The fast answer: The EU AI Act sorts AI systems by risk. It bans a short list of practices outright, puts the heaviest duties on high-risk systems, and asks most chatbots and content generators to be open about being AI. The United States has no single federal AI law. Instead, builders face a growing set of state statutes, city rules and agency enforcement, while Washington pushes for one national standard that would override some of them.

If you build or buy AI, both systems reach you at once. A product sold in Paris and Phoenix has to satisfy a binding EU regulation with fixed application dates and a patchwork of US laws that changed several times in 2026 alone. This guide walks through what each side actually requires, with links to the official text, and closes with what a builder can do this quarter regardless of which rules end up applying.

The EU AI Act in one paragraph

The EU AI Act is Regulation (EU) 2024/1689, published in the Official Journal in July 2024. It is a regulation, not a directive, so it applies directly in every member state without national transposition. It regulates AI by use, not technology: the same model can be minimal risk as a writing aid and high risk when it screens job applicants. In 2026 the EU amended the Act through the Digital Omnibus on AI, which moved the high-risk deadlines back and added new bans. The European Commission's AI Act Service Desk publishes the consolidated text as at 27 July 2026, with every amended and new provision marked, and that is the version this guide quotes.

The four risk tiers of the EU AI Act

The Act never says "four tiers," but its structure produces them, and your tier decides almost everything else.

1. Prohibited practices (Article 5)

Article 5 lists practices that cannot be placed on the EU market at all. In plain terms, the consolidated text bans AI that:

• uses subliminal, manipulative or deceptive techniques to distort a person's decisions in a way that causes, or is likely to cause, significant harm;
• exploits vulnerabilities linked to age, disability or a person's social or economic situation, with the same harm threshold;
• generates or manipulates realistic intimate images, video or audio of an identifiable person without their explicit consent (added by the 2026 amendments);
• generates or manipulates child sexual abuse material (also added in 2026);
• scores people on social behaviour or personal traits where that leads to unjustified or detrimental treatment;
• predicts a person's risk of committing a crime based solely on profiling or personality traits;
• builds facial recognition databases by untargeted scraping of images from the internet or CCTV;
• infers emotions in workplaces and schools, except for medical or safety reasons;
• categorises people by biometrics to deduce sensitive traits such as race, political opinions or sexual orientation;
• uses real-time remote biometric identification in public spaces for law enforcement, outside narrow exceptions.

The original prohibitions have applied since 2 February 2025. The two new bans on non-consensual intimate content and child sexual abuse material apply from 2 December 2026, according to the amended Article 113.

For consumer products, the second bullet matters most. A companion app or AI toy that exploits a child's age to push harmful behaviour is not a high-risk system with paperwork to complete. It is a banned practice.

2. High-risk systems (Article 6, Annex I and Annex III)

High-risk AI comes in two kinds. The first is AI that is a safety component of, or is itself, a product already covered by EU product safety law listed in Annex I, such as toys, machinery and medical devices. The second is AI used in the eight areas listed in Annex III:

1. Biometrics, including remote identification and emotion recognition
2. Critical infrastructure, such as road traffic and water, gas, heating and electricity supply
3. Education and vocational training, including admissions and exam proctoring
4. Employment and workers' management, including CV screening, promotion and dismissal decisions
5. Access to essential private and public services, including credit scoring, public benefits, life and health insurance pricing, and emergency call triage
6. Law enforcement
7. Migration, asylum and border control
8. Administration of justice and democratic processes, including tools meant to influence elections

Providers of high-risk systems carry the bulk of the Act's obligations, covered in detail below.

3. Transparency obligations (Article 50)

Most customer-facing AI is not high risk, but much of it still has duties under Article 50. Systems meant to interact with people must be designed so that people are told they are "interacting with an AI system," unless that is obvious. Providers of systems that generate synthetic audio, images, video or text must ensure outputs are "marked in a machine-readable format" and detectable as AI generated. Deployers of deepfakes must disclose that the content has been artificially generated or manipulated, and deployers of emotion recognition systems must tell the people exposed to them.

