Laws compared
A cell records what a text requires of the parties it defines. It does not show that any developer falls within them, or complied.
What does each binding government instrument that addresses developers of frontier or general-purpose AI models say on the same questions? Each column is one instrument; each row is one question, worded so that it fits every instrument. The questions are Selekoa’s; an instrument’s own terms appear only in quotation marks. Some instruments address providers of services rather than developers, and the first row quotes the party each one defines.
Question Q02Threshold
What threshold or condition brings a model, or its developer, within those duties?
Criterion: A stated condition, quoted with its figure where there is one: training compute, compute cost, revenue, capability, number of users, or designation by an authority. Stated if a threshold or designation condition is quoted; partly if the instrument covers a broad class (for example, all public services of a kind) without any model-level threshold.
What the statuses mean
- Stated
- A passage of the instrument meets the question’s criterion as worded, however conditional or narrow its scope. The condition or scope is in the note.
- Stated, partly
- A passage has what the criterion asks for, but in a weaker form than the criterion’s own words: for example, a power (“may”) where the criterion asks for a duty (“shall”), or a duty without the deadline the question asks about.
- Not found in the recorded search
- No passage meeting the criterion was found. The cell lists the sections read, the search terms and every passage set aside, with the reason. It does not show that the instrument lacks such a provision elsewhere, or that no other law imposes one.
- Does not apply
- The question does not arise for this instrument, for the reason given (for example, there is no framework document whose publication could be asked about).
Four rules decide “partly” in borderline cases:
- An order to an agency to impose a duty counts as partly.
- Protection only by reference to another law counts as partly.
- “Without undue delay” counts as a stated deadline.
- Required security measures for model weights count as risk-management measures under Q03.
Every stated or partly cell quotes the instrument’s own text with its article, section or page; Chinese passages carry Selekoa’s working translation. Three cells give a paraphrase with its locator instead, and say so. A not-found cell lists the sections read, the terms searched and every passage set aside, with the reason.
Paraphrased with its locator, not quoted: quoting it would pass the source’s quotation limit, or the source’s length is not known.
Limits of this comparison
- A cell records what the text requires of the parties it defines. It does not show that any named developer falls within them, or that anyone complied.
- A not-found cell is bounded by its search: the sections read and the terms searched. Another provision, or another law, may still impose the duty.
- The Code of Practice is a compliance route, not a law. Executive Order 14110 is revoked, and the BIS rule’s rescission was announced; each column shows its standing.
- The table compares and does not rank. Statuses are not scores and are not to be summed.
- Biological and chemical risk provisions are recorded at policy level only.
- Dates keep their kind: adoption, signature, approval, publication, entry into force and application are different events. A computed date is marked as computed.
How we searched
The instruments started from those on the frameworks-and-law page. A column was kept where the instrument is binding, or is the AI Act’s stated means of compliance (the Code), and places duties on developers or providers of frontier or general-purpose AI models. Fifty other instruments on that page were set aside, each with its reason, in the research pack: proposed or vetoed texts, guidance and reports, voluntary commitments, binding texts that place no duty on developers, and one standard whose text was not retrieved (GB 45438-2025).
United Kingdom: within the searches made, no UK law or other binding instrument placing duties on developers of such models was located. This is not a finding that none exists. The 2024 King’s Speech announced an intent to legislate; the 2026 briefing notes contain none of the searched terms; the memoranda with developers state that they are not legally binding; and title searches of legislation.gov.uk for 2023 to 2026 found one instrument, which places a duty on the Information Commissioner.
Other US federal texts: as read, none places a binding duty on developers. Executive Order 14409 of 2 June 2026 provides for a voluntary access framework and says that its section 3 authorises no licensing requirement.
Not yet read: the Republic of Korea’s AI Basic Act, Japan’s AI Promotion Act (2025), China’s provisions on deep synthesis (2022) and on algorithm recommendation (2021), and Colorado SB 24-205 (2024). They are not in the dataset; a later pack starts with Korea.
Illinois: the act’s official website refused the preparer’s requests, and the Internet Archive reset the connections. The cross-family checker’s session saved the enacted text from an Internet Archive capture of the official PDF (2026-09-08), and the cells were drafted from that saved text. The two model families read the act independently and agreed on the status of 13 of 15 cells. The approval date rests on the checker’s copy of the official bill status record, which the preparer could not open; it is given as a paraphrase, labelled as that copy.
Texts were read in official copies: the EU texts through the Publications Office’s repository, because EUR-Lex returned empty responses; California SB 53 from the Legislative Counsel’s database file, because its website refused requests; New York’s chapters in the Senate print and the Assembly’s bill records, not the session laws; the Chinese measures in the gov.cn policy library. Six web pages served changed bytes between retrievals; every quotation still matched.
