What the law makes you constrain
About this documentUpdated ShowHide
Sean McDermott, Co-Founder and CEO, UnGovr
Written by Sean McDermott (with AI assistance) using the LexLint law library, which supplied every legal instrument, status and date on these pages.
Every law named here links to its summary page on lexlint.io, translated to English (if needed) and restructured to a standard format for human and code use. Every case links to the court's or the regulator's own record where one could be reached.
© 2026 UnGovr, publishing as LexLint. The text and the figures are licensed under Creative Commons Attribution-ShareAlike 4.0: share and adapt them, including commercially, with credit to LexLint (UnGovr) and under the same license. Please contact LexLint at hello@ungovr.org to discuss other terms. Logos and wordmarks belong to their owners.
Law library as of .
Legal information, not legal advice. This document describes the law as written and dated; it does not apply it to any system. The notice in the footer says what that means.
Eight places a duty can land on a running agent: the model itself, the instructions it runs on, the safeguards around it, the tools it can reach, the environment it runs in, what watches it, the people accountable, and what it is built from. Each with the law that already asks for something there, and how much of it the LexLint law library holds.
1What this document is
An AGENT is a program that picks its own next action, and the law does not reach it as one thing. It reaches the model that generates, the OPERATOR that runs it, the data it touches, the systems it calls and the people it acts on, each through a different statute with a different addressee. Introduction: The 6 parties in AI law sets out who those parties are. Global AI law: 8 common threads sets out the AI law itself, and the older privacy, security and scraping law that binds an AGENT without naming one.
This document cuts the same body of law along the line an engineer actually works in: the eight places a duty can land on a system while it is running. It is the before half of a pair. What the law makes you able to show once the system has run is What the law makes you able to show; when you have to report what happened, and to whom, is Incident reporting clocks; what it has cost the organizations that could not is Does legal action really happen?.
Each of the eight sections below states the layer in its own terms, names the law that already asks for something there, and then shows what the LexLint law library holds: how many requirement lines in force speak to that layer, from how many provisions and jurisdictions, with a sample you can read against the statutes themselves. It describes the law as written and dated. It does not say what any particular system should do, and a layer with a large count is not a layer you are behind on.
2How to read a layer
Reading the layers
- Ours first
- The eight layers are this section's own vocabulary, named for what an engineer constrains rather than for any framework's headings. The document reads the same with every alignment note removed.
- The instruments
- Every law named in the prose is linked to its page on lexlint.io, which carries the citation, the status, the source and the date it was read. The name and the date you see are read from the law library when this page is built, never typed here.
- The count
- A requirement line joins a layer when its own wording speaks to that layer. The test is words, not judgment, and it is applied to the line and not to the instrument around it, so a count is a floor rather than a ceiling: a duty stated in words the test does not carry is missed, and no duty is admitted because the statute it sits in happens to carry a class.
- In force
- The counted lines are the ones whose instrument binds on the law library's date in the byline. Law that is enacted and not yet in force is named in the prose as something coming, with its start date.
- One refusal
- A line about the safeguards a cross-border transfer needs is read as a boundary of the environment the data sits in, not as a guardrail around the model, so the word "safeguard" alone does not move a transfer rule into the third layer.
- A compact, not a law
- The note under each layer quotes SAFE v1.0, dated , the compact of the OSAA's SAFE working group, read from its text. It is not law. It is voluntary to join, and it states that “Participation in SAFE is a commitment to this compact. Open Secure AI Alliance membership alone does not impose it.” It is shown because a reader who has that framework in hand should be able to find the same ground here, not because the law follows it. It scores a layer in its own terms: “For every applicable control layer, state the expected outcome, assessment and supporting evidence, using pass, fail, insufficient evidence or not applicable. These labels describe a specific assessment but do not indicate general system safety or legal compliance.” Nothing on this page is scored that way.
- Not here
- Whether any of this reaches your system is the applicability question, answered per deployment from the roles the system holds and where its parties are. This document does not answer it, and neither does any table on it; the rules for deciding it are Where the parties are, and whose law that makes applicable.
3The eight layers
In order. Each section names the law first and shows what the law library holds second.
1 · The model itself
What the law asks of the model you ship.
Law binds conduct far more readily than it binds an artifact, so the duties that attach to the model itself are few, recent, and aimed at whoever makes it. In the European Union the general-purpose model duties of AI Act, Article 53 (obligations for providers of general-purpose AI models) bind now (since ): publish a training-content summary, adopt and follow a copyright policy that respects mining opt-outs, and keep technical documentation available for the AI Office. Where the Commission classifies a model as carrying systemic risk, AI Act, Article 55 (obligations for providers of general-purpose AI models with systemic risk) adds evaluation with documented adversarial testing, mitigation of those risks at Union level, cybersecurity for the model and its physical infrastructure, and reports of serious incidents to the AI Office without undue delay.
Three more duties on the model itself are enacted and coming, for a system the Act classifies as high-risk. AI Act, Article 9 (risk management system) (from , in 13 months; that date is for a system classified under Annex III, and a system classified under Annex I follows on the later date the instrument's page states) asks the provider for a risk-management system run as a continuous process across the system's lifecycle, with the system tested against pre-defined metrics before it is placed on the market. AI Act, Article 10 (data and data governance) asks that the training, validation and testing data sets be governed, examined for bias, and relevant and representative for the intended purpose. AI Act, Article 15 (accuracy, robustness and cybersecurity) asks for declared accuracy, resilience to error and to attempts to alter the system's use or performance, and cybersecurity measures against data poisoning, model poisoning, adversarial examples and model evasion. None of the three is counted in this layer's figures until it applies.
