Skip to content
OGERIA — Observatory of Global Evidence on Risks in AISynthesis report · 2026 ed.
Updated 2 Oct 2026

Chapter 09

Figures

The observatory's data seen all at once, in small charts. Each has its own axis and unit: they sit side by side, they do not compare. Counting observatory entries is not measuring the world.

By vector

Evidence

Facts by year, stacked by vector

2026 has the highest count, with 247 of 288 facts.

2023 · Biological and CBRN: 12025 · Biological and CBRN: 42026 · Biological and CBRN: 152023 · Cybersecurity and infrastructure: 12024 · Cybersecurity and infrastructure: 22025 · Cybersecurity and infrastructure: 32026 · Cybersecurity and infrastructure: 692024 · Military and autonomous weapons: 22025 · Military and autonomous weapons: 22026 · Military and autonomous weapons: 262025 · Economic and labor: 62026 · Economic and labor: 262019 · Political and power concentration: 12023 · Political and power concentration: 12024 · Political and power concentration: 12025 · Political and power concentration: 112026 · Political and power concentration: 1252023 · Epistemic and information: 12024 · Epistemic and information: 12025 · Epistemic and information: 42026 · Epistemic and information: 262022 · Loss of control: 12023 · Loss of control: 22024 · Loss of control: 42025 · Loss of control: 122026 · Loss of control: 137

An event with several vectors counts once in each.

n = 288 · as of 2 Oct 2026 · counts observatory entries · See in Evidence ↗

Facts by type

“Policy” is the most frequent type, with 122 of 288 facts.

Incident: 50Evaluation: 38Warning: 36Policy: 122Framework: 28Model: 14

n = 288 · as of 2 Oct 2026 · counts observatory entries · See in Evidence ↗

Primary source vs. press only, by year

51.04% of the 288 facts have a primary source.

2019 · Primary: 12022 · Primary: 12023 · Primary: 32024 · Primary: 62025 · Primary: 242026 · Primary: 1122024 · Press: 32025 · Press: 32026 · Press: 135

n = 288 · as of 2 Oct 2026 · counts observatory entries · See in Evidence ↗

Facts by month, last 24 months

278 facts recorded in the last 24 months.

Nov 2024: 1Dec 2024: 3Jan 2025: 0Feb 2025: 1Mar 2025: 1Apr 2025: 2May 2025: 2Jun 2025: 0Jul 2025: 5Aug 2025: 2Sep 2025: 4Oct 2025: 3Nov 2025: 3Dec 2025: 4Jan 2026: 4Feb 2026: 10Mar 2026: 2Apr 2026: 7May 2026: 2Jun 2026: 11Jul 2026: 16Aug 2026: 10Sep 2026: 180Oct 2026: 5

n = 278 · as of 2 Oct 2026 · counts observatory entries · See in Evidence ↗

Indicators

Incidents per year in the AI Incident Database

Latest measurement: 212 (year to date) incidentes, Sep 2026.

Incidents per year in the AI Incident Database: 212 (year to date) incidentes

n = 13 · as of 7 Sep 2026 · See in Indicators ↗

FrontierMath tier 4 saturation

Latest measurement: 97.6 % de acierto, Sep 2026.

FrontierMath tier 4 saturation: 97.6 % de acierto

n = 8 · as of 3 Sep 2026 · See in Indicators ↗

GPQA Diamond saturation

Latest measurement: 95.8 % de acierto, Sep 2026.

GPQA Diamond saturation: 95.8 % de acierto

n = 7 · as of 3 Sep 2026 · See in Indicators ↗

SWE-bench Verified saturation

Latest measurement: 83.5 % de acierto, Apr 2026.

SWE-bench Verified saturation: 83.5 % de acierto

n = 7 · as of 16 Apr 2026 · See in Indicators ↗

AI adoption by US firms (Census BTOS)

Latest measurement: 22.4 % de empresas, Aug 2026.

AI adoption by US firms (Census BTOS): 22.4 % de empresas

n = 5 · as of 9 Aug 2026 · See in Indicators ↗

Largest known training compute per year

Latest measurement: 3.87 × 10²⁵ (year to date) FLOP, Sep 2026.

Largest known training compute per year: 3.87 × 10²⁵ (year to date) FLOP

n = 8 · as of 9 Sep 2026 · See in Indicators ↗

Companies with a published frontier safety framework

Latest measurement: 12 empresas, Dec 2025.

