Free reference · v1.0 · verified 2026-09-07

AI Governance Glossary & Regulation Cheat Sheet

The vocabulary and the dates, in one page. Every regulatory claim carries its instrument and date, and everything here was checked against a primary source or two independent credible ones.

Everyone - incl. individuals and freelancersTeam & up - if you answer for how others use AIOrganisation-wide - if you answer for AI across the business

No term matches that.

Who you are

Get this wrong and every decision downstream is wrong. Most organisations are deployers and wrongly assume the obligations are their vendor's.

ProviderEU AI ActTeam & up
Develops an AI system or general-purpose AI model, or has one developed, and places it on the market or into service under its own name or trademark. You can become a provider by substantially modifying a system, or by putting your own name on someone else's. This catches people out.
DeployerEU AI ActEveryone
Uses an AI system under its own authority, other than for personal non-professional activity. Most organisations are here. Deployer duties are lighter than provider duties but they are not nil - and several US state laws put real obligations here.
Autonomous Systems Officer (ASO)DIFCOrganisation-wide
A named, accountable individual required by DIFC Regulation 10 as a condition of commercially operating a high-risk autonomous or semi-autonomous system, with status and tasks substantially similar to a Data Protection Officer. Notable because neither the EU nor the US mandates a personal role of this kind for AI.

Risk and classification

Prohibited practiceEU AI ActEveryone
Uses banned outright. Applicable since 2 February 2025. A further prohibition covering AI-generated child sexual abuse material and non-consensual intimate imagery was added by the 2026 Omnibus and begins to apply on 2 December 2026.
High-risk AI systemEU AI ActTeam & up
Two routes: Annex III stand-alone systems (biometrics, critical infrastructure, education, employment, essential services, law enforcement, migration, justice), and Annex I systems embedded as safety components in already-regulated products. Obligations deferred by the Omnibus: Annex III to 2 December 2027, Annex I to 2 August 2028. Deferred is not cancelled - conformity assessment has long lead times.
Covered ADMTColoradoTeam & up
"Automated decision-making technology" that processes personal data and generates outputs - predictions, rankings, scores, classifications - used to materially influence a consequential decision. Colorado's replacement framework, effective 1 January 2027.
AEDTConnecticutTeam & up
Automated Employment-Related Decision Technology: processes personal data and produces a rank, score, classification or recommendation that is a substantial factor in an employment decision - hiring, firing, promotion, discipline, renewal, or selection for training.
Frontier developerCaliforniaOrganisation-wide
Under SB 53, a developer of frontier models; "large frontier developers" are those above $500M annual revenue, who carry the transparency and incident-reporting duties. Almost nobody reading this is one - but your vendors may be, and their disclosures are useful to you.

Documents you will be asked for

AI use-case inventoryEveryone
The list of every AI system in use, with owner, role, model/vendor, data touched, decision affected, oversight point and risk tier. Every framework assumes you have one. Almost nobody maintains one. It is the single highest-value artifact in this field.
AI impact assessmentTeam & up
Structured analysis of a system's effects on people and rights. ISO/IEC 42005 addresses this specifically. An assessment with no re-trigger condition is a historical document, not a control. Bind it to change management.
Technical documentationOrganisation-wide
The provider-side dossier describing design, data, performance and risk management, sufficient to demonstrate conformity.
Evidence packTeam & up
The contemporaneous record showing controls actually operated - logs, approvals, reviews, attestations. A control that leaves no trace is an intention. Design the evidence artifact together with the control, not afterwards.

Standards and conformity

The most expensive misconception in this field: that ISO/IEC 42001 certification demonstrates EU AI Act compliance.

The European Commission states that ISO/IEC 42001:2023 does not qualify for presumption of conformity under the AI Act, because its goals and definitions are not aligned with the quality management system the Act requires - which is why a new standard aimed at regulatory compliance is being developed.

Certification is still genuinely worth having. It builds the management system and satisfies buyers. It is simply not a legal shield.

Harmonised standardOrganisation-wide
A European standard cited in the Official Journal of the EU. Under Article 40, conformity with one creates a presumption of conformity with the corresponding AI Act requirements. This is the only route to presumption of conformity. The Commission requested harmonised standards across ten areas; first publications are expected from CEN/CENELEC during 2026, each then reviewed before OJEU citation.
ISO/IEC 42001Team & up
AI management system standard, published December 2023. Certifiable. Tells you how to run the system.
ISO/IEC 42005Organisation-wide
AI system impact assessment.
ISO/IEC 42006:2025Organisation-wide
Requirements for bodies auditing and certifying AI management systems - the accreditation layer that separates credible certification from paper mills. When reading any certificate, the scope statement matters more than the certificate. Most cover far less than the badge implies.
NIST AI Risk Management FrameworkTeam & up
Voluntary US framework organised around four functions: Govern, Map, Measure, Manage. The de facto global reference for how to assess risk.
NIST AI 600-1Organisation-wide
The Generative AI Profile, published July 2024, defining twelve GenAI-specific risk categories against the RMF core - including data provenance, confabulation, and dual-use capability.