Two other jurisdictions set a duty at the model rather than at its use. AI Framework Act, Article 32 (safety-assurance duty for high-compute AI systems) turns on a compute threshold and asks for risk identified, assessed and mitigated across the life cycle, with the results submitted to the Ministry of Science and ICT. Transparency in Frontier Artificial Intelligence Act (SB 53) asks a large frontier developer to write, implement and publish a frontier framework.
Two things follow for anyone building on a model rather than training one. Nearly all of this binds the MAKER, and an OPERATOR inherits none of it directly; what it inherits is whatever the maker's documentation lets it show. And the law library is thin here, which is the finding rather than a gap in the reading: where a framework has a great deal to say about the model, binding law says little, and what it does say it says to somebody else.
99 requirement lines in force · 107 provisions · 51 jurisdictions · from: AI 135 · cybersecurity 29 · scraping 14 · privacy 5
The law behind the layer Show 7 of the 99 requirement lines in forceHide them
Drawn to span jurisdictions rather than to rank them. Each line is the law library's own statement of the requirement, with the instrument it comes from, the page that carries the citation and the source, and the date it was read.
Before placing the software on the market, undergo the conformity assessment for its class (Article 52): for class III, Annex IX, or Annex X coupled with Annex XI; for class IIb, Chapters I and III of Annex IX with an assessment of the technical documentation of at least one representative device per generic device group, or Annex X coupled with Annex XI; for class IIa, Chapters I and III of Annex IX with an assessment of the technical documentation of at least one representative device for each category of devices, or the technical documentation in Annexes II and III coupled with Section 10 or Section 18 of Annex XI.
 European Union
Medical Device Regulation, Annex VIII Rule 11 (classification of software) with Article 52 (conformity assessment by class), Regulation (EU) 2017/745, Annex VIII, Chapter III, Rule 11
Source
as of
|
If you are a large frontier developer (a frontier developer whose group had annual gross revenues over $500,000,000 in the prior calendar year), write, implement and publish on your website a frontier AI framework describing how you define and assess catastrophic-risk thresholds for your frontier models and apply mitigations, and review that framework at least once a year
 California
Transparency in Frontier Artificial Intelligence Act (SB 53), Cal. Bus. and Prof. Code Sections 22757.10 to 22757.16
Source
as of
|
To rely on the defense that you found the violation yourself, discover it through feedback from a developer, deployer or other person who believes a violation has occurred, through testing including adversarial or red-team testing, through following guidelines set by applicable state agencies, or through an internal review process while you substantially comply with the most recent version of the National Institute of Standards and Technology's Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile or another nationally or internationally recognized AI risk management framework (s. 552.105(e)(2)).
 Texas
TRAIGA (H.B. 149, 2025), Attorney General enforcement, civil investigative demand, cure period, penalties and defenses, Tex. Bus. & Com. Code §§ 552.101-552.106
Source
as of
|
Establish and carry out a plan to explain, so far as technically feasible, your AI's final output, the main criteria it used to reach that output, and an overview of the training data you used to develop or use it.
 South Korea
AI Framework Act, Article 34 (business-operator duties for high-impact AI), Act No. 20676, Article 34
Source
as of
|
Ensure, having regard to the state of the art, a level of security of network and information systems appropriate to the risk posed, and prevent and minimise the impact of incidents with a view to ensuring the continuity of the service (regulation 12(2)(a) and (b)).
 United Kingdom
Security Duties of Relevant Digital Service Providers, The Network and Information Systems Regulations 2018, SI 2018/506, reg. 12(1) and (2)
Source
as of
|
Get the rights holder's authorisation before reproducing or reusing copyrighted material, including a copyrighted database or compilation, to train an AI model; the fair-dealing exceptions do not name text-and-data-mining or AI training as a covered purpose.
 India
Copyright Act, No Text-and-Data-Mining Exception, Database Compilations, and Technological Protection Measures, Copyright Act, 1957 (No. 14 of 1957), ss. 2(o), 51, 52, 63, 65A
Source
as of
|
For a high-risk system, complete a conformity assessment before putting it into use through a conformity assessment organization where the system is on the list needing certification, or yourself with a technical dossier where it is not, and repeat it when a significant change affects the earlier result.
 Vietnam
Decree implementing the AI Law, risk classification, notification, conformity assessment and risk management, Decree No. 142/2026/ND-CP, arts. 6, 9, 11-15
Source
as of
|
Model. Model recognition of uncertainty, scope boundaries and stop conditions.
SAFE v1.0 Review Framework, read .
2 · The instructions it runs on
The authority a run is given, and its purpose.
Two bodies of law decide what an AGENT may be told to do. The first is the authority to process at all: General Data Protection Regulation (GDPR), Comprehensive Regime requires a lawful basis established and documented before any personal data of a person in the Union is processed, and the same requirement, in its own words, is the opening line of most of the privacy statutes in the law library. The second is access law, which asks whether the system was allowed on the machine it reached: Computer Fraud and Abuse Act (unauthorized access and the gates-based authorization test) does not reach a public, unauthenticated page, and does reach the same crawl once an operator has revoked access individually or a technical block has been circumvented.
A third sits between them and matters to anything that reads the web to learn: DSM Directive, Article 4 (text-and-data-mining exception and rights reservation) asks a miner to check for machine-readable rights reservations before mining content accessible from the Union, to skip or license what has been reserved, and to mine only content it accessed lawfully to begin with.
The practical point is that a prompt is not an authority. A system prompt telling an AGENT that it may read a site does not make the reading lawful, and "use the data we already hold" is not a lawful basis. What the law calls instructions is narrower and more formal than what an engineer calls them, and the OPERATOR owns the difference.
