Reference
AI/SI Factbook
FACT-001Draftv0.1.0Reviewed 2026-10-15
This is a table-based reference of verifiable facts about artificial intelligence: historical milestones, landmark benchmark and competition results, standard definitions and measurement conventions, major awards, and key statutes and instruments with their status as of 2026-10-15. It favors few rows with full provenance over many rows with partial provenance. The chronological companion is TIME-001.
#How to read the tables
Every row has the same five elements.
| Element | Meaning |
|---|---|
| Source | The page or record the fact was read from (DOI, arXiv ID, statute citation or URL). "Archived copy" means the live page blocked automated access and a web-archive copy was read. |
| Date | The date of the event or of the legal act, at the precision the source supports. |
| Unit | The quantity and unit of any number, or "date", "text", "status" where the row is not numeric. |
| Limitation | What the row does not establish: who reported it, the conditions, what could not be checked. |
Facts about legal status are as of 2026-10-15 unless stated. Benchmark results and vendor-reported figures are claims by the reporting organisation unless the row says they were independently graded. Values that change quickly (model sizes, prices, scores, chip specifications) are not tabulated here; see MODEL-001.
#1. Historical milestones
| ID | Fact | Date | Unit | Source | Limitation |
|---|---|---|---|---|---|
| H-01 | J. McCarthy (Dartmouth College), M. L. Minsky (Harvard University), N. Rochester (I.B.M. Corporation) and C. E. Shannon (Bell Telephone Laboratories) propose "a 2 month, 10 man study of artificial intelligence" for the summer of 1956 at Dartmouth College, Hanover, New Hampshire. | Proposal dated 1955-08-31 | Date; duration "2 month"; "10 man" | https://www-formal.stanford.edu/jmc/history/dartmouth/dartmouth.html | The page is the proposal text. It shows what was planned, not who attended or what was done in 1956. |
| H-02 | Alan Turing's "Computing Machinery and Intelligence" appears in Mind LIX(236), pp. 433-460. | October 1950 | Date; page range | doi:10.1093/mind/LIX.236.433 (Crossref) | Crossref gives the issue month; the paper's day of appearance is not stated. |
| H-03 | Turing's "On Computable Numbers, with an Application to the Entscheidungsproblem", Proc. London Math. Soc. s2-42, pp. 230-265. | 1936-1937 (Crossref issue year 1937) | Year range; page range | doi:10.1112/plms/s2-42.1.230 | The paper was published in parts across two years; Crossref assigns a single issue year. |
| H-04 | Rosenblatt's perceptron paper appears in Psychological Review 65, pp. 386-408. | 1958 | Year; page range | doi:10.1037/h0042519 | The paper describes a probabilistic model; it does not by itself establish what any hardware implementation achieved. |
| H-05 | Rumelhart, Hinton and Williams, "Learning representations by back-propagating errors", Nature 323, pp. 533-536. | 1986-10 | Month; page range | doi:10.1038/323533a0 | Records one influential publication; it does not settle priority for earlier derivations of the same gradient computation. |
| H-06 | IBM's Deep Blue defeats Garry Kasparov 3.5-2.5 in a six-game rematch at the Equitable Center, New York; IBM describes it as the first computer system to defeat a reigning world champion in a match under standard tournament time controls. In the first match, held in Philadelphia in February 1996, Deep Blue won the first game (the match result was not read). | May 1997 | Match score (points; win = 1, draw = 0.5) | https://www.ibm.com/history/deep-blue | Day-level dates of the 1997 games were not read. The IBM page is the developer's own account. |
| H-07 | IBM's Watson (DeepQA) defeats Brad Rutter and Ken Jennings in a televised Jeopardy! contest. | February 2011 | Date | https://www.ibm.com/history/watson-jeopardy | Developer's account; scores and conditions (buzzer, categories) not read. |
