The busiest day did not repeat, and the visitors who submit fastest are the ones who keep the page's number.

Every submission since 1 August, by quiz level

287 submissions arrived in these ten weeks, and the 285 with a quiz level are drawn, so they get the full width; the charts below cover every month. The tall column of dots two weeks from the right-hand edge is 1 October. Hollow dots are readers who took the expert quiz without a pass: those who simply called themselves experts before any check existed, and three who have since fallen short of the check and continued anyway. Filled dark ones are verified: they passed a 30-term vocabulary check before the expert quiz would open. The line is a running median, the median of each submission and the fourteen before it, with identical repeat rows left out, and it is dashed across gaps of more than two days between submissions. Hover or tap a dot for its p10–p90 range; with the chart focused, the arrow keys step through the dots. The two rows in this window that arrived with no quiz behind them are left out, as all six such rows are from the strips below.

Monthly median with interquartile band

Line: median of that month's submissions, every level and every row together; band: 25th–75th percentile. September is complete. October's point is hollow and joined by a dashed line because the month runs only to 14 October, and 37 of its 95 rows are one day, 1 October. The next chart splits the same line by quiz level.

September's median was 49% on 130 submissions. October so far is 39% on 95, and without its first day 54%.The first three months of the site, before any quiz existed, produced a median of 15%. The quiz arrived in March, the monthly median reached a peak of 69% in August, and September, still the busiest month in the series, came back down to 49%. October's 39% is two different halves averaged into one point: 1 October, 37 rows at 23%, and 2 to 14 October, 58 rows at 54%. The same split holds among visitors who set their own number, repeats excluded: 50% in August, 29% in September, 18% on 1 October and 40% in the fortnight after it. Why the visitors changed is not something a table of submissions can say. The author's impression of how the site's reception turned around August is in a note; it is an impression, not a measurement.

The burst did not come back. The thirteen days after it ran at September's pace.58 rows arrived between 2 and 14 October, 4.5 a day, with nine on the busiest days, 6 and 8 October; September averaged 4.3 a day. 1 October, at 37 rows, is still nearly twice the next busiest day in the series. Its low tail did not vanish entirely: two later days with five or more rows also have medians under 25%, 2 October at 20% on seven rows, three of them from one browser, and 8 October at 22% on nine. Those are the only two such days besides 1 October since the chooser went live. Where the 1 October crowd came from is still not in the table.

Monthly median by quiz level

Every row, then each quiz level, on one axis from March 2026, when the level chooser went live; the chart above has the months before it. Every row counts all submissions in the month, including kept ones, so it is the line above. Moved a slider redraws all four lines on the quiz rows where at least one midpoint left the quiz's proposal, identical repeats excluded, which takes the page's own arithmetic out; there the first line is every quiz-taker who moved. A point from fewer than five rows is small and hollow, and the segment into it is dashed: before August each level had between none and four submissions a month, so those stretches are anecdotes, not trends. October is shaded because it runs only to 14 October. Hover or tap a month for every value and its n; the monthly summary further down gives the same medians as a table.

On every row, the expert line has fallen three months running, 59%, 32%, 23%. The beginner line fell and came back, 72%, 48%, 56%.Medium went 69%, 63%, 40% over the same three months. In August the three levels sat within thirteen points of each other; in October beginner and expert are 34 points apart. Most of that gap is the proposal and how often each level keeps it: across the whole export the beginner quiz is kept on 42% of its rows, medium on 39% and expert on 20%, and a kept row is the formula's number. Switch to the rows where the visitor moved a slider and the lines close up: in September they sat at 31%, 27% and 28%, the convergence the last report described.

In October the movers come apart again on one side: medium at 20%, expert at 22%, beginner at 54%.Before reading anything into the beginner point: it is 16 rows, three of them one browser on 2 October that took the same quiz three times in six and a half minutes, and "moved" includes small moves. Of the twelve beginner movers after 1 October, two ended within two points of the proposal and two dragged it up to 100%. Across the whole export the movers' medians are still 36%, 33% and 31% on 54, 75 and 79 rows, and before October the beginner movers stood at 32%. The medium and expert movers are the ones that have kept falling, from 61% and 47% in August, though 1 October supplies 25 of their 57 October rows. The beginner line is something to watch at the next export, not a finding.

Every submission's p(doom), by quiz level

The stats page's histogram, the same twenty bands of five points over all 424 submissions, with each bar split by the quiz the visitor took. From the baseline up: beginner, medium, expert, and in grey the rows with no quiz, the 98 from before the level chooser existed and the six later ones without a level. Hover or tap a bar for its breakdown and for how many of its rows kept the quiz's proposal; with the chart focused, the arrow keys step through the bars.