431 requirement lines in force · 371 provisions · 204 jurisdictions · from: scraping 243 · privacy 199 · AI 20 · cybersecurity 4
The law behind the layer Show 7 of the 431 requirement lines in forceHide them
Drawn to span jurisdictions rather than to rank them. Each line is the law library's own statement of the requirement, with the instrument it comes from, the page that carries the citation and the source, and the date it was read.
|
Establish and document a lawful basis under Article 6 before processing any personal data of a person in the EU.
European Union
General Data Protection Regulation (GDPR), Comprehensive Regime, Regulation (EU) 2016/679
Source
as of
|
|
Do not produce, distribute, or make available the digital replica of a deceased personality's voice or likeness in an expressive audiovisual work or sound recording without prior consent from the person or persons who hold the right of consent
California
Deceased Personality Digital Replica Liability Law (AB 1836), Cal. Civ. Code Section 3344.1 (AB 1836, Ch. 258, Stats. 2024)
Source
as of
|
Collect, use and share student personally identifiable information only for the purposes the contract authorizes or with the consent of the student or the student's parent, and obtain that consent before a use materially inconsistent with your privacy policy or the contract. A student can consent only if at least eighteen years old or legally emancipated (section 22-16-109(1) and (4)).
 Colorado
Student Data Transparency and Security Act: school service contract providers, C.R.S. sections 22-16-101 to 22-16-112 (HB 16-1423, 2016 Colo. Sess. Laws ch. 355)
Source
as of
|
|
If a school district requests deletion of a student's covered information under its control, delete it not later than the 60th day after the request, or as the contract or terms of service specify, unless the student or parent consents to your maintaining it (Tex. Educ. Code sec. 32.156).
Texas
Texas student information: duties of operators of websites and applications used for a school purpose, Tex. Educ. Code secs. 32.151 to 32.157
Source
as of
|
|
An app processing Korean personal data must be able to answer to the PIPC for its lawful basis and safeguards, and an individual harmed by a security failure may bring a private civil claim for statutory damages, or damages up to five times the actual loss, without needing to prove the controller's negligence.
South Korea
Personal Information Protection Act, enforcement and private civil remedy, Act No. 10465 (as amended by Act No. 19234, 2023), Arts. 39, 39-2, 51, 64-2
Source
as of
|
|
Establish and document a lawful basis under UK GDPR Article 6 before processing any personal data of a person in the United Kingdom, including the new closed-list recognised legitimate interests basis where it applies.
United Kingdom
UK GDPR and Data Protection Act 2018, as Amended by the Data (Use and Access) Act 2025, Data Protection Act 2018 (c. 12); UK GDPR, as amended by the Data (Use and Access) Act 2025, c. 18
Source
as of
|
Establish a lawful basis under article 7 before processing personal data, including data the person has made public.
 Brazil
Lei Geral de Proteção de Dados Pessoais (LGPD), Lei nº 13.709, de 14 de agosto de 2018 (LGPD), arts. 1º-10, 15-16, 23-32, 37-41, 46-47, 49-51 (general regime, principles, lawful basis, public-sector processing, agents and governance)
Source
as of
|
Instructions. Authorization and environmental assumptions were explicit and correct.
SAFE v1.0 Review Framework, read .
3 · The safeguards around it
What has to sit between the model and the act.
The law names outcomes, and only rarely the mechanism that secures them. AI Act, Article 50 (transparency obligations for AI systems and synthetic content) (since ) requires that a person interacting with an AI system be told so, and that synthetic content be marked in a machine-readable way; AI Framework Act, Article 31 (transparency obligations for AI outputs) asks the same of AI outputs in Korea. NIS2 Directive, Cybersecurity Risk-Management Measures asks for measures appropriate to the risk covering incident handling, continuity, supply chain and the security of development, with basic cyber hygiene, training, and multi-factor or continuous authentication where appropriate.
That shape is worth taking seriously when a control is being designed. A classifier, an approval step or an action limit is not compliant because it exists, and no statute in the law library asks for one by name. Each is evidence that an outcome the law does require has been secured, which is why this layer cannot be turned into a checklist, by us or by anyone else.
487 requirement lines in force · 392 provisions · 174 jurisdictions · from: privacy 278 · cybersecurity 201 · AI 52 · scraping 5
The law behind the layer Show 7 of the 487 requirement lines in forceHide them
Drawn to span jurisdictions rather than to rank them. Each line is the law library's own statement of the requirement, with the instrument it comes from, the page that carries the citation and the source, and the date it was read.
|
If you are a Commission-designated very large online platform or very large online search engine, commission an independent audit, at your own expense and at least once a year, of your compliance with Chapter III's due diligence obligations, including the systemic-risk mitigation measures for generated and manipulated content, and with any codes of conduct or crisis protocols you have joined.
European Union
Digital Services Act, Article 37 (independent audit of very large online platforms and search engines), Regulation (EU) 2022/2065, Article 37, supplemented by Commission Delegated Regulation (EU) 2024/436
Source
as of
|
|
If you seek to do business with the State of California, be ready to attest to and explain your policies and safeguards to protect public safety while preventing the misuse of your technology, where state contracting incorporates the certifications that Paragraph 1 asks for (Paragraph 1).
California
California Executive Order N-5-26 (vendor certifications, supply chain review and contractor responsibility in state AI procurement), Cal. Exec. Order N-5-26 ()
Source
as of
|
| Process personal data with reasonable and appropriate administrative, technical, organizational and physical safeguards (Rule 6.09). |
|
Respond to a civil investigative demand from the Attorney General, which may ask for a high-level description of your system's purpose, intended use, deployment context and benefits, the type of data used to program or train it, the categories of input data, its outputs, performance metrics, known limitations, and post-deployment monitoring and user safeguards (s. 552.103).