| H-08 | In the ILSVRC 2012 classification results, team SuperVision records top-5 error 0.15315 (using extra ImageNet Fall 2011 data) and 0.16422 (supplied training data only); the next listed team, ISI, records 0.26172. | 2012 | Top-5 error rate (fraction of test images whose correct label is not in the five guesses) | https://image-net.org/challenges/LSVRC/2012/results.html; doi:10.1007/s11263-015-0816-y | The results page gives no announcement date. The runner-up gap is 0.1086 absolute (about 10.9 percentage points) against the best SuperVision entry. |
| H-09 | AlphaGo beat the European champion Fan Hui 5-0 (October 2015), then Lee Sedol 4-1 in Seoul, South Korea (March 2016). | Oct 2015; Mar 2016 | Match score (games) | https://deepmind.google/research/breakthroughs/alphago/; doi:10.1038/nature16961 | Developer's account. Fan Hui is described on the page as the reigning three-time European champion; Lee Sedol as winner of 18 world titles. |
| H-10 | OpenAI is announced in the post "Introducing OpenAI". | 2015-12-11 | Date | https://openai.com/index/introducing-openai/ (archived copy) | Announcement date, not a legal incorporation date. |
| H-11 | Google Brain and DeepMind are combined into Google DeepMind. | 2023-04-20 | Date | https://blog.google/technology/ai/april-ai-update/ | Announcement date; completion of integration not stated. |
| H-12 | OpenAI publishes the ChatGPT announcement. | 2022-11-30 | Date | https://openai.com/index/chatgpt/ (archived copy) | Vendor announcement; user counts are not asserted here. |
| H-13 | The AI Safety Summit is held 1-2 November 2023 and the Bletchley Declaration is issued by the attending countries. | 2023-11-01 to 2023-11-02 | Date range | https://www.gov.uk/government/publications/ai-safety-summit-2023-the-bletchley-declaration/the-bletchley-declaration-by-countries-attending-the-ai-safety-summit-1-2-november-2023 | Non-binding political declaration. |
| H-14 | The Paris AI Action Summit statement "on Inclusive and Sustainable Artificial Intelligence for People and the Planet" is published; the page records participants "from over 100 countries". [C: 2026-10-15] |
2025-02-11 | Date; country count as printed | https://www.elysee.fr/en/emmanuel-macron/2025/02/11/statement-on-inclusive-and-sustainable-artificial-intelligence-for-people-and-the-planet | Non-binding. Which states did or did not endorse the statement is not tabulated here. |
#2. Landmark benchmark and competition results
Each row states the conditions under which the result was obtained. A result under one set of conditions does not transfer to another: the rows below differ in whether the problem was translated by humans, whether training data included the benchmark's public set, and whether the grader was independent.
| ID | Result | Date | Unit and conditions | Source | Limitation |
|---|---|---|---|---|---|
| B-01 | ImageNet / ILSVRC 2012. SuperVision: 0.15315 top-5 error with extra training data, 0.16422 with supplied data only; runner-up 0.26172. | 2012 | Top-5 error on the ILSVRC 2012 classification test set; 5 guesses per image. | See H-08 | Results are for the ILSVRC subset of ImageNet, not for the full ImageNet database. Top-5 error is a more lenient metric than top-1. |
| B-02 | Chess: Deep Blue vs Kasparov. 3.5-2.5 for Deep Blue. | May 1997 | Six games at standard tournament time controls, New York; a special-purpose machine. | https://www.ibm.com/history/deep-blue; doi:10.1016/S0004-3702(01)00129-1 | One match; a single result does not establish general superiority. Match conditions and the dispute about them are not covered by the sources read. |
| B-03 | Go: AlphaGo vs Lee Sedol. 4-1 for AlphaGo. | March 2016 | Five-game match, Seoul, South Korea. | https://deepmind.google/research/breakthroughs/alphago/ | Developer's account. Rules/time controls are not given on the page read. Fan Hui match (5-0) was against a lower-ranked professional. |
| B-04 | IMO 2024. AlphaProof and AlphaGeometry 2 solved 4 of 6 problems for 28 points, a silver-medal standard. | 2024-07-25 (report) | Points out of 42 (6 problems at 7 points). Problems were manually translated into the formal language Lean; the systems took from minutes up to three days per problem, versus 2 sessions of 4.5 hours for students. | https://deepmind.google/discover/blog/ai-solves-imo-problems-at-silver-medal-level/ | Not equal-time or end-to-end natural language. Reported by the developer. |
