from the contemporaneous texts. In particular, the observation that the Clock's 1984 setting coincides with the Able Archer exercise of November 1983 was considered and left out, because I could not establish what the board knew or said at the time and an entry that implied foresight it cannot document would be doing the thing section 4 warns against.
• archive.org and the June 1947 issue. I did not view the original cover. Langsdorf's account of the design — the Beethoven sonatas, the full clock face reduced to fifteen minutes — is from the Bulletin's own virtual-tour material via search summary.

On the boundary. This file names einstein-1905, asilomar-1975, dartmouth-1956, bombe-bletchley-1940, hollerith-census-1890, google-index-1998, eliza-1966, tcp-ip-1983, kurzweil-2005 and goodharts-law-1975 as entries that exist in canon/, and franck-report-1945 and model-laundering-2008 as ids that are unwritten — the first not proposed at all, the second proposed and not yet researched — neither invented into the header. It refers to history/events.jsonl, to the digests, and to METHODOLOGY.md and ARCHITECTURE.md solely to establish the citation occasion in section 2, in the manner asilomar-1975 established its own; all four are the project's own output and are not evidence, here or anywhere. It records the 2026 board's political claims as the board's and grades none of them. It deposits nothing in history/, touches no other entry, and places no needle, no score and no landmark — and section 2 says explicitly why the Clock's own values must never become either.