Roles: one model family (Claude) drafted the cells, audited 50 of them in a fresh context and adjudicated. A model from another family (GPT, through Codex) checked 122 items, answering 85 of them blind before seeing the drafted answers. Where it disagreed, its status was applied. One cell it disputed (Executive Order 14110, Q06) was settled as not found after its second pass. For Illinois the roles were reversed: the checker read the act first, blind, and the drafter then wrote all fifteen cells without seeing its answers. The maintainer decided the remaining questions, including the four rules for “partly” and Illinois Q05 (partly, as for California’s identical wording). No person has checked the cells against the texts.
European Union
EU AI Act
Stated
“presumed to have high impact capabilities pursuant to paragraph 1, point (a), when the cumulative amount of computation used for its training measured in floating point operations is greater than”
Article 51(2)
“greater than 1025”
Article 51(2), closing words
“based on a decision of the Commission, ex officio or following a qualified alert from the scientific panel”
Article 51(1), point (b)
Note: The figure is 10 to the power 25 floating point operations; the cached text renders the superscript as '1025'. The compute figure is a presumption of high-impact capabilities under Article 51(1)(a); a model can also be classified by Commission decision having regard to the Annex XIII criteria. Article 51(3) empowers the Commission to amend the thresholds by delegated act; Regulation (EU) 2026/1744, Article 1, point (37), restates that delegation in Article 97 without changing Article 51. Articles 52(2) to 52(5) provide for rebuttal and reassessment.
Inspect sources and checks: EU AI Act, Q02 ThresholdEU Code of Practice, Safety and Security
Stated, partly
“providers of general-purpose AI models with systemic risk”
Recitals, (b) (PDF p. 3) (part)
Note: The Chapter sets no threshold of its own; its scope is the AI Act category of models with systemic risk, whose conditions are in Article 51 of the Act. Measure 1.1 ties the timing of the Framework to the provider's notification under Article 52(1) of the Act. Search for the rest of the criterion: Whole text: Objectives, Recitals, Commitments 1 to 10 with Measures, Glossary, Appendices 1 to 4. Terms searched: “Article 51”, “Article 52”, “FLOP”, “floating”, “10\^”. Passages read and not counted: Measure 1.1, last paragraph (PDF p. 7): Refers to notification under Article 52(1) AI Act as a timing point; sets no condition of its own. Measure 1.2, point (1)(a) (PDF p. 7): 'Trigger points' (e.g. training compute) for lighter-touch evaluations, set by the Signatory; not a scope threshold.
Inspect sources and checks: EU Code of Practice, Safety and Security, Q02 Threshold
United States: states
California SB 53
Stated
“Frontier model” means a foundation model that was trained using a quantity of computing power greater than 10^26 integer or floating-point operations.”
SEC. 2, Bus. & Prof. Code § 22757.11(i)(1)
“together with its affiliates collectively had annual gross revenues in excess of five hundred million dollars ($500,000,000) in the preceding calendar year.”
SEC. 2, Bus. & Prof. Code § 22757.11(j)
Note: Model threshold: training compute greater than 10^26 operations, which under § 22757.11(i)(2) includes the original training run and subsequent fine-tuning, reinforcement learning or other material modifications. Developer tier: annual gross revenues with affiliates above $500 million in the preceding calendar year. § 22757.14 requires the Department of Technology to recommend, on or before 1 January 2027 and annually thereafter, whether and how to update these definitions; it does not itself change them.
Inspect sources and checks: California SB 53, Q02 ThresholdNew York RAISE Act (As amended, Chapter 96 of 2026)
Stated
"Frontier model" means a foundation model that was trained … using a quantity of computing power greater than”
Bill text, § 2 (GBL § 1420(9)(a) as re-added), p. 3, lines 46–47
"Large frontier developer" means a frontier developer that … collectively had annual gross revenues in excess … of five hundred million dollars in the preceding calendar year”
Bill text, § 2 (GBL § 1420(10) as re-added), p. 3, lines 53–55
Note: The Assembly record renders the figure after “greater than” as an unreadable character between “10” and “26” followed by “integer or floating-point operations”; the Senate PDF of S8828 (cross-check copy, not quoted) prints 10º26, i.e. 10^26. Compute includes subsequent fine-tuning, reinforcement learning and other material modifications (§ 1420(9)(b)). Developer tier: revenues with affiliates above $500 million. § 1425 limits the article to models developed, deployed or operating in whole or in part in New York.
Inspect sources and checks: New York RAISE Act (As amended, Chapter 96 of 2026), Q02 ThresholdNew York RAISE Act (As signed, Chapter 699 of 2025)
Stated
“an artificial intelligence model trained using greater than 10º26 computational operations”
Proposed GBL § 1420(6)(a), p. 2, lines 28–30
“the compute cost of which exceeds one hundred million dollars”
Proposed GBL § 1420(6)(a), p. 2, line 30
"Large developer" means a person that has trained at least one frontier model and has spent over one hundred million dollars in compute costs in aggregate in training frontier models.”