| B-05 | IMO 2025. An advanced version of Gemini with Deep Think solved 5 of 6 problems for 35 of 42 points, a gold-medal score. [C: 2026-10-15] |
2025-07-21 (report) | Points out of 42. End-to-end in natural language within the 4.5-hour time limit; results officially graded and certified by IMO coordinators using the same criteria as for students. | https://deepmind.google/discover/blog/advanced-version-of-gemini-with-deep-think-officially-achieves-gold-medal-standard-at-the-international-mathematical-olympiad/ | Developer's account of the grading; the grader's own statement was quoted on the developer's page, not read independently. |
| B-06 | ARC-AGI-1 semi-private set: OpenAI o3. 75.7% at the public-leaderboard compute limit (USD 10,000) and 87.5% in a high-compute configuration. | 2024-12-20 | Percentage of tasks solved on a 100-task Semi-Private Evaluation set; o3 was trained on the ARC-AGI-1 public training set; high-compute sample count and cost are given on the source page (1,024 samples in the 87.5% configuration, retail cost USD 456,000 for the set as shown). | https://arcprize.org/blog/oai-o3-pub-breakthrough | The source page carries a note dated 3/24/2025 that the released o3 differs from the tested version. Score depends on compute budget and training exposure; it should not be read as "human level" without the benchmark's own caveats. |
| B-07 | Protein structure: AlphaFold. "Highly accurate protein structure prediction with AlphaFold", Nature 596, 583-589. | 2021-07-15 | Published paper (accuracy claims are in the paper, not restated here). | doi:10.1038/s41586-021-03819-2 | The paper's CASP14 numbers were not extracted for this row; consult the paper. |
| B-08 | Poker: Libratus. "Superhuman AI for heads-up no-limit poker: Libratus beats top professionals", Science 359, 418-424. | 2018-01-26 | Heads-up (two-player) no-limit poker only. | doi:10.1126/science.aao1733 | Results do not extend to multi-player poker (see doi:10.1126/science.aay2400 for the multiplayer paper). Match size and win rate were not extracted. |
| B-09 | Introduction of benchmarks. MMLU (arXiv:2009.03300, 2020-09-07), GPQA (arXiv:2311.12022, 2023-11-20), Humanity's Last Exam (arXiv:2501.14249, 2025-01-24), ARC-AGI-2 (arXiv:2505.11831, 2025-05-17), SWE-bench (arXiv:2310.06770, 2023-10-10). | As listed | Date of arXiv v1 | arXiv API records | Dates are for introduction of the benchmark, not for any score. Current scores belong in MODEL-001. |
#3. Standard definitions and measurement conventions
| ID | Term and convention | Date | Unit | Source | Limitation |
|---|---|---|---|---|---|
| D-01 | FLOP (a count) vs FLOP/s (a rate). A FLOP is one floating-point operation; a quantity of training compute is a count of operations, and the speed of a computer is a count per second. The 2023 US Executive Order 14110 uses both in one sentence: a reporting threshold of "a quantity of computing power greater than 10^26 integer or floating-point operations" for a trained model, and for a data-center cluster "a theoretical maximum computing capacity of 10^20 integer or floating-point operations per second". | 2023-10-30 | FLOP (operations); FLOP/s (operations per second) | https://www.govinfo.gov/content/pkg/FR-2023-11-01/html/2023-24283.htm (section 4.2(b)) | EO 14110 was revoked (see S-03). The informal acronym "FLOPS" is used for both quantities in the literature; always state the unit and whether it is peak (theoretical) or achieved. Legal text counts "integer or floating-point operations", which is broader than FLOP alone. |
| D-02 | Petaflop/s-day. Kaplan et al. quote training compute in PF-days, where one PF-day = 10^15 x 24 x 3600 = 8.64 x 10^19 floating-point operations. | 2020-01-23 | FLOP (count) | arXiv:2001.08361 (read via ar5iv.labs.arxiv.org/html/2001.08361) | A unit of work (operations): one petaFLOP/s (10^15 operations per second) sustained for one day. Do not confuse with a rate. |
| D-03 | Approximate training compute for a dense Transformer. Kaplan et al. estimate non-embedding training compute as C ~ 6NBS, where N is the number of parameters excluding all vocabulary and positional embeddings, B is batch size and S is the number of parameter updates; the factor of 6 accounts for the forward and backward passes. | 2020-01-23 | FLOP (count) | arXiv:2001.08361 | An approximation for dense Transformers that ignores attention-length terms and embedding compute; not valid as stated for sparse (mixture-of-experts) models, where N must be the active parameters. |
