TL;DR: ArgMap is a tool I created to make arguments explorable and computable, with every step directly linkable. Its first map is of If Anyone Builds It, Everyone Dies: I mapped the logical structure of the book's argument with LLM help. LLM readers derive confidence numbers from the authors' words through a rubric that was fixed in advance. The computation tests how well the authors' stated confidence fits their argument as I mapped it, rather than reflecting my personal view of the book. Out of 67 claims that the book both rates and argues for, 66 land within 10 points of the authors' range; the exception is the claim that everyone dies, at 0.62 against the authors' range 0.75 to 0.95. The policy conclusion for a global shutdown is at 0.93. You can adjust any probability on the map and see the conclusions change and tensions light up. I'd like you to find a step you disagree with or a toy example where the computation is wrong.
Do I contradict myself?
Very well then I contradict myself,
(I am large, I contain multitudes.)
Walt Whitman, "Song of Myself"
In the art of Kintsugi the broken pieces that no longer quite fit together are joined back into a whole, and the breakage and repair in gold become part of the piece.
In If Anyone Builds It, Everyone Dies (henceforth IABIED), the word "everyone" is meant literally: everyone, everywhere on Earth, will die, killed by an artificial superintelligence (ASI) soon after one is built. The authors add: "We do not mean that as hyperbole." Many people have stated a probability for that outcome, but few explain how they arrived at it. By contrast, IABIED lays out the argument in great detail, and confidence is communicated as natural language or even omitted.
With the numbers read from the authors' natural language through the fixed rubric, the map reports a 62% probability that an ASI is built and kills everyone.[1]
What the question mark here asks is whether the book's argument, as the authors present it, is mapped correctly and whether the computation evaluates it correctly. Whether the premises of that argument are true is a separate question that the map helps the reader to investigate themselves.[2]
The authors deliberately don't state a number for the probability of the extinction claim , because it depends on how society decides to act, and assigning a single probability "strikes us as defeatist".[3]
This probability and are sometimes even conflated with each other.[4]
This is what the map implies where society does not change course.
And this is what needs to be done.
Note that these numbers are not the authors'. For 67 claims the book both states a confidence in words and argues for them in more detail. The map also computes their values from the claims beneath them, without using the reading of the book's confidence at all. Those readings are then checked against the map's computed values. Of the 67, 18 land outside of the authors' range by more than one point (14 above the range, 4 below), and only one by more than 10 percentage points.
That one is also the most important: the extinction claim that everyone dies where the authors' statement is "The default outcome is lethal, but the situation is not hopeless; machine superintelligence doesn’t exist yet, and its creation can still be prevented", reading 0.75 to 0.95 through the rubric, against the map's 0.62.
For the book's central conditional the gap is larger still: the authors' "...then everyone, everywhere on Earth, will die. We do not mean that as hyperbole." reads 0.93 through the rubric against 0.74, the map's probability that everyone dies given an ASI is built.[5] Part of the gaps comes from choices in ArgMap that are not settled yet, see What is not settled.
Everything you see in the map below is LLM-extracted under my steering: the logical structure of the argument, the nesting into the different levels of detail, and the numbers. The numbers were elicited by LLM readers from the book and its supplements' words. The map has not been reviewed by the authors.
The map
This is the map of IABIED. See the custom widget below, or go to the full web app, which has more features and readouts.
It's interactive, so let's go through an example interaction: let's say you're not convinced that artificial superintelligence is possible and that's why not everybody will be killed by an ASI.
Click the "+" symbol on "Anyone builds an artificial superintelligence (ASI)" to unfold its sub-nodes. Drag the probability gauge on the left edge of "A machine superintelligence can be built" down to around 50%. "What-if mode" shows that the conclusion "Everyone, everywhere on Earth, will die" falls from 0.62 to 0.39. Does the extinction claim now read your own personal number? Or are there other underlying disagreements?
The extinction claim's gap on the map lights up in gold, the larger the gap the deeper. That is the golden glue of the Kintsugi bowl.
Long arguments don't fit in working memory. We're damned to local reasoning and it's difficult to integrate the pieces into a coherent whole. The pieces of local reasoning we produce may contradict each other, and they cannot always be glued back together into one beautiful piece of porcelain. Upon inspection, we find that because one of its many fine joints is failing, the piece cannot stand on its own.
It would be much nicer to be able to explore and dive into an argument at the depth you desire, reason locally there, and see the conclusions change across the entire argument map. Conflicts would light up if you assign some credence that clashes with another part of the argument. The system would then calculate a closest consistent picture of the argument, and do that loudly by showing exactly where the glue went and where some parts were off. The rule that decides how much glue to put where is one choice among several; the main point is to highlight conflicts.