The two spikes in the upper half are the page's proposal. The tallest bar, 0 to 5%, is not.The 70–75% band holds 36 rows, 17 of them beginners, and 20 of the 36 kept the proposal: it is where the beginner quiz puts a visitor who ticks most of its boxes. The 95–100% band holds 28, 17 of them medium-quiz takers, and 15 of the 28 kept it. The 0–5% band is the tallest at 54 rows, and none of them kept a proposal: 22 are from before any quiz existed and the other 32 are quiz-takers who took the page's number down to near zero, 15 of them on the medium quiz. The 10–15% band is mostly the site's first months, 25 of its 36 rows. A single-colour histogram, as on the stats page, shows three peaks; split by level, each peak is a different population for a different reason.

Who says what: distribution by quiz level

One mark per submission; the tall tick is the group median. An arrow pointing left is a visitor who moved the page's proposed number down, an arrow pointing right one who moved it up, and a round dot either kept the proposal or, in the bottom row, never had one. The expert rows are split three ways: the 22 who declared themselves experts before the check existed, the 74 who have passed it since, and the three who fell short of it and continued to the expert quiz anyway. Three marks do not make a median, so that row has no tick.

The widest-spread factor tracks the total most. That is arithmetic, not behaviour

Each panel: one factor's midpoint against the submission's own final p(doom). Read this as a sensitivity check, not a finding about people. Final p(doom) is the product of these three numbers, so each factor is being correlated with something it helped compute. What separates the panels is how much each factor varies across this set of submissions, not where it sits in the chain; reordering the chain leaves every correlation unchanged.

What the page proposed against what was registered

One dot per quiz submission. Horizontal: the p(doom) the quiz set the sliders to, recomputed from the stored answers with the page's own rules. Vertical: what was registered. A round dot on the diagonal never had a midpoint moved; off the diagonal the mark is an arrow pointing the way the number went, the same encoding the strips use further down. Identical repeat rows are not drawn.

The kept share has fallen since the page began saying that the quiz proposes a starting point. One sentence cannot claim the fall.Since 21 September the page says, above the sliders, "A short quiz proposes a starting point from what you already know", and 119 quiz rows have arrived under it. 34 of them kept the proposal, 29%, against 37 of the 103 rows in the nineteen days before it, 36%. The window since breaks into three pieces that differ from one another more than any of them differs from the window before: 12 of 24 kept in the last ten days of September, 8 of 37 on 1 October, and 14 of 58, 24%, from 2 to 14 October. By month the kept share was 39% in August, 38% in September and 23% in October. Most of the October fall is the expert quiz, where 1 of 15 rows since 2 October kept the proposal, against 5 of 17 beginner rows and 8 of 26 medium. That fits a sentence that works, and it fits an audience that changed, and neither the timing nor the size tells them apart.

Where the registered numbers land

Every quiz submission, repeats excluded, counted in ten bands of ten points. Top row: the 106 that kept the proposal. Bottom row: the 208 that moved a midpoint. The two rows share the horizontal scale and each bar is labelled with its count.

The commonest place for a mover to land is under 10%.44 of the 208 movers registered less than 10%, against 2 of the 106 kept rows, and no other band of ten points holds as many movers. Their proposals had a median of 64%: these are visitors who were handed a number near two in three and took it to near zero. The kept rows sit where the formula puts them, 38 in the seventies and 24 at 90% or above. The share of movers under 10% has been steady outside one day: 8 of 54 before September, 15 of 81 in September, 9 of 44 from 2 to 14 October, against 12 of 29 on 1 October. By level it is 7 of 54 beginner movers, 19 of 75 medium and 18 of 79 expert. So the movers' median is a median of two kinds of visitor, those who adjust the proposal and those who throw it out, and 40 of the 208 moved it up rather than down.

Time on the sliders, and who keeps the proposal

Since 12 September the page records the moment it applies the quiz's proposal, and 164 quiz submissions carry it. The time here runs from that moment to the submit press: time spent on the sliders, the uncertainty bands and the rest of the page. Each bar is the share of rows in that time band that kept the proposal, labelled with the count. Identical repeat rows are left out.