| D-04 | Estimating training compute from reports. Sevilla et al. estimate training compute for a model "from the forward pass compute, and from GPU time", following the method of OpenAI's "AI and Compute" analysis, and state that they expect most estimates to be "accurate within a factor of about two". | 2022-02-11 | FLOP (count); factor | arXiv:2202.05924 (read via ar5iv) | An uncertainty claim by the authors for their dataset; undisclosed training runs cannot be estimated this way. |
| D-05 | Parameter-counting convention (embeddings). Kaplan et al. define N as parameters "excluding all vocabulary and positional embeddings". Other papers state totals without this exclusion. | 2020-01-23 | Parameters (count) | arXiv:2001.08361 | When comparing model sizes, state whether embeddings are included; the difference is large for small models with large vocabularies. |
| D-06 | Parameter-counting convention (dense vs sparse). GPT-3 has 175 billion parameters, described by its authors as "10x more than any previous non-sparse language model". Mixtral 8x7B: "each token has access to 47B parameters, but only uses 13B active parameters during inference". DeepSeek-V3: "671B total parameters with 37B activated for each token". | 2020-05-28; 2024-01-08; 2024-12-27 | Parameters (count): total vs active per token | arXiv:2005.14165; arXiv:2401.04088; arXiv:2412.19437 | Total and active parameters are different quantities; quote both for mixture-of-experts models. Parameter counts of closed models are generally undisclosed and are not tabulated. |
| D-07 | Compute-optimal training (Chinchilla). "For compute-optimal training, the model size and the number of training tokens should be scaled equally: for every doubling of model size the number of training tokens should also be doubled." Chinchilla has 70B parameters trained on 1.4 trillion tokens, with the same compute budget as Gopher (280B parameters, 300 billion tokens). | 2022-03-29 | Parameters; tokens (count) | arXiv:2203.15556 | An empirical result for the authors' setup, fitted over models of 70 million to over 16 billion parameters and 5 to 500 billion tokens; extrapolation to other regimes is not guaranteed. The paper contrasts its result with the earlier fit of Kaplan et al. |
| D-08 | Legal compute threshold, EU. A general-purpose AI model is presumed to have high-impact capabilities when "the cumulative amount of computation used for its training measured in floating point operations is greater than 10^25" (AI Act, Article 51(2)). | Regulation of 2024-06-13 | FLOP (cumulative training compute) | https://eur-lex.europa.eu/eli/reg/2024/1689/oj (Official Journal text, Art. 51(2), read 2026-10-18) | The Commission may amend the thresholds by delegated act (Article 51(3)); check the current text. A presumption, not a definition of risk. |
| D-09 | Legal compute threshold, California. SB 53 defines "frontier model" as "a foundation model that was trained using a quantity of computing power greater than 10^26 integer or floating-point operations", counting the original training run and later fine-tuning, reinforcement learning or other material modifications. | Approved 2025-09-29 | Operations (cumulative) | Bill text, leginfo.legislature.ca.gov, bill_id=202520260SB53 (archived copy) | The cumulative-including-fine-tuning rule differs from the pretraining-only convention used in much of the research literature. |
| D-10 | ILSVRC top-5 error. The fraction of images whose correct label is not among the five labels a system guesses. | 2012 | Fraction (0-1) | https://image-net.org/challenges/LSVRC/2012/results.html (column "Error (5 guesses)") | Top-1 error is higher. |
#4. Major awards
#4.1 ACM A.M. Turing Award: AI-related laureates
The ACM index lists laureates by award year. Each citation below is quoted from the laureate page or ACM news page named in the Source column. Pages at acm.org block automated fetches; the quoted text was read from archived copies on 2026-10-15.