Technical risk

Prompt injectionEveryone
Untrusted input causing a model to follow instructions it should not. Heads the OWASP GenAI LLM Top 10. A design constraint, not a bug awaiting a patch. Assume it succeeds and constrain the blast radius instead of trying to filter it.
Blast radiusTeam & up
What an AI system can actually reach and change - the union of its tool permissions. For agents this is the real control surface. The model isn't the risk; the tools you granted it are.
Goal hijacking · tool misuse · memory poisoningTeam & up
Agent-specific failure modes named in the OWASP Top 10 for Agentic Applications 2026 - a separate track from the LLM Top 10, not a replacement for it.
Confabulation (hallucination)Everyone
Fluent, confident, wrong output. Named in NIST AI 600-1. You cannot eliminate it. You can make it structurally unable to reach a consequential decision.
Shadow AIEveryone
Unsanctioned AI tools adopted by staff, typically faster than official rollouts. Discover it with amnesty, not audit. Punitive discovery drives usage underground and corrupts your inventory.

Key dates

Only dates confirmed against a primary source or two independent credible secondaries appear here. Red = already applies. Amber = within the next twelve months.

European Union

DateWhat applies
2 Feb 2025Prohibited practices, and Article 4 AI literacy - applies to providers and deployers regardless of size or sector
2 Aug 2025GPAI model obligations, governance, and the penalties regime (Ch. XII, Arts. 99–100)
27 Jul 2026Digital Omnibus on AI - Regulation (EU) 2026/1744 - enters into force; amended Article 4 applies
2 Aug 2026Article 50 transparency obligations - not deferred. Also Art. 101 GPAI-provider fines
2 Dec 2026Marking and detection of AI-generated content; new prohibition (CSAM / non-consensual intimate imagery) takes effect
2 Dec 2027Annex III stand-alone high-risk obligations (deferred from 2 Aug 2026)
2 Aug 2028Annex I embedded high-risk obligations (deferred from 2 Aug 2027)

United States

DateWhat applies
11 Dec 2025Executive Order 14365, Ensuring a National Policy Framework for Artificial Intelligence (FR 16 Dec 2025)
1 Jan 2026Texas TRAIGA (HB 149); California SB 53 - frontier AI frameworks, transparency reports, critical safety incident reporting to Cal OES within 15 days (24h if imminent risk)
9–10 Jan 2026DOJ AI Litigation Task Force established and operational
1 Jul 2026FTC proposed policy statement on suppression of accuracy in AI systems (comments closed 31 Jul 2026)
1 Oct 2026Connecticut SB 5 general effective date - WARN notices must disclose AI-related layoffs; use of AEDT is not a defense to a discrimination claim
1 Jan 2027Colorado SB 26-189 (covered ADMT) - replaced SB 24-205 before it took effect
1 Oct 2027Connecticut AEDT notice and pre-decision disclosure duties

Gulf

DateWhat applies
1 Sep 2023DIFC Regulation 10 enacted - personal data processed through autonomous and semi-autonomous systems
1 Jan 2026DIFC Regulation 10 moves to full enforcement; Autonomous Systems Officer required for high-risk systems
14 Jun 2026UAE Federal Authority for Artificial Intelligence and Data established, reporting to Cabinet, with authority to propose legislation

No GCC country has a horizontal AI statute. Saudi Arabia had no binding AI-specific statute as of 31 July 2026, and no announced legislative process. The binding layer across the Gulf is data protection law plus sectoral rules - but non-binding frameworks become commercially binding the moment they appear in a tender document.

Five things commonly stated that are wrong

Each of these is circulating right now, including in AI-generated summaries. Checking them is cheap; repeating them is expensive.

"The EU delayed the AI Act, so nothing applies until 2027."

The Omnibus deferred high-risk obligations. Article 50 transparency applied from 2 August 2026, Article 4 AI literacy since February 2025, prohibitions since February 2025, and penalties since August 2025.

Regulation (EU) 2026/1744; AI Act Arts. 99, 113

"California SB 53 requires a kill switch and annual third-party audits."

It does not. Those provisions come from SB 1047, which was vetoed. SB 53 requires a published frontier AI framework, transparency reports, critical safety incident reporting, and whistleblower protections.

SB 53 (Transparency in Frontier AI Act), signed 29 Sep 2025

"ISO 42001 certification means you're AI Act compliant."

The Commission states ISO/IEC 42001:2023 does not qualify for presumption of conformity. Only harmonised standards cited in the Official Journal do that.

European Commission, AI Act standardisation FAQ

"Colorado has the leading US state AI law."

Colorado repealed and replaced its AI Act (SB 24-205) with SB 26-189 before the original ever took effect, dropping the duty of care, mandatory risk programmes and impact assessments. The replacement takes effect 1 January 2027 and is under federal legal challenge.

SB 26-189, signed 14 May 2026

"Article 4 AI literacy was scrapped in the Omnibus."

It was softened, not removed and not deferred - from "ensure" a sufficient level of AI literacy to "take measures to support the development of" one. An obligation of means still requires demonstrable means, and it still binds every deployer. It was backed by penalties from August 2025; what the Omnibus changed is the enforcement, moving it to national supervision from 2 August 2026.

Regulation (EU) 2026/1744, Art. 1 point 5

Where to go next

Not sure where you stand? Take the AI Exposure Self-Check - ten to twelve questions depending on what you answer for, about four minutes, no email, nothing leaves your browser. It tells you which of the obligations above actually land on you.

This page is only useful if its dates are right. It is re-verified whenever the law moves - and the law moved three times in the nine months before this version. The governance page carries the same dates with the instrument beside each one, and a public correction on the rare occasions we get one wrong.

Found an error? Tell us by email and we will fix it and credit you. Being corrected in public is cheaper than being wrong in private.