Then once you have such a system, what you get for basically free is that you can play with it: modify the argument, add your own credences, mark where you disagree, and see how the result changes.
So if the argument is mapped in enough detail, down to a level where the links are mechanical, and if the map is complete enough to capture the argument structure you want to engage with, then this becomes a calculation over scenarios that you can argue with, and every number in it is traceable back to an input you can see and change.
The book's claim, and where the numbers come from
In its introduction, IABIED makes the claim: "If any company or group, anywhere on the planet, builds an artificial superintelligence using anything remotely like current techniques, based on anything remotely like the present understanding of AI, then everyone, everywhere on Earth, will die. We do not mean that as hyperbole." (Hard Calls and Easy Calls)
Most of the remainder of the book then lays out the argument for this in more depth. Almost none of the statements of the book carry probability numbers or credences. But the authors did make careful[6] use of confidence language throughout the book. Wherever they do, the number is read from it. Where they don't, which is mostly for the common objections to the argument which the book quotes, a documented default is used.[7] The rubric was fixed in advance and is published here. Several readings were revised later, among them the extinction claim's.[8] Every revision is documented with its reason in the map's source.
In the places where the book makes a credence claim about a statement or reasoning link that is not further argued in more detail, that credence becomes an input to the map. Where the book makes a credence claim about a statement or reasoning link that is further argued in more detail, the credence becomes a check: it never enters the calculation and is only used as a readout on how well the map's finer version delivers the authors' number on the coarse claim. A gap here can mean that the book or the map is off; and if the map is right, at least one of the authors' numbers must be off.
Such a gap shows through one of the following rules, each further explained in its linked section:
Pulling the map's number down:
a link with several necessary premises at once has them taken as independent unless the authors claim otherwise (see item 1 under What is not settled)
objections keep the share the responses leave them, even when outvoted (see Arguments do not vote)
Pushing the map's number up:
multiple supporting links the map takes as independent accumulate (see Arguments do not vote)
a link's number acts as a floor in the cases where it applies, and the rest are filled at even odds (see Where the probability goes missing)
In either case: the rubric sets 0.90 for links the authors argue without hedging (moving it by 0.05 in either direction is worth 8 to 10 points on the extinction claim; see item 4 under What is not settled)
The largest gaps are shortfalls.
How ArgMap computes the joint probability distribution
The authors' input numbers are used in the calculation as soft constraints, so a contradictory map can still get a distribution, and the disagreement that is left over is the glue that lights up in gold.
The reference distribution in an ArgMap is a Bayesian network derived from the logical structure of the map, with the cases removed in which a link for a claim and a link against it would both apply (i.e. where their premises hold; see also Reasoning links compete only where their grounds hold together). It fills probability mass at even odds where nothing is said by the author unless the rest of the map moves it. The result is the distribution closest to that network by minimizing KL divergence, meeting the constraints as well as possible, each weighted by how firmly it is held. With hard constraints this would be an I-projection.
The objective:
with the solved distribution, the reference network, the authors' input numbers, their values under and their counts in coin flips, see the mathematics page 4.2.
ArgMap builds a factor graph over the claims and reasoning links in the map, and solves this objective with a junction tree that runs in the browser.
What is not settled
Some parts of this framework are not settled yet. The first costs the extinction claim 3 to 9 points. The fourth about 8 to 10 points either way. The others don't move the numbers at all.
Should the premises of a reasoning link be independent by default? Everything that the authors don't state as correlated through a shared premise is multiplied. This contributes to the main conclusion "Everyone, everywhere on Earth, will die" falling short of the authors' check.
On this map, full dependence between "The ASI that gets built is misaligned" and "A misaligned ASI causes human extinction" would recover 3 points of the extinction claim, and with the latter's four grounds also fully coupled, a total of 9 points with the extinction claim at 0.71. Note that even with both claims set to 100%, the extinction claim reads 0.82, because "Anyone builds an ASI" stays at 0.84 and still bounds it.
Can there be contradictions between the authors' numbers that don't light anything up? Currently the map only lights up when there is a conflict where a value ends up outside of the authors' range. But there can be conflicts where the computation resolves conflicts while staying in the authors' range, see Reasoning links compete only where their grounds hold together. A full report that tells an author where their numbers contradict is not built yet.
Are the two parameters credence and kind sufficient degrees of freedom for interacting with an argument under contradiction? The parameters of claims and links are credence and kind (see How firmly each number holds). And every disagreement gets mended in a mechanical way only. There will always be examples where this way is wrong, because what makes the difference is the content of the statements and links that's only in the label and that no parameter carries. So the question is whether credence and kind are enough for interacting with an argument under contradiction or more should be added.
