How to use this page
Checked: November 7, 2026. AI rules are moving faster than almost anything else on this site: dates in this page have changed during 2026, and some will change again. Treat every date as "as of" the line above, and check the source before you rely on it. Every row links to the lesson that explains it, with its sources.
Not legal advice. This page helps you read the rules, take part in a review and ask better questions. Anyone with real obligations needs a lawyer who practises in them.
Three habits that make everything here easier:
- Classify by use, not technology. The same model can be minimal risk in one product and high-risk in another. What the output is used for, and whether it decides about people, sets almost every obligation.
- Ask for evidence, not answers. A vendor's "yes", a reviewer's approval and a passing test all need something behind them: a document, a planted error caught, a test that could have failed.
- Date everything. A compliance note without a date is a guess. Write down when you checked, and what.
Key dates
What applies when, oldest first. Bold rows are still to come.
| Date | What | Where | Lesson |
|---|---|---|---|
| Jan 26, 2023 | NIST AI RMF 1.0 released (voluntary) | US | 2 |
| Jul 5, 2023 | NYC Local Law 144: yearly bias audits for automated hiring tools | New York City | 5 |
| Dec 2023 | ISO/IEC 42001 published: a certifiable AI management system | International | 3 |
| Jul 26, 2024 | NIST AI 600-1, the Generative AI Profile | US | 2 |
| Aug 1, 2024 | EU AI Act in force (applies in stages) | EU | 4 |
| Jan 23, 2025 | EO 14179: revoked the 2023 AI order | US federal | 5 |
| Feb 2, 2025 | EU AI Act: prohibited practices (Article 5) and AI literacy (Article 4) | EU | 4 |
| Apr 3, 2025 | OMB M-25-21 (using AI) and M-25-22 (buying AI) replace the 2024 memos | US federal | 5 |
| Jul 23, 2025 | America's AI Action Plan, with EOs 14318, 14319, 14320 | US federal | 5 |
| Aug 2, 2025 | EU AI Act: general-purpose AI model rules; penalties | EU | 4 |
| Oct 1, 2025 | M-25-22 applies to new AI contracts | US federal | 5 |
| Dec 11, 2025 | EO 14365 (challenging state AI laws) and OMB M-26-04 (buying language models) | US federal | 5 |
| Jan 1, 2026 | California SB 53, Texas TRAIGA, Illinois HB 3773 take effect | States | 5 |
| Jul 27, 2026 | EU Digital Omnibus (Regulation (EU) 2026/1744) in force: dates below moved | EU | 4 |
| Aug 2, 2026 | EU AI Act Article 50 transparency, for new systems | EU | 4 |
| Aug 2026 | OWASP Top 10 for LLM Applications, 2026 edition | International | 7 |
| Sep 29, 2026 | EO 14434: agencies say "super intelligence" for AI in their own documents | US federal | 1 |
| Dec 2, 2026 | EU: new bans (nudifiers, child sexual abuse material); content marking for systems already on the market | EU | 4 |
| Jan 1, 2027 | Colorado SB 26-189, New York RAISE Act, California's automated decision-making rules | States | 5, 7 |
| Aug 2, 2027 | EU member states' AI regulatory sandboxes | EU | 4 |
| Dec 2, 2027 | EU high-risk duties for Annex III uses (hiring, credit, education, benefits...) | EU | 4 |
| Dec 11, 2027 | OMB M-26-04 expires, unless renewed | US federal | 5 |
| Aug 2, 2028 | EU high-risk duties for AI in Annex I regulated products | EU | 4 |
Still in motion
What to check before relying on anything above:
| What | Status as of Nov 7, 2026 | Lesson |
|---|---|---|
| NIST AI RMF | Being revised under the AI Action Plan; check which version you cite | 2 |
| M-25-21's high-impact deadline | The memo says 365 days (April 2026); some trackers say September 22, 2026. No official extension found | 5 |
| Federal preemption of state AI laws | EO 14365 lawsuits and funding conditions under way; no court has struck down a state AI law on those grounds; no federal statute | 5 |
| FTC policy statement on AI accuracy | Proposed in July 2026; not final | 5 |
| EU harmonised standards | Being written (CEN and CENELEC, JTC 21); once published, following them shows high-risk compliance | 4 |
| Commission guidelines | Still arriving, on definitions, prohibited practices and high-risk classification | 4 |
Which framework asks what
| Framework | Binding? | Who | What it asks for | Lesson |
|---|---|---|---|---|
| NIST AI RMF 1.0 | Voluntary | Anyone | Govern, Map, Measure, Manage: 19 categories, 72 subcategories | 2 |
| ISO/IEC 42001 | Voluntary; certifiable | Organizations | An AI management system: clauses 4 to 10, 38 Annex A controls, a Statement of Applicability | 3 |
| EU AI Act | Law | Providers and deployers, wherever based, if the output is used in the EU | Duties by risk tier: banned, high-risk, transparency, minimal; plus general-purpose model rules | 4 |
| OMB M-25-21, M-25-22, M-26-04 | Binding on US agencies | Agencies, and their vendors through contracts | Inventories, AI officers, seven practices for high-impact AI, contract terms | 5 |
| State AI laws | Law | By where you operate | Decision laws (notice, explanation, review); frontier model laws (safety frameworks, incident reports) | 5 |
| OWASP LLM Top 10 (2026) | Voluntary | Builders of LLM apps | Security risks: prompt injection, sensitive information disclosure, excessive agency first | 7 |
Which rules touch my system?