Companies with a published frontier safety framework: 12 empresas

n = 1 · as of 1 Dec 2025 · See in Indicators ↗

Task time horizon (METR)

Latest measurement: 17.41 horas, Apr 2026.

Task time horizon (METR): 17.41 horas

n = 19 · as of 7 Apr 2026 · See in Indicators ↗

AI adoption measured by transactions (Ramp AI Index)

Latest measurement: 56.13 % de empresas del panel, Aug 2026.

AI adoption measured by transactions (Ramp AI Index): 56.13 % de empresas del panel

n = 7 · as of 1 Aug 2026 · See in Indicators ↗

Generative AI services registered in China

Latest measurement: 1,112 servicios, Aug 2026.

Generative AI services registered in China: 1,112 servicios

n = 3 · as of 31 Aug 2026 · See in Indicators ↗

Youth employment decline in AI-exposed occupations

Latest measurement: -4.4 % anual desde noviembre de 2022, Aug 2026.

Youth employment decline in AI-exposed occupations: -4.4 % anual desde noviembre de 2022

n = 2 · as of 18 Aug 2026 · See in Indicators ↗

Experts

p(doom) with no horizon, by stance

The stated median is 12.5%, across 6 estimates that give a figure; 9 decline to quantify.

Dario Amodei · Individual opinion: 25%Paul Christiano · Individual opinion: 22%ESPAI 2023 (Grace et al.) · Survey: 5%ESPAI 2023 (Grace et al.) · Survey: 5%ESPAI 2023 (Grace et al.) · Survey: 10%Geoffrey Hinton · Individual opinion: 10%–20%

n = 6 · as of 9 Sep 2026 · counts observatory entries · See in Experts ↗

Arrival of general AI, by publication date

7 estimates published between 2023 and 2026; 0 decline to quantify.

Yoshua Bengio: 2028–2043 (Jun 2023)ESPAI 2023 (Grace et al.): 2047 (Oct 2023)Huang Tiejun: 2045 (Sep 2025)Daniel Kokotajlo: Dec 2030 (Jan 2026)Eli Lifland: 2035 (Jan 2026)Encuesta del Summit on Existential Security: 2033 (median; p25–p75: 2031–2036) (Feb 2026)Zhang Ya-Qin: 2039–2044 (Jun 2024)

n = 7 · as of 1 Feb 2026 · counts observatory entries · See in Experts ↗

Risk map

Risks: severity × horizon

33 risks, cross-tabulated by severity and horizon.

Localised · Already happening: 2Localised · 1–3 years: 0Localised · 3–10 years: 0Localised · More than 10 years: 0Severe · Already happening: 5Severe · 1–3 years: 4Severe · 3–10 years: 0Severe · More than 10 years: 0Catastrophic · Already happening: 3Catastrophic · 1–3 years: 7Catastrophic · 3–10 years: 2Catastrophic · More than 10 years: 0Irreversible · Already happening: 0Irreversible · 1–3 years: 0Irreversible · 3–10 years: 8Irreversible · More than 10 years: 0Existential · Already happening: 0Existential · 1–3 years: 0Existential · 3–10 years: 2Existential · More than 10 years: 0

n = 33 · as of 2 Oct 2026 · counts observatory entries · See in Risk map ↗

Risks by evidence level

“Projected” is the most common level, with 16 of 33 risks.

Observed: 9Lab: 4Projected: 16Speculative: 4

n = 33 · as of 2 Oct 2026 · counts observatory entries · See in Risk map ↗

Protection

Measures: cost × effort

3 of 19 measures are marked as theatre.

No cost · Minutes: 1No cost · Hours: 0No cost · Days: 0No cost · Ongoing: 1Low cost · Minutes: 1Low cost · Hours: 7Low cost · Days: 1Low cost · Ongoing: 0Medium cost · Minutes: 0Medium cost · Hours: 2Medium cost · Days: 1Medium cost · Ongoing: 0High cost · Minutes: 0High cost · Hours: 1High cost · Days: 0High cost · Ongoing: 4

n = 19 · as of 2 Oct 2026 · counts observatory entries · See in Protection ↗

Coverage by scenario

11 of 81 measure-scenario pairs work without reservations.