What else does the map assume where the authors say nothing? Conventions:
the 137 links the authors argue without hedging are set at 0.90, and moving that 0.90 by 0.05 moves the extinction claim down to 0.54 or up to 0.72
a link's number holds at uniform strength wherever its premises are true, no matter what else is true or false there
the direction a link is written in (positive or contrapositive) can change the answer.
What number should the folded link show? Because the computation always operates on the fully expanded map, folded links need to get a derived number. It is currently calculated after the computation from the composition of values of the submap spanned by that coarse link's premises and conclusions. Since the format allows one to nest arbitrary refinements under a reasoning link, additional premises under the link's unfold may pull at the composed value. In the IABIED map, the main implication ("Anyone builds an artificial superintelligence (ASI)" -> "Everyone, everywhere on Earth, will die") unfolds into a finer link that further requires ASI misalignment and the ASI being able to cause human extinction, which are themselves nested.
Here's an example of the last item. The coarse 90% link "a storm cuts the town's power" contains two steps, one of them with a premise that no reasoning link feeds into. The link shows 0.36 folded, and 0.72 with the premise set to 100%, against the authors' coarse 90%. On the IABIED map, 0.93 and 0.67 are the same two kinds of numbers for the "If anyone builds it, everyone dies" link.
The ask
Find a step or claim you disagree with and let me know with the "💬 give feedback on this" button in the selected node's popup, or with the "💬 Feedback" button in the full app's toolbar, also anonymously.
Write a small toy example in the editor (full app -> advanced mode -> editor) that shows where the computed number is wrong. The "semantics audit" example group in the map picker shows what an example could look like. Confirmed errors get fixed and credited by name.
And if you find yourself in some discussion about AI risk that is covered by the book, you can link the exact step: "copy link to this" on a node's popup will load the map as you unfolded it and focus that node, and keeps your dragged what-if numbers.
Appendix
Build date
The numbers in this post follow the map's build from 2026-10-15. The maps embedded here follow the live site, so there may be some drift. Changes are documented in the maps' sources with their reasons.
Where the probability goes missing
Reasoning links carry logical structure and only strengthen the conclusion or negated conclusion where they apply. That means that for each claim, in some percentage of cases none of the reasoning links into this claim apply. These cases count as even odds.
On a claim card, the paler top of a gauge is that half share.
In the first example, we have rain at 80% and the link "rain -> street wet" at 90%. The street reads 0.86 (0.8 × 0.95 + 0.2 × one half: the link's 90% is a floor, so where it rains, at least 90 of 100 cases are wet, the rest at even odds; without rain, nothing is said and these cases also count at even odds). Drag the rain down and the paler strip widens. Drag it to 0% and the gauge goes grey (a claim with a single support at zero and nothing arguing against it just shows a grey stripe).
The second example has the inverse link "no rain -> dry" at 90% and the street is 0.77 at rest. The same drag of rain down to 0% puts the street down to 0.05.
On the IABIED map, there is also an example with no inverse link on purpose: "Enormous economic and strategic incentives drive the race", since an ASI could even be built under other motives. Drag it and "What physics permits, someone eventually builds" to 0 and "Anyone builds an ASI" still lands at 0.45 with the even-odds fill.
Arguments do not vote
Let's look at another example to see how multiple reasons for and against a claim interact and why support and objection on the same case don't touch each other if possible.
Below you see two toy examples arguing for a product launch. The first has one support at 85% against one objection at 15%. The second has five supports at 85% against one objection at 15%. In both the launch is at 0.85.
If you drag the objection to 0%, the claim absent the objection reads 0.925 with one support (1 - (1 - 0.85) × half), and 1.00 with five (1 - (1 - 0.85)^5 × half).
(Drag four of five supports to 0% against the objection at 15% and nothing changes, still 0.85.)
Humans often learn well through stories, where long lists of reasons impress themselves into the listener's memory merely through the length of their serialization into language, and where independence structures are rarely more complicated than simple chains or fans. If this example was told as one, it could be very convincing: five inputs all counted as independent 85:15 evidence multiply to (85/15)^5 = about 5,800 to 1. It would then be easy to overlook the single objection (which if multiplied would only multiply 15/85 to cancel one of the five supports, still very convincing at 1,030 to 1).
To fix this, in ArgMap, objections can be beaten by a response, a lower number, or doubt about their premises. So as long as the numbers fit, a multiplicity of supports outvoting them is not enough.
And if there are many supports against many objections, human reasoning about all of them breaks down.