Five questions, in order. Each "yes" adds to the list.
| Ask | If yes | Lesson |
|---|---|---|
| Does it decide, or help decide, about people (jobs, credit, housing, education, benefits)? | EU high-risk (Annex III); Colorado, Illinois, NYC, California ADMT; civil rights and credit law | 4, 5 |
| Is its output used in the EU? | The EU AI Act can apply, wherever you're based | 4 |
| Does it talk to people, or make content? | EU Article 50: say it's AI, mark generated content | 4 |
| Do you sell it to a US federal agency? | M-25-22 contract terms; M-26-04 if it's a language model | 5 |
| Did you build it on someone else's model? | You're probably its provider, with the provider's duties | 10 |
All "no"? It's probably minimal risk: no new legal duties, but the FTC Act still covers what you claim about it, and an inventory entry still belongs to it.
The engineer's sequence
The order of work, and what you end up with at each step.
| Step | Do | You end up with | Lesson |
|---|---|---|---|
| 1. Inventory | Scan code, check what's running, ask about bought and planned AI | A list of systems with owners, uses, data and review dates | 6 |
| 2. Classify | Use the five questions above | A tier and a rules list per system | 4, 5 |
| 3. Data | Allowlist fields, redact, fence user text, check outputs | A prompt guard and a log of hashes and counts | 7 |
| 4. Test | Noise floor, fairness pairs, red team with a canary, release gates | Tests that can fail, in CI | 8 |
| 5. Oversight | Design the review; plant known errors; watch time and volume | Numbers that show review is real | 9 |
| 6. Vendors | Must-haves, evidence, the right questions for the kind of vendor | A verdict and a list of contract clauses | 10 |
| 7. Policy | Tools, data, banned and reviewed uses, a default; test it with requests | One page people can follow | 11 |
Red flags
Phrases that sound like controls and often aren't.
| When you hear | Ask | Lesson |
|---|---|---|
| "It's not high-risk: a human makes the final decision" | Does it profile people? Then the exemption doesn't apply | 4 |
| "The executive order cancels the state law" | Has a court, Congress or the state changed it? An order alone can't | 5 |
| "We inventoried the repos" | And bought software, switched-on features, and plans? | 6 |
| "The filter blocks injections" | What's in the model's context? Anything there can come out | 7 |
| "We tested it for bias" | How many runs, against what noise? Could the test have failed? | 8 |
| "The model refused" | Is the secret in the refusal anyway? | 8 |
| "Every decision is reviewed by a person" | How many planted errors do reviewers catch, in how many seconds? | 9 |
| "They're ISO 42001 certified, so we're covered" | What does the certificate's scope cover? What are our own duties? | 10 |
| "The vendor scored 100%" | Does choosing them make us the provider? | 10 |
| "We banned it" | Where did the use go instead? | 11 |
Numbers worth knowing
| Number | What | Lesson |
|---|---|---|
| 4 and 72 | NIST AI RMF functions, and subcategories | 2 |
| 12 | Risks in NIST's Generative AI Profile | 2 |
| 38 | ISO/IEC 42001 Annex A controls | 3 |
| 8 | Annex III high-risk areas in the EU AI Act | 4 |
| €35M or 7% | Top EU AI Act fine, for prohibited practices (€15M or 3% for most other duties) | 4 |
| 6 months | Minimum log retention for EU high-risk deployers | 4 |
| 7 | Minimum practices for high-impact AI in M-25-21 | 5 |
| 0.8 | The four-fifths rule: an impact ratio below it suggests adverse impact | 8 |
| 72 hours | A common incident-notification term in AI contracts | 10 |