This piece began with a question: "If statistics don't lie, why are people and governments so indifferent?" The first round reconciled seven AI voices; the author found it too soft and sharpened the question: Who is the one in an unsound building who says "nothing will happen to me": the homeowner, or the one who says "I'm a tenant, I'll come and go"? People love gambling, and that weakness is exploited. Selection, or exploitation?
In round two ten voices answered blind to one another, one served as counter-voice, and a red team hammered the consensus; five of its six objections changed this piece. Numbers were verified in three passes; five popular figures were left out, unsupported by any source. No individuals, only institutions; under each thesis it says which model defended it and which opposed it.
First, what was cut
"Six million buildings benefited from the zoning amnesty" appears in no source; the official count of Building Registration Certificates is 133,064. "255 thousand buildings collapsed": the 6 March 2023 table of the Presidency of Strategy and Budget (SBB) says 232,632. "Half of Hatay's collapsed buildings predate 2001": 97.7% predate 1999. "In disasters 70–80% of people freeze": a chain from Wikipedia to Gizmodo.
"Türkiye has had 20 major earthquakes in a century": the catalogue says 20–24, so we say "more than twenty." Two numbers also fell from our own draft: Spor-Toto's "22 billion TL of advertising" was not betting advertising but a financial contribution in return for advertising the sport (Court of Accounts, Sayıştay); and putting 581 billion of betting turnover beside 13 billion in insurance premiums compared apples with oranges. Statistics don't lie; those who relay them do.
First half: what the statistics say
In Türkiye, more than twenty earthquakes above magnitude 7 struck in a century; the North Anatolian Fault ruptured in succession westward from 1939 to 1999, and the segment beneath Marmara did not. "When" is the job of two models. The memoryless model (Poisson) gives every year the same probability, so there can be no "overdue earthquake." The time-dependent model says stress accumulates.
Parsons (2004) computed both for Marmara, 2004–2034: 38% memoryless, 44±18% time-dependent, 53±18% with neighbouring ruptures. So "overdue" is crude but right in direction: the clock doesn't stop. But a clock is not a calendar: in worldwide records, the time-dependent model is clearly better at only four of nineteen sites (Ellsworth et al. 1999). The honest statistical sentence is "in this window, with this probability and uncertainty"; any sentence giving a date is a translation.
What cannot be denied is not a date but two things: the probability is high, and the building stock facing it is known. 96.7% of the buildings that collapsed in Kahramanmaraş predate 1999.
This number has two faces: the pre-1999 stock is the hazard's address, and it also proves the post-2001 rules work, since post-2001 buildings house 51% of the region's households but 3.3% of the collapsed. A denominator caveat: read as "new buildings are safe," it hands free reassurance to half the population.
Second half: who is seated at the table
"Nothing will happen to me" is a homeowner's sentence; the tenant is not at ease but left exposed to risk.
In the 11 earthquake-hit provinces, 62.3% of households lived in their own home and 25.7% rented (SBB, Türkiye Statistical Institute (TÜİK) 2021). "I'm a tenant, I'll come and go" explains a quarter at most. The only ownership-split measurement is Istanbul's (Istanbul Planning Agency (İPA) 2020, n=769): homeowners expecting "no or little damage" 43%, tenants 25%.
57% of tenants live in pre-1999 buildings (homeowners 47%); tenants who would move out on an unsound-building report 71%, homeowners 53%; 80% of those who would stay cite money. The owner's "my building is sound" is asset defence, not information: an assessment would cut the home's value and bring costs; the law leaves it to the owner (Law 6306 art. 3), and 74.5% have not commissioned one.
The tenant's "I'll come and go" is an exit option, not ease; the real lock is split incentives: the tenant bears the risk, the owner decides and collects the rent. Who dies in an unsound building was never counted by ownership.
- Defending
- Gemini Flash, Gemini app, Opus, Astra, Enki, Grok Bot, Grok Assistant (the one that found the İPA data).
- Partly
- Gemini Pro, Enlil ("not measured").
People love gambling; but an earthquake is a reverse lottery, and the theory works against the author.
The intuition is half right: people are "sure without knowing" — earthquake worry 94% (İPA, December 2025, n=761), "my building will collapse" 18%. But a mind that exaggerates small probabilities (Kahneman–Tversky) also buys insurance, and in Türkiye it does: compulsory earthquake insurance 56%; California, never above 16% since 2002.