Reasoning links compete only where their grounds hold together
Alice laughed. "There's no use trying," she said: "one can't believe impossible things."
"I daresay you haven't had much practice," said the Queen. "When I was your age, I always did it for half-an-hour a day. Why, sometimes I've believed as many as six impossible things before breakfast."
Lewis Carroll, Through the Looking-Glass
What happens when two (or more) links on the same premise claim more than all of a conclusion's cases?
In this example we see two diary entries that promise too much, 90% for Monday and 70% against Monday, given that the diary is reliable. The diary's reliability itself nobody has vouched for, in the non-conflicting case it sits filled at even odds.
In ArgMap, a for- and against-link resting on the same grounds never apply together and the excess 0.60 (90% + 70% = 160%) cannot happen. The cases producing the contradiction become rarer, and the doubt lands on the diary's reliability that is free to give. It falls from 0.50 to 0.21 and Monday reads 0.55.
Now drag the diary to 100%. It holds and the two links bend instead until they fit, with no excess left. Monday now reads 0.72.
Contrapositively, this gives rise to another authoring rule: If you want to author two links with contradicting conclusions and you want them to not compete with each other, you must put them on different grounds that cannot hold together: by explicitly naming the variable that distinguishes the two.
How firmly each number holds
Under disagreement of the authored numbers something has to give. This was already visible in the diary example above when the diary's reliability was held high.
What gives how much is determined by an additional parameter to links and claims.
Each reasoning link has a kind that is also read from the authors' language: formal, deductive, mechanism, empirical, testimony, analogy, hope.
The currency of these kinds is the number of coin flips at which their link is held when you drag something: formal never bends; deductive 1000; mechanism 64; empirical 16 or an optional sample size; testimony 16; analogy 4; hope 4; default 16. A number of 4 coin flips gives easily and 1000 moves little. The number held at n coin flips resists a change like n observations at that rate would.
Claims with an authors' range give depending on the width of that range: 18 flips for a range of 0.85 to 0.95, 8 for 0.7 to 0.9, 198 for a point interval. More specifically, claims with a width of four points or more give before any mechanism or deductive link.
Prior art
There are already similar systems, each lacking something an author needs:
Markov logic networks have weights instead of probabilities, and the joint distribution is normalized by the partition function over the whole model. A weight's effect depends on every other one and that's why they lack the ability to be elicited locally. ArgMap takes probabilities as inputs and solves for the factor weights.
Probabilistic conditional logic under maximum entropy (Paris 1994; Kern-Isberner 1998, 2001) is a close relative, has authored numbers as a single conditional, solved with a similar projection, but with hard constraints, which needs a consistent knowledge base. Later work (Potyka and Thimm 2014, 2015) handles inconsistency by first minimizing the violation and then choosing among the distributions at the minima, whereas ArgMap has a single weighted objective. In ArgMap, a number's meaning differs (a lower bound on the conditional, forcing some mass on the worlds where the conclusion holds given the premises, and leaving the remainder to be filled; not an exact conditional), it starts from the Bayesian network induced by the map, and how firmly each number holds is elicited from the source. This has an author who is locally honest and globally inconsistent as a legitimate end state, without hiding the glue.
MTAIR (Modeling Transformative AI Risks) is a quantitative model based on numbers from modellers through literature review and engagement with experts. ArgMap reads the numbers from an author's words.
AMTAIR (Automating Transformative AI Risk Modeling, 2025) has LLMs extract argument structure from AI safety papers into an ArgDown-like syntax, probabilities into a "BayesDown" syntax, then a Bayesian network with CPTs and prediction markets for live updates. The numbers come from the literature via the LLM and from markets, where ArgMap reads them from an author's confidence language. The output is a Bayesian network, where ArgMap uses for/against links with argument nesting.
ArgMap is very different from other types of argument mapping like Argdown, Argunet, Kialo (these are almost purely structural) or Concludia (non-probabilistic), and wiki approaches like Arbital (explorable explanations) with its "lenses" as tabs explaining a page at several depths and where users could cast probability votes on claims but which weren't further processed.
The conjunction and differ by here. This difference is the map filling the cases where nobody builds an ASI with anything like current methods and everyone is killed by one anyway, for example by an ASI built elsewhere or with other methods. The map's link against the claim ("if nobody builds it, it kills nobody", read at 0.9 from the book's words) leaves roughly one in ten cases open, and these cases count at even odds, which contributes (1-0.835) × 0.038 = 0.006.