Catastrophe through misuse · Works: 5Power grab by a small group · Works: 1AI as normal technology · Works: 5Catastrophe through misuse · Partial: 11Gradual disempowerment · Partial: 3Power grab by a small group · Partial: 3The intelligence curse · Partial: 3Rapid loss of control · Partial: 5AI as normal technology · Partial: 9Catastrophe through misuse · Does not work: 2Gradual disempowerment · Does not work: 10Power grab by a small group · Does not work: 9The intelligence curse · Does not work: 5Rapid loss of control · Does not work: 6AI as normal technology · Does not work: 2Gradual disempowerment · Backfires: 1The intelligence curse · Backfires: 1

n = 81 · as of 2 Oct 2026 · counts observatory entries · See in Protection ↗

Sources

Sources by year of publication

2026 has the highest count, with 710 of 1146 sources.

1996: 11997: 01998: 01999: 12000: 02001: 02002: 02003: 02004: 02005: 02006: 02007: 02008: 12009: 02010: 02011: 12012: 32013: 52014: 12015: 62016: 12017: 42018: 52019: 32020: 42021: 82022: 182023: 392024: 992025: 2362026: 710

n = 1146 · as of 2 Oct 2026 · counts observatory entries · See in Sources ↗

Sources by type

“Press” is the most frequent type, with 342 of 1146 sources.

Paper: 96Preprint: 75Report: 126System card: 18Survey: 7Book: 7Official document: 220Press: 342Interview: 12Forum post: 53Social media: 11Institutional blog: 107Dataset: 21Web: 51

n = 1146 · as of 2 Oct 2026 · counts observatory entries · See in Sources ↗

Verification status

1090 of 1146 sources are checked through some route.

Sources · Direct HTTP: 1002Sources · Authorised route: 88Sources · Archived copy only: 51Sources · Could not be checked: 5

n = 1146 · as of 2 Oct 2026 · counts observatory entries · See in Sources ↗

Sources by language

“English” is the most frequent language, with 949 of 1146 sources.

Spanish: 59English: 949Other: 138

n = 1146 · as of 2 Oct 2026 · counts observatory entries · See in Sources ↗

OGERIA

Changelog entries by week

The week of 14 Sep 2026 has the highest count, with 14 of 44 entries.

7 Sep 2026: 814 Sep 2026: 1421 Sep 2026: 928 Sep 2026: 13

n = 44 · as of 2 Oct 2026 · counts observatory entries · See in Changelog ↗

View as table

Facts by year, stacked by vector

2026 has the highest count, with 247 of 288 facts.

Biological and CBRNCybersecurity and infrastructureMilitary and autonomous weaponsEconomic and laborPolitical and power concentrationEpistemic and informationLoss of control
20190000100
20200000000
20210000000
20220000001
20231100112
20240220114
2025432611412
20261569262612526137

An event with several vectors counts once in each.

n = 288 · as of 2 Oct 2026 · counts observatory entries · See in Evidence ↗

Facts by type

“Policy” is the most frequent type, with 122 of 288 facts.

Incident50
Evaluation38
Warning36
Policy122
Framework28
Model14

n = 288 · as of 2 Oct 2026 · counts observatory entries · See in Evidence ↗

Primary source vs. press only, by year

51.04% of the 288 facts have a primary source.

PrimaryPress
201910
202000
202100
202210
202330
202463
2025243
2026112135

n = 288 · as of 2 Oct 2026 · counts observatory entries · See in Evidence ↗

Facts by month, last 24 months

278 facts recorded in the last 24 months.

Nov 20241
Dec 20243
Jan 20250
Feb 20251
Mar 20251
Apr 20252
May 20252
Jun 20250
Jul 20255
Aug 20252
Sep 20254
Oct 20253
Nov 20253
Dec 20254
Jan 20264
Feb 202610
Mar 20262
Apr 20267
May 20262
Jun 202611
Jul 202616
Aug 202610
Sep 2026180
Oct 20265

n = 278 · as of 2 Oct 2026 · counts observatory entries · See in Evidence ↗

Incidents per year in the AI Incident Database

Latest measurement: 212 (year to date) incidentes, Sep 2026.

201413 incidentes
201524 incidentes
201641 incidentes
201750 incidentes
201846 incidentes
201944 incidentes
202091 incidentes
202180 incidentes
2022106 incidentes
2023174 incidentes
2024299 incidentes
2025449 incidentes
Sep 2026212 incidentes (year to date)

n = 13 · as of 7 Sep 2026 · See in Indicators ↗

FrontierMath tier 4 saturation

Latest measurement: 97.6 % de acierto, Sep 2026.