Texas
TRAIGA (H.B. 149, 2025), Attorney General enforcement, civil investigative demand, cure period, penalties and defenses, Tex. Bus. & Com. Code §§ 552.101-552.106
Source
as of
|
|
Build a risk-management system that monitors and responds to AI-related safety accidents involving your system.
South Korea
AI Framework Act, Article 32 (safety-assurance duty for high-compute AI systems), Act No. 20676, Article 32
Source
as of
|
|
Where you take a significant decision about a person, one with a legal effect or a similarly significant effect, based solely on automated processing of their personal data, have safeguards in place that give the person information about the decision, let them make representations, let them obtain human intervention from you and let them contest the decision.
United Kingdom
Automated Decision-Making: Significant Decisions Based Solely on Automated Processing (UK GDPR Articles 22A to 22D), Data (Use and Access) Act 2025, c. 18, s. 80, substituting Articles 22A to 22D into the UK GDPR
Source
as of
|
|
Do not build behavioral profiles of child and adolescent users from their personal data, including data obtained in age verification, to direct commercial advertising.
Brazil
ECA Digital (Lei nº 15.211/2025), children's personal data and profiling, Lei nº 15.211, de 17 de setembro de 2025 (Estatuto Digital da Criança e do Adolescente), arts. 7º, 16, 22 and 26
Source
as of
|
Safeguards. Classifiers, policies, approvals, action limits and trust-boundary checks were operating as intended, including for the human languages, scripts and usage conditions involved.
SAFE v1.0 Review Framework, read .
4 · The tools it can reach
Limits on what it may reach, send, spend or publish.
Every tool an AGENT holds reaches a different body of law. Credentials and access: NIS2 Directive, Cybersecurity Risk-Management Measures names access control and multi-factor authentication among its minimum measures. Getting past a barrier: Computer Fraud and Abuse Act (unauthorized access and the gates-based authorization test) is the United States example, and the law library's scraping topic is largely this layer, one national computer-misuse statute at a time. Anything the AGENT sends reaches communications and marketing law; anything it spends reaches payments law.
This is the layer where an agentic system differs most from the software the statutes were written for. A tool list is a list of legal exposures, and the exposure attaches at the moment the tool is available rather than at the moment it is used: a capability the AGENT could reach without a second authorization is one the OPERATOR has already granted. A subagent is one more way to reach a tool. No instrument in the law library names subagents; the limits at this layer bind the system as a whole, so what a subagent can reach counts against them.
647 requirement lines in force · 536 provisions · 217 jurisdictions · from: cybersecurity 315 · scraping 177 · privacy 160 · AI 14
The law behind the layer Show 7 of the 647 requirement lines in forceHide them
Drawn to span jurisdictions rather than to rank them. Each line is the law library's own statement of the requirement, with the instrument it comes from, the page that carries the citation and the source, and the date it was read.
|
If you are a provider of a high-risk AI system, you may exceptionally process special categories of personal data to the extent strictly necessary to detect and correct bias under Article 10(2), points (f) and (g), provided all six conditions of Article 4a(1) are met: the bias cannot be effectively addressed with other data, including synthetic or anonymised data; the data is subject to technical re-use limits and state-of-the-art security, including pseudonymisation; the data is secured with strict access controls and documented access; the data is not transmitted, transferred, or otherwise accessed by other parties; the data is deleted once the bias is corrected or its own retention period ends, whichever comes first; and your records of processing activities under the GDPR, the EU institutions data protection regulation, or the law enforcement directive state why the processing was strictly necessary and why other data could not achieve the same result.
European Union
AI Act, Article 4a (processing of special categories of personal data for bias detection and correction), Regulation (EU) 2024/1689, Article 4a
Source
as of
|
|
Do not require an individual to use their social security number to access an internet website unless a password, unique personal identification number or other authentication device is also required (section 1798.85(a)(4)).
California
Confidentiality of Social Security Numbers, Cal. Civ. Code sections 1798.85 to 1798.89
Source
as of
|
| Do not require an individual to transmit a social security number over the internet unless the connection is secure or the number is encrypted, and do not require an individual to use one to access an internet website unless a password, unique personal identification number or other authentication device is also required (section 6-1-715(1)(c) and (d)). |
|
Unless you are a government, do not require an individual to transmit the individual's social security number over the Internet unless the connection is secure or the number is encrypted, and do not require it for access to a website unless a password, personal identification number or other authentication device is also required (Tex. Bus. & Com. Code sec. 501.001(a)(3), (4)).
Texas
Texas limits on public display, Internet transmission and mailing of Social Security numbers, Tex. Bus. & Com. Code sec. 501.001
Source
as of
|
|
Where unauthorized access to a personal-information system or to a handler's device suggests a breach but the affected data subjects are hard to identify, or where some personal information is found traded unlawfully so that other data subjects' data may also have been breached, notify every data subject who may be affected within 72 hours of learning of it, stating that a further notice follows if the breach is confirmed (Article 34(2); Enforcement Decree Articles 39-2 and 39-3).
South Korea
Personal Information Protection Act, breach notification duties, Act No. 10465 (as amended by Act No. 19234, 2023), Art. 34; Enforcement Decree Arts. 39-40
Source
as of
|
|
It reaches you if you are a firm carrying on retail market business, meaning the regulated activities and ancillary activities, payment services and issuing of electronic money of a firm in a distribution chain that involves a retail customer, or if you communicate or approve a financial promotion addressed to or likely to be received by a retail customer, to the extent you are responsible for determining or materially influencing retail customer outcomes: act to deliver good outcomes for retail customers (Principle 12, PRIN 2.1.1R, and PRIN 3.2.6R to 3.2.7R).