Compare Aumann's agreement theorem, where two people with the same prior whose posteriors are common knowledge will arrive at the same posterior. Aumann names "errors in calculating posteriors" as one reason for lasting disagreement. In ArgMap the two are the authors' stated confidence for a claim and the map's computation based on the authors' own numbers in the detailed reasoning. They should agree, and a gap means that one of the calculations is off. But the theorem says nothing about the agreed-upon posterior being right; the two could still be wrong together. So whether the authors' premises are true is for the reader to judge (and adjust in What-if mode!).
From IABIED's online supplement, introduction chapter, "AI Experts on Catastrophe Scenarios": "We have concerns with the practice of trying to assign a “p(doom).” Assigning a single probability — as opposed to multiple probabilities that each assume a different response society could choose — strikes us as defeatist."
The book itself draws this distinction in chapter 12 on Toby Ord's 10%: "if you look into the details, Ord says the reason he estimates 'only' a 10 percent chance ... is because he expects humanity to come to its senses".
The map shows 0.67 on the folded link, which is the strength the finer steps multiply to, a different readout from . Which to show there is open, see What is not settled.
Chapter 11, endnote 10: "While we authors think that disaster is a strong default outcome, five nines is a ridiculous level of confidence that neither of us endorse."
Out of the 115 objection links here (excluding nine inverses that argue a negated conclusion from its negated premise), 81 remain at the rubric's 0.15, or 0.20 where the book engages with the objection at length or grants some part of it (counted 2026-10-12).
The extinction claim's reading changed from 0.88 to 0.98 down to 0.75 to 0.95 once the range was read from the book's unconditional sentences instead of the IABIED conditional title sentence.
TL;DR: ArgMap is a tool I created to make arguments explorable and computable, with every step directly linkable. Its first map is of If Anyone Builds It, Everyone Dies: I mapped the logical structure of the book's argument with LLM help. LLM readers derive confidence numbers from the authors' words through a rubric that was fixed in advance. The computation tests how well the authors' stated confidence fits their argument as I mapped it, rather than reflecting my personal view of the book. Out of 67 claims that the book both rates and argues for, 66 land within 10 points of the authors' range; the exception is the claim that everyone dies, at 0.62 against the authors' range 0.75 to 0.95. The policy conclusion for a global shutdown is at 0.93. You can adjust any probability on the map and see the conclusions change and tensions light up. I'd like you to find a step you disagree with or a toy example where the computation is wrong.
In the art of Kintsugi the broken pieces that no longer quite fit together are joined back into a whole, and the breakage and repair in gold become part of the piece.
Photo: "Hand Pinted Kintsugi Pottery Bowl" by Ruthann Hurwitz, CC BY-SA 4.0.
Whose number, and what "everyone" means
In If Anyone Builds It, Everyone Dies (henceforth IABIED), the word "everyone" is meant literally: everyone, everywhere on Earth, will die, killed by an artificial superintelligence (ASI) soon after one is built. The authors add: "We do not mean that as hyperbole." Many people have stated a probability for that outcome, but few explain how they arrived at it. By contrast, IABIED lays out the argument in great detail, and confidence is communicated as natural language or even omitted.
With the numbers read from the authors' natural language through the fixed rubric, the map reports a 62% probability that an ASI is built and kills everyone.[1]
What the question mark here asks is whether the book's argument, as the authors present it, is mapped correctly and whether the computation evaluates it correctly. Whether the premises of that argument are true is a separate question that the map helps the reader to investigate themselves.[2]
The authors deliberately don't state a number for the probability of the extinction claim
This probability and
This is what the map implies where society does not change course.
And this is what needs to be done.
Note that these numbers are not the authors'. For 67 claims the book both states a confidence in words and argues for them in more detail. The map also computes their values from the claims beneath them, without using the reading of the book's confidence at all. Those readings are then checked against the map's computed values. Of the 67, 18 land outside of the authors' range by more than one point (14 above the range, 4 below), and only one by more than 10 percentage points.
That one is also the most important: the extinction claim that everyone dies where the authors' statement is "The default outcome is lethal, but the situation is not hopeless; machine superintelligence doesn’t exist yet, and its creation can still be prevented", reading 0.75 to 0.95 through the rubric, against the map's 0.62.
For the book's central conditional the gap is larger still: the authors' "...then everyone, everywhere on Earth, will die. We do not mean that as hyperbole." reads 0.93 through the rubric against 0.74, the map's probability that everyone dies given an ASI is built.[5] Part of the gaps comes from choices in ArgMap that are not settled yet, see What is not settled.
Everything you see in the map below is LLM-extracted under my steering: the logical structure of the argument, the nesting into the different levels of detail, and the numbers. The numbers were elicited by LLM readers from the book and its supplements' words. The map has not been reviewed by the authors.
The map
This is the map of IABIED. See the custom widget below, or go to the full web app, which has more features and readouts.