| Award year | Laureate(s) | Citation as published by ACM | Source | Limitation |
|---|---|---|---|---|
| 1969 | Marvin Minsky | "For his central role in creating, shaping, promoting, and advancing the field of Artificial Intelligence." | amturing.acm.org laureate page (Minsky) and byyear index (archived copies) | The laureate page's honors list gives "Turing Award ... (1970)", the year it was presented; ACM's index assigns award year 1969. |
| 1971 | John McCarthy | ACM's page cites his lecture, "The Present State of Research on Artificial Intelligence", and no separate citation sentence. | amturing.acm.org laureate page (McCarthy) (archived copy) | ACM's published citation for this year is not a conventional citation sentence. |
| 1975 | Allen Newell and Herbert A. Simon | "In joint scientific efforts extending over twenty years, initially in collaboration with J. C. Shaw at the RAND Corporation, and subsequentially with numerous faculty and student colleagues at Carnegie-Mellon University, Newell and co-recipient Herbert A. Simon made basic contributions to artificial intelligence, the psychology of human cognition, and list processing." | amturing.acm.org laureate page (Newell) (archived copy) | The Simon page was not read separately; the citation is joint. |
| 1994 | Edward A. Feigenbaum and Dabbala Rajagopal (Raj) Reddy | "For pioneering the design and construction of large scale artificial intelligence systems, demonstrating the practical importance and potential commercial impact of artificial intelligence technology." | amturing.acm.org laureate page (Feigenbaum) (archived copy) | The Reddy page was not read separately; the citation is joint. |
| 2010 | Leslie G. Valiant | "For transformative contributions to the theory of computation, including the theory of probably approximately correct (PAC) learning, the complexity of enumeration and of algebraic computation, and the theory of parallel and distributed computing." | amturing.acm.org laureate page (Valiant) (archived copy) | Included because PAC learning underpins the theory of machine learning; the citation covers more than learning. |
| 2011 | Judea Pearl | "For fundamental contributions to artificial intelligence through the development of a calculus for probabilistic and causal reasoning." | amturing.acm.org laureate page (Pearl) (archived copy) | None beyond the archived-copy status. |
| 2018 | Yoshua Bengio, Geoffrey Hinton, Yann LeCun | "For conceptual and engineering breakthroughs that have made deep neural networks a critical component of computing." (ACM news text: "recipients of the 2018 ACM A.M. Turing Award for ...") | https://awards.acm.org/about/2018-turing (archived copy) | The quoted words come from ACM's announcement text, not from a separate citation field. |
| 2024 | Andrew G. Barto and Richard S. Sutton | "For developing the conceptual and algorithmic foundations of reinforcement learning." | https://awards.acm.org/about/2024-turing (archived copy) | The quoted words come from ACM's announcement text. |
Not AI-related (listed to prevent error). The 2025 ACM A.M. Turing Award went to Charles H. Bennett and Gilles Brassard "for their essential role in establishing the foundations of quantum information science and transforming secure communication and computing" (https://awards.acm.org/about/2025-turing, archived copy, read 2026-10-15). The most recent AI-related Turing Award as of this date is the 2024 award. The 1995 laureate Manuel Blum is cited for computational complexity theory and cryptography and is not listed above as AI-related.
#4.2 Nobel Prizes, 2024
| Prize | Laureates and citation | Announcement date | Source | Limitation |
|---|---|---|---|---|
| Physics | John J. Hopfield (Princeton University, NJ, USA) and Geoffrey Hinton (University of Toronto, Canada), "for foundational discoveries and inventions that enable machine learning with artificial neural networks". | 8 October 2024 | https://www.nobelprize.org/prizes/physics/2024/press-release/ | The prize is in Physics, awarded for work on artificial neural networks; it is not a prize "for AI" in general. |
| Chemistry | One half to David Baker (University of Washington, Seattle, USA; Howard Hughes Medical Institute) "for computational protein design"; the other half jointly to Demis Hassabis and John Jumper (both Google DeepMind, London, UK) "for protein structure prediction". | 9 October 2024 | https://www.nobelprize.org/prizes/chemistry/2024/press-release/ | Affiliations are as printed in the press release at the time of the prize. |
#5. Key statutes and instruments, status as of 2026-10-15
"Status" is the legal state found in the sources read on 2026-10-15. Where a source was a law-firm summary or news report rather than an official text, the row says so. Binding instruments are listed before voluntary ones.
