selekoa

A documentary history of machine intelligence.

Found and cross-checked by machines · How this works →

Comparison · 15 questions · 9 instruments

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.

The same kind of questions put to developers’ own frameworks: Compare frameworks →

Download every cell as CSV

Question Q05Risk evaluation

Must the covered party evaluate or test its models for risks?

Criterion: A duty to assess, evaluate or test the model for risks, before or after it is made available. Partly if the duty is to assess something other than the model (for example, a service or a content category) or is stated only as an aim.

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:

  1. An order to an agency to impose a duty counts as partly.
  2. Protection only by reference to another law counts as partly.
  3. “Without undue delay” counts as a stated deadline.
  4. 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

  1. EU AI Act

    EU AI ActBinding regulation

    Stated

    • “perform model evaluation in accordance with standardised protocols and tools reflecting the state of the art, including conducting and documenting adversarial testing”

      Article 55(1), point (a)

    Note: Applies to providers of general-purpose AI models with systemic risk. The provision does not tie the evaluation to a point before or after the model is placed on the market.

    Inspect sources and checks

  2. EU Code of Practice, Safety and Security

    EU Code of Practice, Safety and SecurityCompliance route, not binding

    Stated

    • “Signatories will conduct at least state-of-the-art model evaluations”

      Measure 3.2, first paragraph (PDF p. 12)

    • “at least before placing the model on the market”

      Measure 1.2, last paragraph (PDF p. 8)

    Note: A full systemic risk assessment and mitigation process (identification, analysis including model evaluations, acceptance determination, mitigation) is carried out at least before placing on the market and when the Measure 7.6 conditions are met; lighter-touch evaluations and post-market monitoring continue along the model lifecycle (Measures 1.2 and 3.5).

    Inspect sources and checks

United States: states

  1. California SB 53

    California SB 53Binding statute

    Stated, partly

    • Paraphrase: The framework must describe how the developer approaches defining and assessing thresholds used to identify and assess whether a frontier model has capabilities that could pose a catastrophic risk.

      SEC. 2, Bus. & Prof. Code § 22757.12(a)(2)

    • “A large frontier developer shall write, implement, comply with, and clearly and conspicuously publish on its internet website a frontier AI framework”

      SEC. 2, Bus. & Prof. Code § 22757.12(a), opening clause

    Note: The framework, which a large frontier developer must implement and comply with, describes how the developer approaches defining and assessing thresholds to identify and assess whether a frontier model has capabilities that could pose a catastrophic risk (§ 22757.12(a)(2)). The text sets no free-standing duty to test each model; § 22757.12(c)(2)(A)–(B) requires summaries of assessments of catastrophic risks conducted pursuant to the framework, and § 22757.12(d) summaries of any assessment of catastrophic risk from internal use. Search for the rest of the criterion: SEC. 2 in full (Bus. & Prof. Code §§ 22757.10–22757.16); SEC. 4 (Lab. Code §§ 1107–1107.2). Terms searched: “assess”, “evaluat”, “test”. Passages read and not counted: SEC. 2, Bus. & Prof. Code § 22757.12(c)(2)(A): Requires summaries of assessments in the transparency report; it presupposes assessments under the framework rather than imposing them. SEC. 2, Bus. & Prof. Code § 22757.12(d): Transmission of summaries of any internal-use assessment to the Office of Emergency Services; no duty to assess is stated.

    Inspect sources and checks

  2. New York RAISE Act (As amended, Chapter 96 of 2026)

    New York RAISE ActAs amended, Chapter 96 of 2026Binding statute; the amended duties take effect on 1 January 2027

    Stated, partly

    • “developer to identify and assess whether a frontier model has”

      Bill text, § 2 (GBL § 1421(1)(b) as re-added), p. 4, line 22

    • “A large frontier developer shall … write, implement, comply with, and clearly and conspicuously publish on … its internet website a frontier AI framework”