Jan 20250 % de acierto
Apr 20254.9 % de acierto
Aug 202522 % de acierto
Dec 202546 % de acierto
Mar 202658.5 % de acierto
Apr 202678 % de acierto
Jun 202690.2 % de acierto
Sep 202697.6 % de acierto

n = 8 · as of 3 Sep 2026 · See in Indicators ↗

GPQA Diamond saturation

Latest measurement: 95.8 % de acierto, Sep 2026.

Mar 202335.7 % de acierto
Jun 202454 % de acierto
Dec 202476.8 % de acierto
Mar 202583.8 % de acierto
Nov 202592.6 % de acierto
Feb 202694.4 % de acierto
Sep 202695.8 % de acierto

n = 7 · as of 3 Sep 2026 · See in Indicators ↗

SWE-bench Verified saturation

Latest measurement: 83.5 % de acierto, Apr 2026.

Nov 202431 % de acierto
Feb 202561 % de acierto
May 202570.7 % de acierto
Aug 202573.6 % de acierto
Nov 202576.7 % de acierto
Feb 202678.7 % de acierto
Apr 202683.5 % de acierto

n = 7 · as of 16 Apr 2026 · See in Indicators ↗

AI adoption by US firms (Census BTOS)

Latest measurement: 22.4 % de empresas, Aug 2026.

Nov 202517.3 % de empresas
Dec 202517.7 % de empresas
Apr 202619.8 % de empresas
Jul 202621.8 % de empresas
Aug 202622.4 % de empresas

n = 5 · as of 9 Aug 2026 · See in Indicators ↗

Largest known training compute per year

Latest measurement: 3.87 × 10²⁵ (year to date) FLOP, Sep 2026.

20191.08 × 10²³ FLOP
20203.14 × 10²³ FLOP
20211.7 × 10²⁴ FLOP
20222.74 × 10²⁴ FLOP
20235 × 10²⁵ FLOP
20243.8 × 10²⁵ FLOP
20255 × 10²⁶ FLOP
Sep 20263.87 × 10²⁵ FLOP (year to date)

n = 8 · as of 9 Sep 2026 · See in Indicators ↗

Companies with a published frontier safety framework

Latest measurement: 12 empresas, Dec 2025.

Dec 202512 empresas

n = 1 · as of 1 Dec 2025 · See in Indicators ↗

Task time horizon (METR)

Latest measurement: 17.41 horas, Apr 2026.

Feb 20191.7 × 10⁻³ horas
May 20201.7 × 10⁻³ horas
Mar 20220.01 horas
Mar 20230.07 horas
May 20240.12 horas
Jun 20240.19 horas
Sep 20240.34 horas
Dec 20240.65 horas
Feb 20251.01 horas
Apr 20252 horas
May 20251.67 horas
Aug 20253.38 horas
Nov 20253.74 horas
Nov 20254.88 horas
Dec 20255.87 horas
Feb 202611.98 horas
Feb 20266.4 horas
Mar 20265.7 horas
Apr 202617.41 horas

n = 19 · as of 7 Apr 2026 · See in Indicators ↗

AI adoption measured by transactions (Ramp AI Index)

Latest measurement: 56.13 % de empresas del panel, Aug 2026.

Jan 20237.46 % de empresas del panel
Dec 202331.68 % de empresas del panel
Dec 202436.66 % de empresas del panel
Jun 202542.72 % de empresas del panel
Dec 202545.92 % de empresas del panel
Apr 202652.98 % de empresas del panel
Aug 202656.13 % de empresas del panel

n = 7 · as of 1 Aug 2026 · See in Indicators ↗

Generative AI services registered in China

Latest measurement: 1,112 servicios, Aug 2026.

Dec 2025748 servicios
Jun 2026988 servicios
Aug 20261,112 servicios

n = 3 · as of 31 Aug 2026 · See in Indicators ↗

Youth employment decline in AI-exposed occupations

Latest measurement: -4.4 % anual desde noviembre de 2022, Aug 2026.

Jul 2026-4.5 % anual desde noviembre de 2022
Aug 2026-4.4 % anual desde noviembre de 2022

n = 2 · as of 18 Aug 2026 · See in Indicators ↗

p(doom) with no horizon, by stance

The stated median is 12.5%, across 6 estimates that give a figure; 9 decline to quantify.