United Kingdom
The Consumer Duty: Principle 12 and the cross-cutting rules and outcome rules for retail customers, FCA Handbook, PRIN 2A and Principle 12
Source
as of
|
|
Adopt technical and administrative security measures suited to protect personal data against unauthorized access and accidental or unlawful destruction, loss, alteration, communication, or improper processing, from the design of the product or service through its execution.
Brazil
Lei Geral de Proteção de Dados Pessoais (LGPD), Lei nº 13.709, de 14 de agosto de 2018 (LGPD), arts. 1º-10, 15-16, 23-32, 37-41, 46-47, 49-51 (general regime, principles, lawful basis, public-sector processing, agents and governance)
Source
as of
|
Tools. Credentials, permissions, spending, publishing and execution were appropriately constrained.
SAFE v1.0 Review Framework, read .
5 · The environment it runs in
Where the data may sit, and what it may touch.
Two questions live here: what the system may touch, and where the data may sit. The first is ordinary security law, stated as measures appropriate to the risk. The second is the transfer and localization rules, which are the sharpest constraint on an agent's architecture that most teams meet: GDPR Chapter V, Cross-Border Transfer Restrictions conditions any movement of personal data out of the Union, and the rest of the law library holds the same question answered differently in over a hundred jurisdictions.
Localization goes further than conditioning a transfer. CERT-In Cyber Security Directions, Incident Reporting, Logging and Time Synchronisation requires the logs an organization is told to keep to be maintained within Indian jurisdiction, which is a duty about where a record lives rather than about what it says. An AGENT that moves work to whichever region has capacity is making a legal choice in a scheduler.
386 requirement lines in force · 354 provisions · 180 jurisdictions · from: privacy 230 · cybersecurity 176 · AI 6 · scraping 6
The law behind the layer Show 7 of the 386 requirement lines in forceHide them
Drawn to span jurisdictions rather than to rank them. Each line is the law library's own statement of the requirement, with the instrument it comes from, the page that carries the citation and the source, and the date it was read.
|
If you are a provider of a high-risk AI system, you may exceptionally process special categories of personal data to the extent strictly necessary to detect and correct bias under Article 10(2), points (f) and (g), provided all six conditions of Article 4a(1) are met: the bias cannot be effectively addressed with other data, including synthetic or anonymised data; the data is subject to technical re-use limits and state-of-the-art security, including pseudonymisation; the data is secured with strict access controls and documented access; the data is not transmitted, transferred, or otherwise accessed by other parties; the data is deleted once the bias is corrected or its own retention period ends, whichever comes first; and your records of processing activities under the GDPR, the EU institutions data protection regulation, or the law enforcement directive state why the processing was strictly necessary and why other data could not achieve the same result.
European Union
AI Act, Article 4a (processing of special categories of personal data for bias detection and correction), Regulation (EU) 2024/1689, Article 4a
Source
as of
|
|
Do not require an individual to transmit their social security number over the internet unless the connection is secure or the number is encrypted (section 1798.85(a)(3)).
California
Confidentiality of Social Security Numbers, Cal. Civ. Code sections 1798.85 to 1798.89
Source
as of
|
| Do not require an individual to transmit a social security number over the internet unless the connection is secure or the number is encrypted, and do not require an individual to use one to access an internet website unless a password, unique personal identification number or other authentication device is also required (section 6-1-715(1)(c) and (d)). |
|
Unless you are a government, do not require an individual to transmit the individual's social security number over the Internet unless the connection is secure or the number is encrypted, and do not require it for access to a website unless a password, personal identification number or other authentication device is also required (Tex. Bus. & Com. Code sec. 501.001(a)(3), (4)).
Texas
Texas limits on public display, Internet transmission and mailing of Social Security numbers, Tex. Bus. & Com. Code sec. 501.001
Source
as of
|
|
You may process pseudonymized information of people in Korea without their consent for statistics, scientific research, archiving in the public interest and similar purposes, and when you give it to a third party for those purposes you must not include any information that could be used to identify a specific individual (Article 28-2).
South Korea
Personal Information Protection Act, pseudonymized information, Act No. 10465 (as amended by Act No. 21445, 2026), Arts. 28-2 to 28-5 and 28-7; Arts. 64-2 and 71 (penalties)
Source
as of
|
|
Do not do anything that circumvents an effective technological measure, such as encryption, scrambling or another access control or protection process, or a copy control mechanism, applied to a copyright work other than a computer program, if you know or have reasonable grounds to know that you are circumventing it.
United Kingdom
Copyright, Designs and Patents Act 1988, circumvention of technological measures, Copyright, Designs and Patents Act 1988, c. 48, ss. 296ZA to 296ZF
Source
as of
|
|
Transfer personal data outside Brazil only to a country or organization with an adequate level of protection, or under contractual clauses, corporate rules, or another article 33 safeguard.
Brazil
LGPD, international transfer of data, Lei nº 13.709, de 2018 (LGPD), arts. 33-36 (international transfer of data)
Source
as of
|
Environment. Network paths, isolation, targets and data boundaries were independently verified.
SAFE v1.0 Review Framework, read .
6 · What watches it
The duty to notice, and to be able to stop it.