It's interactive, so let's go through an example interaction: let's say you're not convinced that artificial superintelligence is possible and that's why not everybody will be killed by an ASI.
Click the "+" symbol on "Anyone builds an artificial superintelligence (ASI)" to unfold its sub-nodes. Drag the probability gauge on the left edge of "A machine superintelligence can be built" down to around 50%. "What-if mode" shows that the conclusion "Everyone, everywhere on Earth, will die" falls from 0.62 to 0.39. Does the extinction claim now read your own personal number? Or are there other underlying disagreements?
The extinction claim's gap on the map lights up in gold, the larger the gap the deeper. That is the golden glue of the Kintsugi bowl.
This map is also available in German, Traditional Chinese and Simplified Chinese, with identical structure and numbers.
Why map an argument?
Long arguments don't fit in working memory. We're damned to local reasoning and it's difficult to integrate the pieces into a coherent whole. The pieces of local reasoning we produce may contradict each other, and they cannot always be glued back together into one beautiful piece of porcelain. Upon inspection, we find that because one of its many fine joints is failing, the piece cannot stand on its own.
It would be much nicer to be able to explore and dive into an argument at the depth you desire, reason locally there, and see the conclusions change across the entire argument map. Conflicts would light up if you assign some credence that clashes with another part of the argument. The system would then calculate a closest consistent picture of the argument, and do that loudly by showing exactly where the glue went and where some parts were off. The rule that decides how much glue to put where is one choice among several; the main point is to highlight conflicts.
Then once you have such a system, what you get for basically free is that you can play with it: modify the argument, add your own credences, mark where you disagree, and see how the result changes.
So if the argument is mapped in enough detail, down to a level where the links are mechanical, and if the map is complete enough to capture the argument structure you want to engage with, then this becomes a calculation over scenarios that you can argue with, and every number in it is traceable back to an input you can see and change.
The book's claim, and where the numbers come from
In its introduction, IABIED makes the claim: "If any company or group, anywhere on the planet, builds an artificial superintelligence using anything remotely like current techniques, based on anything remotely like the present understanding of AI, then everyone, everywhere on Earth, will die. We do not mean that as hyperbole." (Hard Calls and Easy Calls)
Most of the remainder of the book then lays out the argument for this in more depth. Almost none of the statements of the book carry probability numbers or credences. But the authors did make careful[6] use of confidence language throughout the book. Wherever they do, the number is read from it. Where they don't, which is mostly for the common objections to the argument which the book quotes, a documented default is used.[7] The rubric was fixed in advance and is published here. Several readings were revised later, among them the extinction claim's.[8] Every revision is documented with its reason in the map's source.
In the places where the book makes a credence claim about a statement or reasoning link that is not further argued in more detail, that credence becomes an input to the map. Where the book makes a credence claim about a statement or reasoning link that is further argued in more detail, the credence becomes a check: it never enters the calculation and is only used as a readout on how well the map's finer version delivers the authors' number on the coarse claim. A gap here can mean that the book or the map is off; and if the map is right, at least one of the authors' numbers must be off.
Such a gap shows through one of the following rules, each further explained in its linked section:
The largest gaps are shortfalls.
How ArgMap computes the joint probability distribution
The authors' input numbers are used in the calculation as soft constraints, so a contradictory map can still get a distribution, and the disagreement that is left over is the glue that lights up in gold.
The reference distribution in an ArgMap is a Bayesian network derived from the logical structure of the map, with the cases removed in which a link for a claim and a link against it would both apply (i.e. where their premises hold; see also Reasoning links compete only where their grounds hold together). It fills probability mass at even odds where nothing is said by the author unless the rest of the map moves it. The result is the distribution closest to that network by minimizing KL divergence, meeting the constraints as well as possible, each weighted by how firmly it is held. With hard constraints this would be an I-projection.
The objective:
with
ArgMap builds a factor graph over the claims and reasoning links in the map, and solves this objective with a junction tree that runs in the browser.
What is not settled
Some parts of this framework are not settled yet. The first costs the extinction claim 3 to 9 points. The fourth about 8 to 10 points either way. The others don't move the numbers at all.
Whether that splits the book's reasoning into too many independent stages is Yudkowsky's "Multiple stage fallacy" question: Does splitting an outcome into stages with a probability for each and multiplying drive the result too low? Joe Carlsmith's reply to it says "compressing an argument into very few premises (or just directly forecasting the conclusion) risks hiding conjunctiveness, too." In ArgMap, shared grounds have to be named as a common premise.