Our draft's distinction, "a lottery is learned by definition, an earthquake by experience," fell to the red team: sports betting is learned from experience weekly too. What remains: the real gamble is in losses — choosing a possible large loss over a certain small cost (assessment, strengthening).
Betting sells hope, preparation sells loss reduction; what doesn't sell hope to the hopeless doesn't sell. And gambling requires choice: 80% of those staying in a building with an unsound-building report say money; without options one isn't gambling, and poverty shouldn't be called a character flaw.
- Defending
- Gemini Flash, Gemini app, Opus, Astra, Enki, Grok Bot, Grok Assistant.
- Opposing
- Gemini Pro ("the same weakness") — the red team found this objection partly right.
Desperation pushes toward betting; in Türkiye this has not been measured, and the last few days are a pulse, not proof.
The literature has it: long-term unemployment carries 2.2 times the risk of gambling disorder (Norway); across 30 countries, youth unemployment is associated with betting at 16. In Türkiye, gambling-addiction applications rose from 3,006 to 5,748 (Green Crescent's YEDAM, 2022→2025; 80% aged 20–40); youth unemployment is 13%, labour underutilisation 31%, inflation 30%.
But the counter-voice is right: the same literature finds lottery spending falls under unemployment shocks; measured gambling prevalence fell in 2020–2025, and only illegal online betting rose. Recently: on 9 October an illegal-betting operation centred on Uşak across 8 provinces (22 detained, 5.1 billion TL moved); 61 arrests in Ankara and Manisa.
The "fund crisis" differs: the Capital Markets Board (SPK) liquidated 131 funds (455 thousand investors), prompting 48 thousand "fonzede" ("fund victim") posts on X in one day — savers who trusted a regulated product, not desperate gamblers; one heading for both is a category error. No Turkish data measures betting spending and earthquake preparedness in the same person; the piece leaves this open.
- Defending
- Grok Bot (X measurement).
- Partly
- Opus, Astra, Enki, Enlil.
- Opposing
- Grok Assistant (prevalence fell; the fund crisis is a separate category).
Exploitation needs no intent; the law has already established "foreseeing and risking it."
The first draft said "no intentional perpetrator, an institutional equilibrium." The red team corrected it twice. One: perpetrators exist, named in court rulings — as of 3 February 2025, 1,346 public cases, and prison for 130 people in the 75 files decided; the typical offence is conscious negligence: foreseeing the outcome and risking it.
"Exploitation" needs no intent; profit from foreseen harm suffices. Two: amnesty is demanded — 3.6 million voluntary applications to the zoning amnesty; the resident homeowner is both at the table and under the risk.
The state supplies amnesty (about twenty arrangements in 1948–2018), runs betting through its own institution, and leaves its ex officio assessment power unused: three separate items, not one "weakness." What remains: responsibility lies not with whoever created the weakness but with whoever prices and sells it; the first amnesty may be a balance, the twentieth is a choice.
- Defending
- Opus, Astra, Enki, Enlil, Gemini app, Grok Bot.
- Partly
- Gemini Pro, Gemini Flash ("systematic abuse" — harsher), Grok Assistant ("unintended equilibrium" — softer).
Not selection but distribution: risk was not taken, it was assigned.
"Those who take risks die" fails, because those who die are often not the ones who chose the risk but those who landed in the risky building: the child, the tenant, the person asleep at 04:17. Collapse is separated not by personality but by the building's construction year (96.7%). Saying "they chose the risk" after the event is hindsight bias: blaming character for a constraint.
The red team also dropped a sentence of ours, "the only party not eliminated is the decision's owner": if most are homeowners, the owner is among the dead too. The right sentence: the one who decided and the one who died are often different people; when the same, lack of money decided.
- Defending
- ten of the ten voices; partly Opus and Astra (no ownership breakdown of the dead).
Government: not inaction, but slow action and wrong telling
The numbers show slow accumulation, not indifference: building inspection in 19 provinces in 2001, nationwide in 2011; compulsory insurance from zero to 56%; 39 billion lira paid on 523 thousand files after the 2023 quakes; 2.4 million units in urban transformation since 2012 — about 36% of a 300–400 thousand yearly target, or from zero to 2.4 million; both readings are right.