#5.1 Binding instruments
| ID | Instrument | Date | Status as of 2026-10-15 | Source | Unit / key parameter | Limitation |
|---|---|---|---|---|---|---|
| S-01 | Regulation (EU) 2024/1689 (AI Act) of 13 June 2024. | Published in the Official Journal 2024-07-12; in force 2024-08-01 | In force. Application dates: Chapters I and II (prohibitions, AI literacy) from 2 February 2025, except new Article 5(1)(ba), (bb), (1a), (1b) from 2 December 2026; general-purpose AI model obligations from 2 August 2025; high-risk obligations of Chapter III, Sections 1-3 from 2 December 2027 (Annex III systems) and 2 August 2028 (Annex I systems), per Regulation (EU) 2026/1744. Providers of generative systems placed on the market before 2 August 2026 have until 2 December 2026 to comply with Article 50(2). | https://eur-lex.europa.eu/eli/reg/2024/1689/oj (Art. 113, original dates); http://data.europa.eu/eli/reg/2026/1744/oj (amending text); Official Journal texts | Dates (see status); penalties not tabulated | The implementation-timeline page predates the Omnibus and still shows 2 August 2026 for the high-risk rules; the Omnibus text controls. The application date of the AI Act's other provisions (for example the general Article 50 transparency duties and penalties) under Article 113 was not re-verified against a consolidated text. [VERIFICATION NEEDED: consolidated application dates after Regulation (EU) 2026/1744] Secondary reproduction of the Official Journal; EUR-Lex pages could not be fetched. |
| S-02 | Regulation (EU) 2026/1744 (Digital Omnibus on AI), amending Regulation (EU) 2024/1689 and Regulations (EU) 2018/1139 and (EU) 2023/1230. [C: 2026-10-15] |
Council decision 2026-06-29; entered into force 2026-07-27 | In force since 27 July 2026. Its text provides entry into force on the third day after publication in the Official Journal. | http://data.europa.eu/eli/reg/2026/1744/oj | Postponed dates as in S-01 | Signed 8 July 2026 and published in the Official Journal on 24 July 2026 (Official Journal text read 2026-10-15); in force on the third day after publication. |
| S-03 | US Executive Order 14110 (88 FR 75191), signed 2023-10-30. | 2023-10-30 | Revoked on 2025-01-20 by Executive Order 14148 (90 FR 8237). | https://www.govinfo.gov/content/pkg/FR-2025-01-28/html/2025-01901.htm (EO 14148 text lists "Executive Order 14110 of October 30, 2023") | Compute threshold 10^26 operations (see D-01) | Reporting requirements and thresholds of EO 14110 are no longer in force. Some agency actions that implemented it may persist; they are not tabulated. |
| S-04 | US Executive Order 14179, "Removing Barriers to American Leadership in Artificial Intelligence" (90 FR 8741). [C: 2026-10-15] |
2025-01-23 | No revocation found in the sources read. | Federal Register API record (document published 2025-01-31) | Policy direction | An executive order is not a statute; its continued status was not independently checked against later orders. |
| S-05 | US Executive Order 14365, "Ensuring a National Policy Framework for Artificial Intelligence" (90 FR 58499). [C: 2026-10-15] |
2025-12-11 | Issued; no revocation found in the sources read. Directs the Attorney General to establish an AI Litigation Task Force to challenge State AI laws it considers inconsistent with the order's policy. | https://www.govinfo.gov/content/pkg/FR-2025-12-16/html/2025-23092.htm | 30-day deadline for the Task Force | Whether the Task Force has filed suit, and the outcome of any challenge, was not researched. |
| S-06 | California SB 53, Transparency in Frontier Artificial Intelligence Act. [C: 2026-10-15] |
Approved by the Governor 2025-09-29 | Enacted. [VERIFICATION NEEDED: operative date (California statutes generally take effect on 1 January of the following year) and any court challenge] |
Bill text, leginfo.legislature.ca.gov, bill_id=202520260SB53 (archived copy) | "Frontier model": > 10^26 integer or floating-point operations (see D-09) | The text read is the chaptered bill; subsequent amendment was not checked. |
| S-07 | Colorado SB 24-205 (Colorado AI Act). [C: 2026-10-15] |
Effective date moved from 2026-02-01 to 2026-06-30 by SB25B-004 (signed 2025-08-28) | [VERIFICATION NEEDED: status after the 2026 legislative session] A June 2026 source says the framework was subsequently reshaped by further legislation; the enacted text and effective date were not found. |
Law-firm summaries (Greenberg Traurig, Clark Hill) of the 2025 delay; no official text read | Date | Do not rely on this row for compliance. |
| S-08 | Council of Europe Framework Convention on Artificial Intelligence and Human Rights, Democracy and the Rule of Law (CETS No. 225). [C: 2026-10-15] |