Proposed GBL § 1420(9), p. 3, lines 4–6
Note: The PDF text layer renders the exponent as “10º26”, i.e. 10^26 operations. Model condition: more than 10^26 operations and compute cost above $100 million; a model produced by knowledge distillation from such a model is also a frontier model where its compute cost exceeds $5 million (§ 1420(6)(b)). Developer condition: over $100 million in aggregate compute costs for training frontier models. § 1424 limits the article to frontier models developed, deployed or operating in whole or in part in New York.
Inspect sources and checks: New York RAISE Act (As signed, Chapter 699 of 2025), Q02 ThresholdIllinois AI Safety Measures Act
Stated
“Frontier model” means a foundation model that was trained using a quantity of computing power greater than 10 26 integer or floating-point operations.”
Section 5, definition of “frontier model”, first sentence, PDF p. 4
“Large frontier developer” means a frontier developer that, together with its affiliates, collectively had annual gross revenues in excess of $500,000,000 in the preceding calendar year.”
Section 5, definition of “large frontier developer”, PDF p. 4
Note: Model threshold: training compute greater than 10^26 operations, which under the second sentence of the definition includes the original training run and any subsequent fine-tuning, reinforcement learning or other material modifications the developer applies to a preceding foundation model. Developer threshold for the large tier: gross revenues with affiliates above $500 million in the preceding calendar year. The PDF prints the exponent as a superscript; its text layer gives “10 26”, which the quotation keeps (the figure is 10^26, as with the EU row's “1025”). The cross-family checker, reading the act independently, quoted the same words with the exponent printed as “10²⁶” (maintainer decision 2a of 2026-11-05).
Inspect sources and checks: Illinois AI Safety Measures Act, Q02 Threshold
China
These measures address providers of services, not developers as such.
China, generative AI services (2023)
Stated, partly
“利用生成式人工智能技术向中华人民共和国境内公众提供生成文本、图片、音频、视频等内容的服务（以下称生成式人工智能服务），适用本办法。”
Working translation: These Measures apply to services that use generative AI technology to provide content such as text, images, audio and video to the public within the territory of the People's Republic of China (hereinafter 'generative AI services').
Article 2 (第二条), first paragraph
“未向境内公众提供生成式人工智能服务的，不适用本办法的规定”
Working translation: where generative AI services are not provided to the public within the territory, the provisions of these Measures do not apply
Article 2 (第二条), third paragraph
Note: Coverage turns on providing generative AI services to the public in China; there is no compute, revenue, capability or user-number threshold. Article 2(3) excludes research, development and application by industry organisations, enterprises, education and research institutions and others that do not provide services to the public in China; Article 2(2) defers to other rules for news publishing, film and television production and literary and artistic creation. Article 17 adds a further condition (services 'with public-opinion attributes or social mobilisation capacity') for security assessment and algorithm filing only; see Q05.
Inspect sources and checks: China, generative AI services (2023), Q02 ThresholdChina, AI content labelling (2025)
Stated, partly
“符合…规定情形的网络信息服务提供者（以下简称“服务提供者”）”
Working translation: network information service providers that fall within the circumstances provided for in [the three named rules] (hereinafter 'service providers')
Article 2 (第二条)
Note: Coverage is defined by reference to the circumstances of three other rules (Algorithmic Recommendation Provisions, Deep Synthesis Provisions, Interim Measures for Generative AI Services), with no model-level threshold. The explicit-labelling duty of Article 4 applies where the service falls within Article 17(1) of the Deep Synthesis Provisions.
Inspect sources and checks: China, AI content labelling (2025), Q02 Threshold
Revoked, or rescission announced
Kept as part of the record, each with its standing shown.
US Executive Order 14110
Stated
“Until such technical conditions are defined, the Secretary shall require compliance with these reporting requirements for: (i) any model that was trained using a quantity of computing power greater than”
Sec. 4.2(b)(i), 88 FR 75197
Note: Interim condition, until the Secretary of Commerce defines technical conditions under Sec. 4.2(b): more than 10^26 integer or floating-point operations, or more than 10^23 for a model trained primarily on biological sequence data. The quote stops before the figures, which the PDF text layer renders without superscript (“1026”, “1023”); the reading follows the dataset’s record of this threshold. Sec. 4.2(b)(ii) sets a separate condition for computing clusters.
Inspect sources and checks: US Executive Order 14110, Q02 ThresholdUS BIS AI diffusion rule
Stated
“4E091 ‘Parameters’ for an artificial intelligence model trained utilizing”
ECCN 4E091 heading, 90 FR 4578 (PDF p. 35)
“ECCN 4E091 excludes from the control any open-weight models”
Supplementary information, section A, 90 FR 4555
Note: The heading continues with a figure that the PDF text layer renders as “1026” before “or more ‘operations’”, read as 10^26 operations; the quote stops before it. Notes 1 and 2 to 4E091 exclude published weights, models derived from them with no more than 2.5 × 10^25 additional operations or 25 percent (whichever is higher), and models trained with fewer operations than needed to match the most advanced published model.
Inspect sources and checks: US BIS AI diffusion rule, Q02 Threshold