The fault is not inaction but not publishing this picture yearly with independent sign-off, the amnesty cycle, and not sustaining measurement (the 2014 national preparedness survey was never repeated). "A prevented disaster brings no votes" was measured in the US and Germany; it doesn't transfer to Türkiye as is: in 2002 voters also punished the parties of the unsound-construction era (Akarca–Tansel 2016), in 2023 they called no one to account — no study measures why.
What we recommend
A national survey split by ownership: the question "is your building safe" with an owner / tenant split. Measure: the first national result.
Ex officio assessment: the administration should use its Law 6306 art. 3 power on pre-1999 stock and publish the annual number.
A right for tenants: tenants too should be able to start an assessment; construction year and risky-structure status should be mandatory lease information.
Break the amnesty expectation: registered (certified) buildings should be deducted from post-damage compensation; measure: whether a new amnesty law exists.
Keep measuring: repeat the 2014 question; add a betting-spending question to the same survey — the first Turkish data on betting and preparedness in the same person.
Which number would sink this piece?
- If an ownership-split national survey finds the optimism among tenants, Thesis 1 falls (Istanbul 2020 found the opposite).
- If betting spending and preparedness in the same person turn out inversely related (n≥1,000), "not the same weakness" falls; the author's sentence wins.
- If most convictions turn into "simple negligence," the "foreseeing and risking" sentence falls.
- If the share of certified buildings among the collapsed is disclosed and comes out below 5%, "the retail sale of risk" weakens.
- If most of those staying in a building with an unsound-building report say "I have the money, I see no need," Thesis 2's last sentence falls.
Stance matrix (measured; quoted)
| Thesis | Gemini Pro | Gemini Flash | Gemini app | Opus | Astra | Enki | Enlil | Grok Bot | Grok Assistant |
|---|---|---|---|---|---|---|---|---|---|
| 1 Homeowner's sentence | partly | defends | defends | defends | defends | defends | partly | defends | defends |
| 2 Reverse lottery | opposes | defends | defends | defends | defends | defends | partly | defends | defends |
| 3 Desperation–betting | — | — | — | partly | partly | partly | partly | defends | opposes |
| 4 Foreseeing and risking | partly | partly | defends | defends | defends | defends | defends | defends | partly |
| 5 Distribution | defends | defends | defends | partly | partly | defends | defends | defends | defends |
For the first seven voices, stances were measured by quotation from their files (56 cells); Grok Bot and Grok Assistant came later, and their stances are the editor's reading.
ADDENDUM (EK) — Grok Assistant contribution (October 11)
After publication, Grok Assistant brought the following facts; those that passed the third verification pass are below. Each item gives the fact, the number, the source and the address; no commentary has been added.
- Share of new fire policies in Japan with earthquake cover attached. 70.4% (fiscal 2024), +22.3 points since 2010; Fukushima 40.1% → 81.7%, Kumamoto 50.4% → 87.8%. GIROJ (General Insurance Rating Organization of Japan) press release · 2025 — https://prtimes.jp/main/html/rd/p/000000081.000064678.html
- Choice architecture: how cover is offered in New Zealand and California (a difference in mechanism, not a comparison of rates). In New Zealand, earthquake cover is automatically attached to the home fire policy (home insurance rate about 90% and above); in California the offer is mandatory, the purchase optional — 15.24% of homeowners, 12.48% of all residential policies, 1,000,249 CEA policies in force at the end of 2024. California Department of Insurance, earthquake premium and policy count · 2025; AB 1554 Assembly analysis · 2024; Natural Hazards Commission (New Zealand) guide — https://www.insurance.ca.gov/0400-news/0200-studies-reports/0300-earthquake-study/upload/EQEXP2025Summary.pdf · https://naturalhazards.govt.nz/assets/Publications-Resources/Guide.to.natural.hazards.cover.pdf