The duty to notice is mostly a shadow of the duty to report. Very few instruments in the law library tell an OPERATOR to watch a system in a particular way; a great many tell it to report inside a fixed number of hours from the moment it became aware, which is only possible if something was watching. NIS2 Directive, Reporting Obligations and Cyber Resilience Act, Manufacturer Reporting Obligations are the clearest European examples, the second with a 24-hour early warning, a 72-hour notification and a final report within fourteen days of a corrective measure becoming available.
Because the duty is stated as a deadline, the clocks are where it can be read properly, and they have a document of their own: Incident reporting clocks discusses them, and the clocks page draws every reporting deadline in the law library on one time axis, with the sentence each was read from.
398 requirement lines in force · 336 provisions · 160 jurisdictions · from: cybersecurity 255 · privacy 102 · AI 91 · scraping 6
The law behind the layer Show 7 of the 398 requirement lines in forceHide them
Drawn to span jurisdictions rather than to rank them. Each line is the law library's own statement of the requirement, with the instrument it comes from, the page that carries the citation and the source, and the date it was read.
|
Have effective processes to identify, manage, monitor and report the risks you are or might be exposed to, including ESG risks in the short, medium and long term (Article 74(1), point (b)).
European Union
Capital Requirements Directive, Robust Governance Arrangements of Institutions (Article 74(1)), Directive 2013/36/EU, Article 74(1)
Source
as of
|
|
Do not allow an AI system to interact with a client in any form of therapeutic communication unless the regulated professional, the AI system and the client interact synchronously in real time, and do not let it generate therapeutic recommendations or treatment plans without the professional's review and approval
Colorado
HB 26-1195 (2026), Psychotherapy Artificial Intelligence Restrictions, Colo. HB 26-1195, 2026 Colo. Sess. Laws ch. 358
Source
as of
|
|
Respond to a civil investigative demand from the Attorney General, which may ask for a high-level description of your system's purpose, intended use, deployment context and benefits, the type of data used to program or train it, the categories of input data, its outputs, performance metrics, known limitations, and post-deployment monitoring and user safeguards (s. 552.103).
Texas
TRAIGA (H.B. 149, 2025), Attorney General enforcement, civil investigative demand, cure period, penalties and defenses, Tex. Bus. & Com. Code §§ 552.101-552.106
Source
as of
|
|
Build a risk-management system that monitors and responds to AI-related safety accidents involving your system.
South Korea
AI Framework Act, Article 32 (safety-assurance duty for high-compute AI systems), Act No. 20676, Article 32
Source
as of
|
|
If you identify, through complaints, your internal monitoring or any other source, that retail customers have suffered foreseeable harm as a result of your acts or omissions, act in good faith and take appropriate action to rectify the situation, including providing redress where appropriate (PRIN 2A.2.5R).
United Kingdom
The Consumer Duty: Principle 12 and the cross-cutting rules and outcome rules for retail customers, FCA Handbook, PRIN 2A and Principle 12
Source
as of
|
|
When you process the data of children and adolescents, above all for purposes not strictly necessary to run the product or service, map the risks, make efforts to mitigate them and prepare a data-protection impact, monitoring and evaluation report to share when the authority requests it.
Brazil
ECA Digital (Lei nº 15.211/2025), children's personal data and profiling, Lei nº 15.211, de 17 de setembro de 2025 (Estatuto Digital da Criança e do Adolescente), arts. 7º, 16, 22 and 26
Source
as of
|
|
Make your security-monitoring systems collect logs from deployed systems and network activity, correlate them through SIEM tools, keep SIEM dashboards under continuous monitoring and generate alerts for abnormal activity (paragraph 215).
India
RBI Commercial Banks Cybersecurity Directions, Cybersecurity Policy, Testing and Security Operations Centre, Reserve Bank of India (Commercial Banks - Cybersecurity, Technology: Risk, Resilience and Assurance Framework) Directions, 2026, paras. 15, 149 to 153 and 212 to 216
Source
as of
|
Monitoring. Operators detected and interrupted unexpected behavior in real time.
SAFE v1.0 Review Framework, read .
7 · The people accountable
The named person a duty lands on.
The law asks for a person, by name or by role, more often than it asks for a mechanism. General Data Protection Regulation (GDPR), Comprehensive Regime requires controller and processor responsibilities allocated and documented in a written agreement, and a data protection officer where core activities involve large-scale monitoring or large-scale special-category processing. NIS2 Directive, Cybersecurity Risk-Management Measures puts the management body itself in the loop: it approves the risk-management measures, oversees their implementation and completes training, and an entity outside the Union offering a covered service inside it designates a representative in a member state.
Human oversight arrives through privacy law rather than through AI law across most of the law library, as the right not to be subject to a solely automated decision and the review that follows from it. AI Act, Article 14 (human oversight) (from , in 13 months; that date is for a system classified as high-risk under Annex III) will change that in the Union: the provider must design the system so the people assigned to oversee it can understand its capacities and limits, interpret its output, decide not to use it, override or reverse it, and stop it, and the deployer must assign that oversight to people with the competence, training and authority to exercise it. For an AGENT that acts between a decision and a person, the question is not whether a human is available somewhere but whether the person affected can reach one.
380 requirement lines in force · 371 provisions · 167 jurisdictions · from: privacy 198 · cybersecurity 158 · AI 79 · scraping 10
The law behind the layer Show 7 of the 380 requirement lines in forceHide them
Drawn to span jurisdictions rather than to rank them. Each line is the law library's own statement of the requirement, with the instrument it comes from, the page that carries the citation and the source, and the date it was read.
|
It reaches you if you are an investment firm that engages in algorithmic trading, meaning trading in financial instruments where a computer algorithm automatically determines individual parameters of orders such as whether to initiate the order, the timing, price or quantity of the order or how to manage the order after its submission, with limited or no human intervention; a system used only to route orders to one or more trading venues, to process orders involving no determination of any trading parameters, or to confirm orders or process executed transactions after the trade is outside that meaning (Directive 2014/65/EU, Article 4(1), point (39)). Have in place effective systems and risk controls suitable to the business you operate to ensure that your trading systems are resilient and have sufficient capacity, are subject to appropriate trading thresholds and limits, and prevent the sending of erroneous orders or the systems otherwise functioning in a way that may create or contribute to a disorderly market (Article 17(1)).