WhoFigureStanceDate
Dario Amodei25%ConcernedSep 2025
Paul Christiano22%ConcernedApr 2023
ESPAI 2023 (Grace et al.)5%Middle groundOct 2023
ESPAI 2023 (Grace et al.)5%Middle groundOct 2023
ESPAI 2023 (Grace et al.)10%Middle groundOct 2023
Geoffrey Hinton10%–20%ConcernedJun 2025

n = 6 · as of 9 Sep 2026 · counts observatory entries · See in Experts ↗

Arrival of general AI, by publication date

7 estimates published between 2023 and 2026; 0 decline to quantify.

WhoFigureStanceDate
Yoshua Bengio2028–2043ConcernedJun 2023
ESPAI 2023 (Grace et al.)2047Middle groundOct 2023
Huang Tiejun2045Middle groundSep 2025
Daniel KokotajloDec 2030ConcernedJan 2026
Eli Lifland2035ConcernedJan 2026
Encuesta del Summit on Existential Security2033 (median; p25–p75: 2031–2036)ConcernedFeb 2026
Zhang Ya-Qin2039–2044Middle groundJun 2024

n = 7 · as of 1 Feb 2026 · counts observatory entries · See in Experts ↗

Risks: severity × horizon

33 risks, cross-tabulated by severity and horizon.

Already happening1–3 years3–10 yearsMore than 10 years
Localised2000
Severe5400
Catastrophic3720
Irreversible0080
Existential0020

n = 33 · as of 2 Oct 2026 · counts observatory entries · See in Risk map ↗

Risks by evidence level

“Projected” is the most common level, with 16 of 33 risks.

Observed9
Lab4
Projected16
Speculative4

n = 33 · as of 2 Oct 2026 · counts observatory entries · See in Risk map ↗

Measures: cost × effort

3 of 19 measures are marked as theatre.

MinutesHoursDaysOngoing
No cost1001
Low cost1710
Medium cost0210
High cost0104

n = 19 · as of 2 Oct 2026 · counts observatory entries · See in Protection ↗

Coverage by scenario

11 of 81 measure-scenario pairs work without reservations.

WorksPartialDoes not workBackfires
Catastrophe through misuse51120
Gradual disempowerment03101
Power grab by a small group1390
The intelligence curse0351
Rapid loss of control0560
AI as normal technology5920

n = 81 · as of 2 Oct 2026 · counts observatory entries · See in Protection ↗

Sources by year of publication

2026 has the highest count, with 710 of 1146 sources.

19961
19970
19980
19991
20000
20010
20020
20030
20040
20050
20060
20070
20081
20090
20100
20111
20123
20135
20141
20156
20161
20174
20185
20193
20204
20218
202218
202339
202499
2025236
2026710

n = 1146 · as of 2 Oct 2026 · counts observatory entries · See in Sources ↗

Sources by type

“Press” is the most frequent type, with 342 of 1146 sources.

Paper96
Preprint75
Report126
System card18
Survey7
Book7
Official document220
Press342
Interview12
Forum post53
Social media11
Institutional blog107
Dataset21
Web51

n = 1146 · as of 2 Oct 2026 · counts observatory entries · See in Sources ↗

Verification status

1090 of 1146 sources are checked through some route.

Direct HTTPAuthorised routeArchived copy onlyCould not be checked
Sources100288515

n = 1146 · as of 2 Oct 2026 · counts observatory entries · See in Sources ↗

Sources by language

“English” is the most frequent language, with 949 of 1146 sources.

Spanish59
English949
Other138

n = 1146 · as of 2 Oct 2026 · counts observatory entries · See in Sources ↗

Changelog entries by week

The week of 14 Sep 2026 has the highest count, with 14 of 44 entries.

7 Sep 20268
14 Sep 202614
21 Sep 20269
28 Sep 202613

n = 44 · as of 2 Oct 2026 · counts observatory entries · See in Changelog ↗

Ask OGERIA

It answers only with what the observatory publishes and can be wrong: check the entries it cites. Your questions are sent to an AI model, so don't write personal data. More in the privacy policy.

Up to 500 characters.

Support OGERIA on Ko-fi

The payment is processed by Ko-fi, not by this site. Open on ko-fi.com