Of the visitors who submitted within a minute, 63% kept the page's number. Of those who took more than three minutes, 14% did.25 of the 40 rows submitted within a minute kept the proposal; 3 of the 21 that took more than three minutes. The median time is 67 seconds for a kept row and 102 for a moved one. It is not one level's habit: split each level at its own median time and the faster half kept the proposal on 12 of 21 beginner rows, 15 of 32 medium and 11 of 28 expert, the slower half on 4 of 22, 9 of 32 and 4 of 29. No row came in under 20 seconds, and 52 of the 55 kept rows that stored the proposed band changed it, so a quick kept row is not a visitor pressing submit blind: it is one who looked at the page's number briefly and let it stand. Among the movers, time does not predict where they land: the rank correlation with the number registered is +0.02, and with how far the number moved +0.07, so the slow ones do not drag further. Time cannot say which way the arrow runs: a quick submitter may be someone already satisfied by the proposal, or someone in a hurry who would have kept any number.

Kept means the three midpoints are exactly where the quiz put them, and nothing more. The uncertainty band is a separate control: of the 55 kept rows that stored the proposed band, 52 registered a different one. Identical repeat rows from one browser are left out of every kept and moved count, which is why the totals here are 314 and not 320. Since 12 September the page stores the proposal with each submission, and 164 rows carry it. For the quiz rows before that, the proposal is recomputed from the stored answers with the page's own tables, and the build checks itself two ways: every kept row sits at exactly the per-factor value the page would have set, and on the 164 rows that have both, the stored and the recomputed proposal agree to within a tenth of a point. Seven rows have three equal midpoints at a value no proposal produces, a visitor's own doing, and are counted as moved. One moved row was submitted 53 minutes after the proposal; a tab left open is the likeliest reading, and it sits in the last band.

What became of everyone who clicked Expert quiz

A score is recorded exactly when the check ran, which is exactly when the expert card was clicked. What the table cannot hold is anyone who took the check, saw the verdict and closed the tab.

Of the 17 new finishers, 14 cleared the check, two took the medium quiz it recommended, and one continued to the expert quiz anyway.The two who took medium scored 22 and 25 and both kept the medium quiz's proposal, at 67% and 95%. The one who continued scored 25 on 8 October and registered 10%, joining the two from the last report, who scored 24 and 25 and registered 27%, the expert proposal kept, and 0.4%. All nine who took a recommended quiz scored 19 to 25; eight were offered medium and the one who scored 19 was offered beginner. Their submissions span 33% to 99%, so falling short of a vocabulary bar says nothing about the number someone then registers. Everyone else scored 26 to 30, at a median of 29. The expert share of submissions that declared a level is 30% since the check, 77 of 254, against 33% before it, 22 of 66; without 1 October it is 27%, and from 2 to 14 October 26%, 15 of 58. Medium is still the largest level, 42% of everything since 19 August. Whether the check deters people from clicking the expert card at all is not something completed submissions can show: a visitor who took the check and closed the tab leaves no row.

The level chooser, before and since the check

Completed submissions per path in each window. In the first window a visitor could still skip the quizzes and go straight to the sliders, or take a standalone knowledge check; both were retired on 19 August, when the check moved onto the expert card, and since then every path on the page runs through a quiz. Two rows from 21 August arrived with no level anyway. They are unsigned, which is what a tab loaded before the 19 August deploy would send. They are left out of the right-hand panel, and since 2 September the database refuses any row without a level.

Term by term: what the 86 finishers ticked

Decoy-to-term pairings are an editorial reading of the term list; the AI-related flags come from the page. Scores here are recomputed from the ticked terms against the list as it stood on the taker's day; Mary Shelley was replaced by Embedding on 2 September, so the four earlier finishers were shown her and the eighty-two later ones were shown Embedding; which list a row saw is resolved from its page version, and the recomputed score agrees with the score the page stored on every one of the 86. The one submission through the retired standalone path (July 2026) is shown separately in the tooltip.

Verified experts against self-declared ones

Left: the share of options ticked on each expert-quiz question, for all 99 expert submissions. Right: how many of the 12 named organisations, commentators and creators each recognised, against the p(doom) they submitted. Dark dots are verified experts, grey ones self-declared, and the two groups sit to either side of each column; the three hollow dots are the readers who fell short of the check and continued.