On this map, full dependence between "The ASI that gets built is misaligned" and "A misaligned ASI causes human extinction" would recover 3 points of the extinction claim, and with the latter's four grounds also fully coupled, a total of 9 points with the extinction claim at 0.71. Note that even with both claims set to 100%, the extinction claim reads 0.82, because "Anyone builds an ASI" stays at 0.84 and still bounds it.
Here's an example of the last item. The coarse 90% link "a storm cuts the town's power" contains two steps, one of them with a premise that no reasoning link feeds into. The link shows 0.36 folded, and 0.72 with the premise set to 100%, against the authors' coarse 90%. On the IABIED map, 0.93 and 0.67 are the same two kinds of numbers for the "If anyone builds it, everyone dies" link.
The ask
Find a step or claim you disagree with and let me know with the "💬 give feedback on this" button in the selected node's popup, or with the "💬 Feedback" button in the full app's toolbar, also anonymously.
Write a small toy example in the editor (full app -> advanced mode -> editor) that shows where the computed number is wrong. The "semantics audit" example group in the map picker shows what an example could look like. Confirmed errors get fixed and credited by name.
And if you find yourself in some discussion about AI risk that is covered by the book, you can link the exact step: "copy link to this" on a node's popup will load the map as you unfolded it and focus that node, and keeps your dragged what-if numbers.
Appendix
Build date
The numbers in this post follow the map's build from 2026-10-15. The maps embedded here follow the live site, so there may be some drift. Changes are documented in the maps' sources with their reasons.
Where the probability goes missing
Reasoning links carry logical structure and only strengthen the conclusion or negated conclusion where they apply. That means that for each claim, in some percentage of cases none of the reasoning links into this claim apply. These cases count as even odds.
On a claim card, the paler top of a gauge is that half share.
In the first example, we have rain at 80% and the link "rain -> street wet" at 90%. The street reads 0.86 (0.8 × 0.95 + 0.2 × one half: the link's 90% is a floor, so where it rains, at least 90 of 100 cases are wet, the rest at even odds; without rain, nothing is said and these cases also count at even odds). Drag the rain down and the paler strip widens. Drag it to 0% and the gauge goes grey (a claim with a single support at zero and nothing arguing against it just shows a grey stripe).
The second example has the inverse link "no rain -> dry" at 90% and the street is 0.77 at rest. The same drag of rain down to 0% puts the street down to 0.05.
On the IABIED map, there is also an example with no inverse link on purpose: "Enormous economic and strategic incentives drive the race", since an ASI could even be built under other motives. Drag it and "What physics permits, someone eventually builds" to 0 and "Anyone builds an ASI" still lands at 0.45 with the even-odds fill.
Arguments do not vote
Let's look at another example to see how multiple reasons for and against a claim interact and why support and objection on the same case don't touch each other if possible.
Below you see two toy examples arguing for a product launch. The first has one support at 85% against one objection at 15%. The second has five supports at 85% against one objection at 15%. In both the launch is at 0.85.
If you drag the objection to 0%, the claim absent the objection reads 0.925 with one support (1 - (1 - 0.85) × half), and 1.00 with five (1 - (1 - 0.85)^5 × half).
(Drag four of five supports to 0% against the objection at 15% and nothing changes, still 0.85.)
Humans often learn well through stories, where long lists of reasons impress themselves into the listener's memory merely through the length of their serialization into language, and where independence structures are rarely more complicated than simple chains or fans. If this example was told as one, it could be very convincing: five inputs all counted as independent 85:15 evidence multiply to (85/15)^5 = about 5,800 to 1. It would then be easy to overlook the single objection (which if multiplied would only multiply 15/85 to cancel one of the five supports, still very convincing at 1,030 to 1).
To fix this, in ArgMap, objections can be beaten by a response, a lower number, or doubt about their premises. So as long as the numbers fit, a multiplicity of supports outvoting them is not enough.
And if there are many supports against many objections, human reasoning about all of them breaks down.
Reasoning links compete only where their grounds hold together
What happens when two (or more) links on the same premise claim more than all of a conclusion's cases?
In this example we see two diary entries that promise too much, 90% for Monday and 70% against Monday, given that the diary is reliable. The diary's reliability itself nobody has vouched for, in the non-conflicting case it sits filled at even odds.
In ArgMap, a for- and against-link resting on the same grounds never apply together and the excess 0.60 (90% + 70% = 160%) cannot happen. The cases producing the contradiction become rarer, and the doubt lands on the diary's reliability that is free to give. It falls from 0.50 to 0.21 and Monday reads 0.55.
Now drag the diary to 100%. It holds and the two links bend instead until they fit, with no excess left. Monday now reads 0.72.