      Bill text, § 2 (GBL § 1421(1) as re-added), p. 4, lines 14–16

    Note: The framework, which a large frontier developer must implement and comply with, describes how it handles assessing whether a model has capabilities that could pose a catastrophic risk. No free-standing duty to test each model was found; § 1421(3)(b) requires summaries of assessments conducted under the framework and § 1422(2)(a) summaries of internal-use assessments. Search for the rest of the criterion: Bill text §§ 1–5, including GBL §§ 1420–1429 as re-added. Terms searched: “assess”, “evaluat”, “test”. Passages read and not counted: GBL § 1421(3)(b)(i)–(ii) as re-added, p. 5: Summaries of assessments in the transparency report; presupposes rather than imposes assessment. GBL § 1422(2)(a) as re-added, p. 5: Transmission of internal-use assessment summaries; no duty to assess is stated.

    Inspect sources and checks

  3. New York RAISE Act (As signed, Chapter 699 of 2025)

    New York RAISE ActAs signed, Chapter 699 of 2025Binding statute; the amended duties take effect on 1 January 2027

    Stated, partly

    • “Describe in detail the testing procedure to evaluate if the frontier model poses an unreasonable risk of critical harm”

      Proposed GBL § 1420(12)(c), p. 3, lines 31–32

    • “specific tests and test results used in any assessment of the frontier model required by this section or the developer's safety and security protocol”

      Proposed GBL § 1421(1)(d), p. 4, lines 28–30

    Note: The protocol the large developer must implement describes a testing procedure to evaluate whether the model poses an unreasonable risk of critical harm, and the developer must record and retain test information. The text sets no free-standing duty to test each model; § 1421(2) prohibits deployment where it would create an unreasonable risk of critical harm. Search for the rest of the criterion: Whole print S6953-B, pp. 1–5. Terms searched: “test”, “evaluat”, “assess”. Passages read and not counted: Proposed GBL § 1421(2), p. 4: A prohibition on deployment, not a duty to evaluate.

    Inspect sources and checks

  4. Illinois AI Safety Measures Act

    Illinois AI Safety Measures ActBinding statute

    Stated, partly

    • “identify and assess whether a frontier model has capabilities that could pose a catastrophic risk”

      Section 10(a)(2), PDF p. 5

    • “assessments of catastrophic risks from the frontier model conducted pursuant to the large frontier developer's frontier AI framework”

      Section 10(c)(2)(A), PDF p. 7

    Note: The framework, which a large frontier developer must implement and comply with, describes how the developer approaches defining and assessing thresholds to identify and assess whether a frontier model has capabilities that could pose a catastrophic risk (Section 10(a)(2)); the transparency report summarises assessments conducted under the framework and their results (Section 10(c)(2)(A)–(B)), and Section 10(e) requires summaries of any assessment of catastrophic risk from internal use to be sent to the Agency. The text sets no free-standing duty to assess or test each model, so the duty runs through the framework. The cross-family checker read the cell as stated, because the framework the developer must comply with covers assessing capabilities and reviewing assessments before deployment. Search for the rest of the criterion: Sections 5, 10, 15, 17, 18, 20 and 25 in full. Terms searched: “assess”, “evaluat”, “test”, “shall”. Passages read and not counted: Section 10(e), PDF pp. 10–11: Requires a summary of “any assessment” of internal-use risk to be sent to the Agency; it presupposes an assessment and does not itself require one. Section 10(a)(10), PDF p. 6: A matter the framework describes (assessing and managing internal-use risk), not a free-standing duty to assess the model. Maintainer decision 1a of 2026-11-05 (Illinois row): partly, as for California SB 53, whose duty is in identical words; the cross-family checker's reading, stated, is noted.

    Inspect sources and checks

China

These measures address providers of services, not developers as such.

  1. China, generative AI services (2023)

    China, generative AI services (2023)Binding measures, addressed to service providers

    Stated, partly

    • “提供具有舆论属性或者社会动员能力的生成式人工智能服务的，应当按照国家有关规定开展安全评估，并按照《互联网信息服务算法推荐管理规定》履行算法备案和变更、注销备案手续。”

      Working translation: Those providing generative AI services with public-opinion attributes or social mobilisation capacity shall carry out a security assessment in accordance with relevant state provisions, and complete algorithm filing, modification and cancellation procedures in accordance with the Provisions on the Administration of Algorithmic Recommendation for Internet Information Services.