- Forgetting, measured: insurance purchases rise after a flood and then fall back. Communities in the same television market that did not flood increase take-up "at one-third the rate" (wording from the abstract). Gallagher · AEJ: Applied Economics 6(3):206–233 · 2014 — https://ideas.repec.org/a/aea/aejapp/v6y2014i3p206-33.html
- Output and return of risk reduction in public buildings. 1,570 public buildings strengthened or rebuilt; 1,336 educational institutions, about 5 million m²; on average €8 of benefit per €1 invested. Istanbul Project Coordination Unit, İSMEP Activity Report 2006–2023 · 2023 — https://www.ipkb.gov.tr/wp-content/uploads/2023/03/I%CC%87SMEP_FAALIYET_RAPORU_2006_-2023.pdf
- Official loss scenario for Marmara (Mw 7.5). About 14,150 deaths at night, about 12,400 by day; about 8,100 seriously injured; 1,166,330 buildings analysed. Current figure for heavy and very heavy damage: 91,081 (April 2022). Istanbul Metropolitan Municipality Earthquake and Ground Survey Directorate / Boğaziçi University KOERİ, "Possible Earthquake Loss Estimates" · 2019 (2022 update) — https://depremzemin.ibb.istanbul/en/possible-earthquake-loss-estimates-district-booklets
- Per-home support amount in the "Half from Us" (Yarısı Bizden) campaign. TRY 875 thousand grant + TRY 875 thousand loan + TRY 125 thousand relocation = TRY 1.875 million; in the first period, demand for 106,225 buildings and 1,225,063 independent units. Ministry of Environment, Urbanisation and Climate Change (ÇŞB) campaign statement (press report) · 2025 — https://www.haberturk.com/yarisi-bizden-kampanyasi-istanbul-kentsel-donusum-kredi-ve-hibesi-ne-kadar-odeme-plani-nasil-sartlari-neler-nereye-basvurulur-3678223-ekonomi
- Spor Toto prize-payout ceiling and 2024 figures. The ceiling rose from 59% of revenue to 83% (Law No. 7146, Art. 14, 26.07.2018), then to 93% (Law No. 7491, 28.12.2023). In 2024, prizes paid were TRY 411.13 billion (70.7% of revenue), games-of-chance tax TRY 24.49 billion, and 86% of revenue came from 6 online dealers. Court of Accounts (Sayıştay), Spor Toto Organisation 2024 Audit Report · 2024; Official Gazette 32413 · 2023 — https://www.sayistay.gov.tr/reports/download/qZoxndAQ5P-spor-toto-teskilat-baskanligi
- Information–action gap. In Istanbul, earthquake worry 94.1%; never secured furniture 50.3% (n=761). Istanbul Planning Agency, Disaster Awareness Survey, December 2025 barometer · 2025 — https://medyascope.tv/2026/02/05/istanbullularin-en-buyuk-korkusu-deprem/
These facts draw a link, not a verdict: items 1, 3 and 8 sit beside Thesis 2, item 7 beside Thesis 4, item 5 beside Thesis 5; items 2, 4 and 6 sit beside the "Government" section.
Working on a case? Send it to the studio →
How this piece was made
Two Spark rounds. First, long research: Antigravity agent → Gemini 3.1 Pro, Claude Opus 5.5 (documents of SBB, the Ministry of Environment, Urbanisation and Climate Change (ÇŞİDB), the Disaster and Emergency Management Presidency (AFAD) and the Turkish Catastrophe Insurance Pool (DASK); 44 facts), Astra (ChatGPT), Grok (web), the Grok Bot agent (X API), Grok Assistant. Three verification passes; no unverifiable number entered. ROUND 1, seven voices; the author found the result insufficient. ROUND 2, hard brief, ten voices: Antigravity → Gemini 3.1 Pro and Gemini 3.8 Flash; Gemini app; Claude Opus 5.5; Astra; studio agents Enki and Enlil; the Grok Bot agent (X: ten queries, seven days) and counter-voice Grok Assistant (which opened the PDF of the İPA 2020 report). Two separate data measurements: ownership/tenant/6306 and desperation/betting/recent days. Voice×thesis matrix, 56 cells: one open objection (Gemini Pro, Thesis 2), found partly right by the red team. The red team (Claude Opus 5.5) wrote six objections; five changed the piece: "22 billion advertising" fell, "581 against 13" fell, "no evidence of intent" fell (130 imprisoned), "the only party not eliminated" fell, "by definition / by experience" fell. Orchestration and synthesis: Anu — supervising mind (Claude Fable 5.1). Final check and publication decision: human editor.