The 74 who passed the check submitted a median 32%. The 22 who called themselves experts before the check existed, 68%.Among visitors who set their own number the gap is smaller, 25% against 44%, on 60 and 17 rows. The pooled gap includes the kept rows on both sides, five self-declared experts at a median 77% and fourteen verified ones at 69%; leaving them out changes who is being compared as well as what, so the two gaps are two different comparisons rather than a decomposition into proposal and belief. The last report saw the verified readers' recognition of names in the field fall, to a median 3 of 12, and put it down to the readers verified since 20 September, who recognised 2. This export says whose it was: the 18 verified on 1 October recognised a median 2, and the 14 verified since recognise a median 4, the same as the 34 verified before 20 September. The verified median over all 74 stays at 3, against 4.5 for the self-declared. The other difference has held too: all 22 self-declared experts ticked every option on the self-improvement question, against 43 of the 74 who passed the check. Ticking everything is consistent with a list that is half-recognised, though a checklist nobody can fail cannot tell that from thoroughness. The two groups are also separated in time, since the check only exists from 19 August, and the site's overall median fell over the same period, so some part of this is the audience changing rather than the check selecting.

The expert quiz's funding question is an opinion, not a recognition test, so it is left out of both panels. It appears in full in the expert section below.

Ticked most and fewest, across all three beginner checklists

Every beginner's ticks on the film, catastrophe and capability lists added together, out of 50 possible, and the cohort cut into thirds on that total. Left: one mark per submission, the tall tick is the group median; an arrow points the way the visitor moved the quiz's proposed number, a dot kept it. Right: what each third did with that proposed number.

Pooled, the thirds sit at 57%, 61% and 51%. Among the 54 who moved a slider, the third who ticked most registered 40%, the middle 40% and the third who ticked fewest 31%.The pooled figures are mostly the page's proposal, which rises with the number of boxes ticked; 39 of 93 beginners kept it. At the last report the movers' thirds were level, at 32%, 34% and 29%. All three have risen since, the upper two most, which is the October beginner movers seen from another side; on cohorts of 14 to 21 that is a direction, not a slope. The catastrophe list is still the one place a gradient runs the other way, and it has held through two exports. The third who ticked most catastrophes and set their own number registered a median 25% (n = 31), the middle third 47% (n = 12) and the third who ticked fewest 52% (n = 11): the visitors who recognise most of the ways the world could end set the lowest number. Ticks on a checklist are a count of recognised names, not a measure of knowledge.

How strongly each answer tracks the p(doom) submitted

Rank correlation per question. The filled mark uses all 128 rows; the hollow mark drops the six identical repeats; the square uses only the 75 rows where the visitor moved a slider.

On all rows the vulnerability checklist tracks p(doom) at ρ = +0.18. On the 75 rows where the visitor moved a slider, it is −0.02.That is the page's own formula being measured, not a fact about people: the share of vulnerabilities recognised is one of the inputs that moves the proposed number up, and 47 of the 122 medium rows without repeats kept the proposal. The pooled figure was +0.66 in the August report, +0.29 in September and +0.25 at the last one, and on the movers it has gone from +0.12 to nothing. What survives on the moved rows is the three opinion questions, governance at +0.41, competition at +0.32 and speed at +0.30, which pick the band and are near restatements of the sliders. System prompts sits at −0.02, and "seen AI go off the rails" falls from +0.23 pooled to +0.07 among movers. On this quiz, what a visitor recognises has no measurable association with the number they set for themselves.

The three opinion questions are not findings: a belief about whether governance can keep AI safe correlating with an estimate of catastrophe is close to tautological.

Ticked most and fewest on the vulnerability checklist

The one multi-select question on this quiz, so the overall split is the same as the per-question one. Thirds by how many of the 16 entries were recognised. Left: one mark per submission, the tall tick is the group median; an arrow points the way the visitor moved the quiz's proposed number, a dot kept it. Right: what each third did with that proposed number.

Pooled: 67%, 62% and 63%. Among those who moved a slider, repeats excluded: the third who ticked most registered 35%, the middle 22%, the third who ticked fewest 49%.At the last report the movers' thirds read 36%, 31% and 27%, a shallow slope in the pooled direction. With nineteen more movers the slope has gone and the third who ticked fewest is now the highest, on cohorts of 22 to 28. The ordering has changed at every export so far, which is what a checklist with no signal looks like when it is cut into thirds. Most of the pooled figure is still the proposal: the top third without repeats has 41 rows, 19 that kept the page's number at a median 94% and 22 that moved it, to 34%. The entries fewest people recognise are side-channel attacks (53 of 128), adversarial attacks (73) and model stealing (75).

Ticked most and fewest, across all seven expert checklists

Every expert's ticks on the seven multi-select questions added together, out of 24 possible, and the cohort cut into thirds on that total. Left: one mark per submission, the tall tick is the group median; an arrow points the way the visitor moved the quiz's proposed number, a dot kept it. Right: what each third did with that proposed number.