European Union
MiFID II Algorithmic Trading Controls: Resilient Systems, Testing, Kill Functionality, Records and Notification (Article 17 and RTS 6), Directive 2014/65/EU, Article 17, supplemented by Commission Delegated Regulation (EU) 2017/589
Source
as of
|
|
Let a consumer opt out of your use of ADMT to make a significant decision about them, unless you offer an appeal to a human reviewer with authority to overturn the decision or another exception listed in the regulation applies
California
CCPA Automated Decisionmaking Technology Regulations, Cal. Code Regs. tit. 11, Sections 7200 to 7222
Source
as of
|
| Honor a consumer's request to opt out of profiling in furtherance of decisions that produce legal or similarly significant effects where the profiling rests on solely automated processing or human reviewed automated processing (Rule 9.04). |
|
Designate one or more employees to maintain the program, identify and assess reasonably foreseeable internal and external risks to records containing personal data, train employees and contractors, require employee compliance, and provide discipline for violations.
Texas
Texas data broker information security program, Tex. Bus. & Com. Code secs. 510.007, 510.009
Source
as of
|
|
An app that makes a decision about a Korean data subject using a fully automated system, including one applying artificial intelligence, that significantly affects the subject's rights or obligations, must let the subject object to or request an explanation of the decision, must not apply the decision over an objection absent a compelling reason or must reprocess it with human involvement, and must disclose the criteria and procedures it uses for automated decisions.
South Korea
Personal Information Protection Act, rights regarding automated decisions, Act No. 10465 (as amended by Act No. 19234, 2023), Art. 37-2; Enforcement Decree of the Personal Information Protection Act (Presidential Decree No. 34309, ), Arts. 44-2 to 44-4
Source
as of
|
|
Where you take a significant decision about a person, one with a legal effect or a similarly significant effect, based solely on automated processing of their personal data, have safeguards in place that give the person information about the decision, let them make representations, let them obtain human intervention from you and let them contest the decision.
United Kingdom
Automated Decision-Making: Significant Decisions Based Solely on Automated Processing (UK GDPR Articles 22A to 22D), Data (Use and Access) Act 2025, c. 18, s. 80, substituting Articles 22A to 22D into the UK GDPR
Source
as of
|
|
Comply with an ANPD warning's corrective-measures deadline before facing escalation to a fine, publicity of the infraction, blocking or deletion of the data involved, or suspension of the processing activity.
Brazil
LGPD, civil liability, administrative sanctions and the ANPD, Lei nº 13.709, de 2018 (LGPD), arts. 42-45, 52-54, 55-A a 58-B (civil liability, administrative sanctions and the ANPD)
Source
as of
|
Human operations. Responsibilities, escalation paths and stop procedures were clear, and people had an opportunity to understand, contest or recover from the outcome.
SAFE v1.0 Review Framework, read .
8 · What it is built from
Duties that follow the parts and the parties behind them.
What an AGENT is built from carries duties of its own. Processor arrangements are contractual by statute under General Data Protection Regulation (GDPR), Comprehensive Regime. Supply-chain security is one of the minimum measures of NIS2 Directive, Cybersecurity Risk-Management Measures. Cyber Resilience Act, Essential Requirements and Manufacturer Obligations (from , in 13 months) puts essential requirements on a product with digital elements and on the manufacturer behind it, which reaches software an agent is assembled from rather than the agent's own behavior. And training data is a supply question before it is a model question, which is where DSM Directive, Article 4 (text-and-data-mining exception and rights reservation) sits.
What an open-source project owes the people who run it takes the other end of this layer: what a project that agents are built from owes the people who run it, and what it can usefully ship to reduce their exposure rather than its own.
1,178 requirement lines in force · 934 provisions · 236 jurisdictions · from: cybersecurity 573 · privacy 503 · AI 199 · scraping 56
The law behind the layer Show 7 of the 1,178 requirement lines in forceHide them
Drawn to span jurisdictions rather than to rank them. Each line is the law library's own statement of the requirement, with the instrument it comes from, the page that carries the citation and the source, and the date it was read.
|
If you provide direct electronic access to a trading venue, have in place effective systems and controls that ensure a proper assessment and review of the suitability of clients using the service, that those clients are prevented from exceeding appropriate pre-set trading and credit thresholds, that their trading is properly monitored, and that appropriate risk controls prevent trading that may create risks to you or contribute to a disorderly market; direct electronic access without such controls is prohibited (Directive 2014/65/EU, Article 17(5)). Apply the controls of Articles 13, 15 and 17 and the real-time monitoring of Article 16 of Delegated Regulation (EU) 2017/589 to the order flow of each of your direct electronic access clients, conduct a due diligence assessment of prospective clients, and review your due diligence assessment processes and reassess your clients' systems and controls annually (Articles 20, 22 and 23).