Contrapositively, this gives rise to another authoring rule: If you want to author two links with contradicting conclusions and you want them to not compete with each other, you must put them on different grounds that cannot hold together: by explicitly naming the variable that distinguishes the two.
How firmly each number holds
Under disagreement of the authored numbers something has to give. This was already visible in the diary example above when the diary's reliability was held high.
What gives how much is determined by an additional parameter to links and claims.
Each reasoning link has a kind that is also read from the authors' language: formal, deductive, mechanism, empirical, testimony, analogy, hope.
The currency of these kinds is the number of coin flips at which their link is held when you drag something: formal never bends; deductive 1000; mechanism 64; empirical 16 or an optional sample size; testimony 16; analogy 4; hope 4; default 16. A number of 4 coin flips gives easily and 1000 moves little. The number held at n coin flips resists a change like n observations at that rate would.
Claims with an authors' range give depending on the width of that range: 18 flips for a range of 0.85 to 0.95, 8 for 0.7 to 0.9, 198 for a point interval. More specifically, claims with a width of four points or more give before any mechanism or deductive link.
Prior art
There are already similar systems, each lacking something an author needs:
Markov logic networks have weights instead of probabilities, and the joint distribution is normalized by the partition function over the whole model. A weight's effect depends on every other one and that's why they lack the ability to be elicited locally. ArgMap takes probabilities as inputs and solves for the factor weights.
Probabilistic conditional logic under maximum entropy (Paris 1994; Kern-Isberner 1998, 2001) is a close relative, has authored numbers as a single conditional, solved with a similar projection, but with hard constraints, which needs a consistent knowledge base. Later work (Potyka and Thimm 2014, 2015) handles inconsistency by first minimizing the violation and then choosing among the distributions at the minima, whereas ArgMap has a single weighted objective. In ArgMap, a number's meaning differs (a lower bound on the conditional, forcing some mass on the worlds where the conclusion holds given the premises, and leaving the remainder to be filled; not an exact conditional), it starts from the Bayesian network induced by the map, and how firmly each number holds is elicited from the source. This has an author who is locally honest and globally inconsistent as a legitimate end state, without hiding the glue.
MTAIR (Modeling Transformative AI Risks) is a quantitative model based on numbers from modellers through literature review and engagement with experts. ArgMap reads the numbers from an author's words.
AMTAIR (Automating Transformative AI Risk Modeling, 2025) has LLMs extract argument structure from AI safety papers into an ArgDown-like syntax, probabilities into a "BayesDown" syntax, then a Bayesian network with CPTs and prediction markets for live updates. The numbers come from the literature via the LLM and from markets, where ArgMap reads them from an author's confidence language. The output is a Bayesian network, where ArgMap uses for/against links with argument nesting.
ArgMap is very different from other types of argument mapping like Argdown, Argunet, Kialo (these are almost purely structural) or Concludia (non-probabilistic), and wiki approaches like Arbital (explorable explanations) with its "lenses" as tabs explaining a page at several depths and where users could cast probability votes on claims but which weren't further processed.
The conjunction
Compare Aumann's agreement theorem, where two people with the same prior whose posteriors are common knowledge will arrive at the same posterior. Aumann names "errors in calculating posteriors" as one reason for lasting disagreement. In ArgMap the two are the authors' stated confidence for a claim and the map's computation based on the authors' own numbers in the detailed reasoning. They should agree, and a gap means that one of the calculations is off. But the theorem says nothing about the agreed-upon posterior being right; the two could still be wrong together. So whether the authors' premises are true is for the reader to judge (and adjust in What-if mode!).
From IABIED's online supplement, introduction chapter, "AI Experts on Catastrophe Scenarios": "We have concerns with the practice of trying to assign a “p(doom).” Assigning a single probability — as opposed to multiple probabilities that each assume a different response society could choose — strikes us as defeatist."
The book itself draws this distinction in chapter 12 on Toby Ord's 10%: "if you look into the details, Ord says the reason he estimates 'only' a 10 percent chance ... is because he expects humanity to come to its senses".
The map shows 0.67 on the folded link, which is the strength the finer steps multiply to, a different readout from
Chapter 11, endnote 10: "While we authors think that disaster is a strong default outcome, five nines is a ridiculous level of confidence that neither of us endorse."
Out of the 115 objection links here (excluding nine inverses that argue a negated conclusion from its negated premise), 81 remain at the rubric's 0.15, or 0.20 where the book engages with the objection at length or grants some part of it (counted 2026-10-12).
The extinction claim's reading changed from 0.88 to 0.98 down to 0.75 to 0.95 once the range was read from the book's unconditional sentences instead of the IABIED conditional title sentence.