The third who ticked most submitted a median 45%, the middle third 28%, and the third who ticked fewest 31%.That is the shape of the last report, 44%, 25% and 32%, and among those who set their own number the thirds sit at 42%, 20% and 31%, on 32, 30 and 17 rows. A total across seven lists does not sort this cohort in one direction. One of the seven does, and it held through the fifteen new rows. On the campaign-organisations question, the 45 who recognised two or three of PauseAI, ControlAI and Microcommit submitted a median 54%, the 33 who recognised one 25%, and the 21 who recognised none 23%; among those who set their own number it is 42%, 23% and 14%. Knowing the organisations that campaign about AI risk and expecting a catastrophe go together, which is close to saying one thing twice. The mechanism questions are still ticked in full by most takers. 20 of 99 expert rows kept the proposal.

Claiming to know terms that are not AI terms, by how many real ones were spotted

The check shows fifteen real AI terms and fifteen decoys such as Agent Orange and Reinforced concrete. Ticking a decoy is overclaiming: saying you recognise something you cannot. The 86 who finished the check are cut into three groups by how many of the real terms they spotted, and each bar is the share of the group who ticked at least one decoy. Spotting a real term and ticking a decoy are counted separately, so one does not mechanically fix the other.

Check score against how sure the visitor said they were

One dot per check-taker. Horizontal: the score on the check, out of 30; the bar to clear it is 26. Vertical: the uncertainty band the visitor registered, the mean of the three sliders' ± spread, so lower is more certain. Colour is the quiz that followed, as in the legend: the expert quiz for those who cleared the check or went on anyway, otherwise the quiz the check recommended. The short horizontal lines are the medians of three score groups. The check score is not an input to the band the page proposes, so a slope here would be the visitors' own.

Self-rated skill with system prompts against what was recognised

Medium-quiz takers, identical repeats excluded. The self-rating is the visitor's own answer to "How familiar are you with system prompts?". The next two columns are from the vulnerability checklist: how many of its five AI-specific entries were ticked (prompt injection, sleeper agents, data poisoning, adversarial attacks, model stealing; which entries count as AI-specific is an editorial reading of the list), and how many ticked prompt injection, the attack that goes through a system prompt. The last three columns are the band, as a median ± in points: what the page proposed and the visitor's change to it, both on the 64 rows that stored the proposal, and what was registered, on every row. A negative change means the visitor narrowed the band.

The visitors who call themselves experts at system prompts recognise no more AI-specific vulnerabilities than regular users, and 3 of 18 did not tick prompt injection.Self-rated experts tick a mean 3.3 of the five AI-specific entries, against 3.7 for those who use system prompts regularly; 15 of the 18 ticked prompt injection, against 37 of 38 regular users. Across the four ratings the self-rating goes with ticking more of the general security entries (ρ = +0.34) more than with the AI-specific ones (+0.14), so part of what a higher self-rating predicts is ticking more boxes. That is the Dunning–Kruger shape at the top of the scale, but it rests on three people and on a checklist without decoys, where an unticked box may be oversight as easily as ignorance. The self-rated experts also register the narrowest bands, ±28 against ±60 for those who do not know what a system prompt is, and that part is not theirs: the page proposes a narrower band to a higher self-rating, and on the rows that stored the proposal the visitors' own change is unrelated to the rating (ρ = +0.07). The certainty belongs to the formula.

The p(doom) of the visitors who ticked a decoy

Expert-quiz takers who took the check, identical repeats excluded, split by whether they ticked any decoy. One mark per submission; a hollow mark kept the quiz's proposal, and the tall tick is the group median. The visitors who fell short and took the recommended quiz are left out, because the medium quiz proposes much higher numbers than the expert quiz and would make the comparison about the quiz.

Verdict: one Dunning–Kruger-like pattern, in what people claim to recognise, and none in how certain they are.Those who spot fewer real terms are about twice as likely to claim a fake one, which is the overclaiming half of the effect, and it holds on a measure built to catch it. Certainty does not follow. Among people who claimed expertise, those with the weakest vocabulary set wider bands, not narrower; among those who rate their own system-prompt skill highly, the narrow bands are the page's. The popular reading of the effect, that the people who know least are the most sure, is not what this data shows. Two limits apply to all of it. The check measures vocabulary, which the site's own definition calls an approximation of hands-on expertise, since vocabulary can be read up on. And nobody here estimated their own score, which is the measurement the effect is defined on; the directions section proposes adding it.