European Union
MiFID II Algorithmic Trading Controls: Resilient Systems, Testing, Kill Functionality, Records and Notification (Article 17 and RTS 6), Directive 2014/65/EU, Article 17, supplemented by Commission Delegated Regulation (EU) 2017/589
Source
as of
|
|
Make available, at no cost to the user, a disclosure verification tool that lets a user assess whether image, video, or audio content was created or altered, other than by a minor modification, by your generative AI system and that outputs any system provenance data detected in the content
California
California AI Transparency Act (SB 942, as amended by AB 853, SB 1000 and AB 2713), Cal. Bus. and Prof. Code Sections 22757 to 22757.6
Source
as of
|
|
Make every component of your governance and risk management framework available to the Division on request, as of December 1 each year if you are a life insurer and as of July 1 each year if you are a private passenger automobile or health benefit plan insurer.
Colorado
Amended Regulation 10-1-1 (2025), Governance and Risk Management Framework for Insurers' Use of External Consumer Data and Information Sources, Algorithms, and Predictive Models, 3 CCR 702-10, Regulation 10-1-1
Source
as of
|
|
To rely on the defense that you found the violation yourself, discover it through feedback from a developer, deployer or other person who believes a violation has occurred, through testing including adversarial or red-team testing, through following guidelines set by applicable state agencies, or through an internal review process while you substantially comply with the most recent version of the National Institute of Standards and Technology's Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile or another nationally or internationally recognized AI risk management framework (s. 552.105(e)(2)).
Texas
TRAIGA (H.B. 149, 2025), Attorney General enforcement, civil investigative demand, cure period, penalties and defenses, Tex. Bus. & Com. Code §§ 552.101-552.106
Source
as of
|
|
It reaches you if you operate as a firm with respect to consumer credit lending, other than under an agreement in which you take an article in pawn on the terms described in CONC 5.2A.2R(1), a current account agreement that may allow the account-holder to overdraw without a pre-arranged overdraft or to exceed a pre-arranged overdraft limit where doing so would be a regulated credit agreement (overrunning), a non-commercial agreement, or a small borrower-lender-supplier agreement which is a restricted-use credit agreement: undertake a reasonable assessment of the creditworthiness of a customer before entering into a regulated credit agreement, significantly increasing the amount of credit provided under one, or significantly increasing a credit limit for running-account credit under one, and do not take that step unless you can demonstrate that you have done so in accordance with CONC 5.2A and had proper regard to the outcome in respect of affordability risk (CONC 5.2A.1R, 5.2A.2R, 5.2A.4R, 5.2A.5R).
United Kingdom
FCA Handbook, CONC 5.2A, assessing the creditworthiness of a customer before consumer credit is granted or significantly increased, FCA Handbook, CONC 5.2A
Source
as of
|
|
Transfer personal data outside Brazil only to a country or organization with an adequate level of protection, or under contractual clauses, corporate rules, or another article 33 safeguard.
Brazil
LGPD, international transfer of data, Lei nº 13.709, de 2018 (LGPD), arts. 33-36 (international transfer of data)
Source
as of
|
|
Embed synthetically generated information with permanent metadata or another technical provenance mechanism, including a unique identifier, to the extent technically feasible, and do not enable removal, suppression or modification of that label or metadata.
India
Synthetically Generated Information Labelling Duty for Intermediaries, Information Technology (Intermediary Guidelines and Digital Media Ethics Code) Rules, 2021, rule 3(3), as inserted by the Information Technology (Intermediary Guidelines and Digital Media Ethics Code) Amendment Rules, 2026 (G.S.R. 120(E), dated )
Source
as of
|
Supply chain. A partner or third-party dependency (cloud, evaluation, data or tooling provider) upheld assumed controls.
SAFE v1.0 Review Framework, read .
Some layers in the alliance's list are not the subject of any section above. They are quoted here, in its words, rather than filed under a section that is not theirs, and the law library is not read against them.
Serving path. The invoked model, runtime, gateway, cache, proxy and wrapper matched configurations and produced expected results.
Authorization and commitment. The consequential action matches valid authorization at commitment, including delegation, expiry and revocation.
Workflow composition. The combined sequence, shared state and interactions did not cross a trust boundary and did not exceed the originating scope.
SAFE v1.0 Review Framework, read .
4Where the law is quiet, and where we are
The distribution across the eight layers is itself a finding, and it runs the opposite way to most governance frameworks. The layer with the most binding law behind it is what it is built from (1,178 requirement lines in force across 236 jurisdictions). The layer with the least is the model itself (99). A framework has most to say about the model and the safeguards; statutes have most to say about the parties, the boundaries and the parts, because those are the things a legislature can name without describing an architecture.
The law library holds the EU AI Act as separate instruments per article, and the high-risk duties that sit in this document's model and people layers, AI Act, Article 9 (risk management system), AI Act, Article 10 (data and data governance), AI Act, Article 14 (human oversight) and AI Act, Article 15 (accuracy, robustness and cybersecurity), are enacted and not yet in force (from , in 13 months; that date is for a system classified as high-risk under Annex III, and a system classified under Annex I follows on the later date each instrument's page states, both written into the Regulation by the Digital Omnibus on AI). They are named in the layers above as something coming, and the counts above do not include them, because a counted line is one whose instrument binds on the law library's date in the byline. No figure on this page is adjusted for law that is enacted and waiting.
5What this document does not claim
It does not say that any layer is covered, or that a control satisfying one is compliant. The counts describe how much of the law library speaks to a layer, not how much of a layer is regulated: a single sentence in one statute can bind more of your system than forty lines elsewhere. The samples are samples, drawn to span jurisdictions rather than to rank them.
It also does not adopt anyone's framework. The alliance's review criteria are quoted because a reader holding them should be able to find the same ground here, with the law underneath it. They are the compact as it stood on the date they were read, the alliance may revise them, and the law they sit beside is dated on every line.