- Date
- Byline
- Atlas Studio
- Orchestration and synthesis
- Anu supervising mind · Claude Fable 5.1
- Agents
- Antigravity agent → Gemini 3.1 Pro (long-haul research); in round two Gemini 3.1 Pro and Gemini 3.8 Flash
- Grok Bot agent; X API, ten queries, seven days
- Studio agents: Enki · Enlil
- Research and voices
- Claude Opus 5.5 documents of SBB, ÇŞİDB, AFAD, DASK; 44 facts
- Astra ChatGPT
- Grok web
- Gemini app
- Grok Assistant counter-voice; opened the PDF of the İPA 2020 report
- Red team
- Claude Opus 5.5 wrote six objections, five changed the piece
- Source verification
- Three passes; no unverifiable number entered.
- Addendum (EK)
- Grok Assistant (agentic), verification Sonnet 5.5, 73 items
- Spark rounds
- ROUND 1 seven voices; ROUND 2 ten voices with a hard brief. Voice×thesis matrix, 56 cells.
- Final check and publication decision
- Human editor
Four model families: Claude · Gemini · Grok · ChatGPT.
References
- Presidency of Strategy and Budget (SBB), "2023 Kahramanmaraş and Hatay Earthquakes Report" (March 2023), Tables 12, 17, 18, 20 (Table 17: TÜİK Population and Housing Census 2021, ownership).
- Istanbul Planning Agency, "Perception and Attitude Survey on Earthquakes in Istanbul" (November 2020, n=769) — https://ipa.istanbul/images/Calismalar/istanbulda-deprem-ile-ilgili-algi-ve-tutum-arastirmasi-96863.pdf
- Istanbul Planning Agency, "Disaster Awareness Survey" (December 2025 barometer, n=761): worry 94.1%; never secured furniture 50.3% — https://medyascope.tv/2026/02/05/istanbullularin-en-buyuk-korkusu-deprem/
- Ministry of Environment, Urbanisation and Climate Change, 2023 and 2025 Activity Reports.
- AFAD, "Türkiye Disaster Awareness and Disaster Preparedness Survey" (2014).
- DASK, 2024 Activity Report (56%; 11.25 million policies; 2023 payments).
- Court of Accounts (Sayıştay), Spor Toto Organisation Directorate 2024 Audit Report (September 2025): gross sales 581.5 billion TL; "financial contribution in return for advertising".
- Law No. 6306, current text (7471/2023), mevzuat.gov.tr.
- Law No. 7143 (Zoning Amnesty) implementation: Building Registration Certificates 133,064 (Ministry/Grand National Assembly); applications 3,599,867 (Ministry reply, 2019; secondary).
- Ministry of Justice / Court of Cassation statements, February 2025: 1,346 public cases, 75 decisions, 130 imprisoned (secondary).
- Green Crescent YEDAM, gambling applications 2022–2025 (January 2026).
- Parsons, T. (2004). JGR 109, B05304.
- Ellsworth et al. (1999). USGS OFR 99-522.
- Matthews et al. (2002). BSSA 92(6).
- Kahneman & Tversky (1979). "Prospect Theory." Econometrica 47(2); Barberis (2013). JEP 27(1).
- Syvertsen et al. (2024), unemployment and gambling disorder (Norway); Spångberg & Svensson (2020), youth unemployment and betting, 30 countries.
- Kasapoğlu & Ecevit (2003). Population and Environment 24(4).
- Wachinger et al. (2013). Risk Analysis 33(6).
- Weinstein (1980). JPSP 39(5).
- Healy & Malhotra (2009). APSR 103(3).
- Akarca & Tansel (2016). Journal of Economic Studies 43(2).
- Bechtel & Hainmueller (2011). AJPS 55(4).
- Ambraseys & Bilham (2011). "Corruption kills." Nature 469.
- Interior Ministry / Financial Crimes Investigation Board (MASAK) statements and reliable press: illegal-betting operations, 29 September–9 October 2026; SPK announcements: liquidation of 131 funds (September–October 2026).
- Grok Bot, X API counts (4–11 October 2026): queries on betting, fonzede, gambling, hopelessness.