Opened for signature 2024-09-05; ratified by the European Union 2026-05-15 | Open for signature and ratification; the European Union ratified on 15 May 2026. [VERIFICATION NEEDED: whether the treaty has entered into force; one secondary source dated 5 October 2026 says it has not] |
https://www.coe.int/en/web/artificial-intelligence/the-framework-convention-on-artificial-intelligence; https://www.coe.int/en/web/artificial-intelligence/-/european-union-ratifies-the-council-of-europe-framework-convention-on-artificial-intelligence (archived copy) | Date | The Treaty Office status table could not be read; the number of ratifications is not stated here. |
| S-09 | China: Interim Measures for the Management of Generative Artificial Intelligence Services. [C: 2026-10-15] |
Effective 2023-08-15 | In effect per law-firm summaries (PwC China, Bird & Bird, Hankun). [VERIFICATION NEEDED: issuing date (reported as 10 or 13 July 2023) and subsequent amendments; Cyberspace Administration text not read] |
https://www.pwccn.com/en/tmt/interim-measures-for-generative-ai-services-implemented-aug2023.pdf | Date | Secondary sources only. |
| S-10 | South Korea: AI Basic Act. [C: 2026-10-15] |
Effective 2026-01-22 | In effect per law-firm and press analysis; fines reported to begin after a grace period. [VERIFICATION NEEDED: official text on law.go.kr not read; English titles vary] |
https://www.cooley.com/news/insight/2026/2026-01-27-south-koreas-ai-basic-act-overview-and-key-takeaways | Date | Secondary sources only. |
#5.2 Voluntary and non-binding instruments
| ID | Instrument | Date | Status as of 2026-10-15 | Source | Limitation |
|---|---|---|---|---|---|
| V-01 | NIST AI Risk Management Framework 1.0. | Released 2023-01-26 | Published; voluntary. | https://www.nist.gov/itl/ai-risk-management-framework | A framework, not a standard of compliance; any later revision was not checked. |
| V-02 | ISO/IEC 42001:2023, "Information technology - Artificial intelligence - Management system". | Edition 1, 2023-12 | Listed as published on the ISO catalog page (archived copy). | https://www.iso.org/standard/81230.html | The status label was read from an archived page, not the live catalog. |
| V-03 | OECD AI Principles. | Adopted May 2019; updated May 2024 | Adopted instruments of the OECD adherents. | https://oecd.ai/en/ai-principles (archived copy) | Non-binding. The OECD legal-instrument page could not be read. |
| V-04 | UNESCO Recommendation on the Ethics of Artificial Intelligence. | November 2021 | Adopted by UNESCO Member States (page text: 193 Member States). | https://www.unesco.org/en/artificial-intelligence/recommendation-ethics (archived copy) | Non-binding recommendation; day of adoption not read. |
| V-05 | Bletchley Declaration. | 2023-11-01 to 2023-11-02 | Issued; political declaration. | gov.uk page named in H-13 | Non-binding. |
| V-06 | General-Purpose AI Code of Practice (EU). [C: 2026-10-15] |
Published 2025-07-10 | Published. The Commission page states the Commission and the AI Board have confirmed it as an adequate voluntary tool for general-purpose AI providers to demonstrate compliance with the AI Act. | https://digital-strategy.ec.europa.eu/en/policies/contents-code-gpai | Voluntary; adherence is by signature of individual providers, which is not tabulated here. |
| V-07 | UN General Assembly resolution A/RES/79/325, establishing the Independent International Scientific Panel on AI and the Global Dialogue on AI Governance. [C: 2026-10-15] |
Adopted 2025-08-26 (by consensus, per search summary) | Adopted. [VERIFICATION NEEDED: adoption date from a search summary of the UN PDF; panel membership and activities in 2026 are reported only by unverified secondary sources and are not listed] |
https://www.un.org/pga/wp-content/uploads/sites/110/2025/12/A_RES_79_325-EN.pdf | General Assembly resolutions are not binding on member states. |
| V-08 | America's AI Action Plan ("Winning the AI Race"). [C: 2026-10-15] |
Released 2025-07-23 | Published policy document. | https://www.whitehouse.gov/articles/2025/07/white-house-unveils-americas-ai-action-plan/ | A policy plan, not a statute or regulation. |
#Open items
- The 1973 Lighthill report, the Samuel/Newell-Simon-era programs and the AI winter dates are not included because no source page was read.
- The Treaty Office status table for CETS 225, the EUR-Lex consolidated AI Act and the UN General Assembly PDF should be read directly at the next review.
- ACM and several vendor sites block automated access; acm.org citations were read from archived copies and should be compared with the live pages.
- No row in this factbook states a current benchmark score or model size. Those belong in
MODEL-001.
#Revision history
v0.1.0 (2026-10-15): Initial draft. Citations checked against arXiv/Crossref/publisher pages where marked; not independently fact-checked.