      Article 17 (第十七条)

    Note: The security assessment (安全评估) is required only for services with public-opinion attributes or social mobilisation capacity, and is of the service, by reference to 'relevant state provisions' not set out in the Measures; no duty to evaluate or test the model for risks was found. Article 8 requires annotation quality assessment (set aside: data annotation, not risk). Search for the rest of the criterion: Articles 4, 7, 8, 14, 17 and 19. Terms searched: “评估”, “测试”, “检测”. Passages read and not counted: Article 8 (第八条): 数据标注质量评估 (annotation quality assessment) concerns data annotation, not risks.

    Inspect sources and checks

  2. China, AI content labelling (2025)

    China, AI content labelling (2025)Binding measures, addressed to service providers and others

    Not found in the recorded search

    Read
    Articles 1–14 in full.
    Terms searched
    “评估”, “测试|检测”, “风险”, “核验”.
    Passages read and set aside
    Article 6 (第六条), first paragraph, point (1) and (3): Dissemination services check file metadata for implicit labels (核验) and detect explicit labels or other traces: label verification on content, not a risk evaluation. Article 12 (第十二条): Refers to security assessment (安全评估) procedures under other rules, during which labelling materials are supplied; it does not itself impose an assessment.

    Note: No duty to evaluate or test models for risks, or to assess a service or content category for risks, was found.

    Open the search record

Revoked, or rescission announced

Kept as part of the record, each with its standing shown.

  1. US Executive Order 14110

    US Executive Order 14110Revoked by Executive Order 14148 (2025-01-20)

    Not found in the recorded search

    Read
    Whole cached text of the Federal Register printing (88 FR 75191–75226), with close reading of Secs. 3, 4.1, 4.2, 12 and 13 and the signature block.
    Terms searched
    “red-team”, “evaluat”, “test”, “results”, “shall require”.
    Passages read and set aside
    Sec. 4.2(a), opening words, 88 FR 75197: “the Secretary of Commerce shall require:”: a reporting duty on results of testing conducted, imposed through the Secretary; no command to conduct testing. Sec. 4.2(a)(i)(C), 88 FR 75197: “the results of any developed dual-use foundation model’s performance in relevant AI red-team testing”: a reporting duty on results of testing conducted, imposed through the Secretary; no command to conduct testing. Sec. 4.1(a)(ii), 88 FR 75196: Directs NIST guidelines to enable developers to conduct red-teaming tests; guidance, not a duty.

    Note: No duty on companies to evaluate or test their models was found. The Secretary of Commerce is directed to require reports that include the results of any red-team testing a company has conducted, on named risk categories including biological weapons (recorded at policy level only); a duty to report results of testing conducted is not a duty to test, and is set aside below.

    Open the search record

  2. US BIS AI diffusion rule

    US BIS AI diffusion ruleRescission announced on 2025-05-13; no formal rescission rule inspected

    Stated, partly

    • “Assess the capabilities, intentions, and opportunity of an adversary”

      Supplement No. 10 to part 748, para. 14(e)(i), 90 FR 4566 (PDF p. 23)

    • “Every interface by which model weights can be accessed, directly or indirectly, must be reviewed”

      Supplement No. 10 to part 748, para. 15(c)(ii), 90 FR 4566 (PDF p. 23)

    Note: Assessments required concern data-centre security threats and interfaces to model weights (to set and limit output rates), not the risks of the model itself. Search for the rest of the criterion: Federal Register PDF, 90 FR 4544–4584: DATES, SUPPLEMENTARY INFORMATION sections III and IV and Rulemaking Requirements; regulatory text §§ 740.27, 742.6(a)(13) and (b)(14), Supplement No. 10 to part 748 in full, ECCN 4E091 entry; signature and filing lines. Terms searched: “assess”, “evaluat”, “test”, “review”. Passages read and not counted: Supplement No. 10 to part 748, para. 11, 90 FR 4566: “ongoing monitoring, evaluation, and end user due diligence” of compliance with the VEU conditions; not of models.

    Inspect sources and checks