Article 50 applies from 2 August 2026. The Council of the EU's AI timeline records that the 2026 amendments cut the grace period for the AI-generated content marking duty from six months to three, giving a deadline of 2 December 2026.

4. Minimal risk

Everything else, from spam filters to recipe recommendations, carries no specific duties beyond the AI literacy measures in Article 4.

A note on general-purpose AI models

Large general-purpose models have their own chapter, which has applied since 2 August 2025. If you build on someone else's model, those duties fall mainly on the model provider, while your application's risk tier still depends on its use.

EU AI Act dates: what applies now and what moved

Deadlines are where most published summaries go stale. Here is the schedule from the consolidated Article 113, cross-checked against the Council's record of the 29 June 2026 final approval of the amending regulation:

• 2 February 2025: general provisions, AI literacy and the original prohibited practices.
• 2 August 2025: general-purpose AI model rules, governance, notified body provisions and most penalties.
• 2 August 2026: the Act's general date of application, including the Article 50 transparency duties.
• 2 December 2026: the two new prohibitions on non-consensual intimate content and child sexual abuse material, and the end of the grace period for machine-readable marking of AI-generated content.
• 2 December 2027: high-risk rules for stand-alone systems in Annex III areas. This was originally 2 August 2026.
• 2 August 2028: high-risk rules for AI embedded in Annex I products such as toys and medical devices.

The delay to high-risk obligations is real, but it is not a pause. Prohibitions, general-purpose AI rules and transparency duties are already live or arrive within weeks. And because Article 9 requires testing "prior to their being placed on the market," a high-risk product launching in early 2028 needs its testing evidence well before then.

What the EU AI Act requires of high-risk systems

Chapter III, Section 2 sets the core requirements for high-risk systems. In practice they fall into a handful of workstreams.

Risk management and testing (Article 9)

Providers must run a risk management system across the whole lifecycle. Article 9 says high-risk systems must be "tested for the purpose of identifying the most appropriate and targeted risk management measures," and that testing must be performed against "prior defined metrics and probabilistic thresholds." That last phrase matters. A one-off demo where the system behaved well is not evidence. Regulators expect a test plan written before testing, measurable pass criteria, and results that can be repeated.

Accuracy, robustness and cybersecurity (Article 15)

Article 15 requires high-risk systems to "achieve an appropriate level of accuracy, robustness, and cybersecurity" and to "perform consistently in those respects throughout their lifecycle." It names specific attacks the system must resist, including data poisoning, model poisoning, "adversarial examples or model evasion." For a language model product, that translates directly into jailbreak, prompt injection and adversarial testing.

Data, documentation, logging and human oversight

The same section adds data governance (Article 10), technical documentation (Article 11), event logging (Article 12), transparency to deployers (Article 13) and human oversight (Article 14).

Conformity assessment (Article 43)

Before a high-risk system goes on sale, it needs a conformity assessment. Under Article 43, most Annex III systems (points 2 to 8) follow internal control under Annex VI, which "does not provide for the involvement of a notified body." Biometric systems may need a notified body, particularly where harmonised standards have not been applied. AI inside Annex I products follows the existing sectoral procedure, such as the toy or medical device route. Any substantial modification triggers a new assessment.

Internal control does not mean light control. When no outside body signs off, the provider's own test evidence is what a market surveillance authority will examine, and because models get retrained, evaluation has to recur rather than happen once.

EU AI Act penalties

Article 99 sets the ceilings. Using a prohibited practice can bring fines of up to EUR 35,000,000 or 7% of total worldwide annual turnover, whichever is higher. Breaching most other operator obligations, including the high-risk requirements and the Article 50 transparency duties, can bring up to EUR 15,000,000 or 3%. Supplying incorrect, incomplete or misleading information to authorities can cost up to EUR 7,500,000 or 1%. National authorities set the actual amounts within those limits.

AI regulation in the United States

The US is moving in two directions at once: states keep passing AI laws while the federal government tries to replace many of them with one national framework.

The federal position

Congress has not passed a comprehensive AI statute. On 11 December 2025 the President signed the executive order Ensuring a National Policy Framework for Artificial Intelligence. According to the White House fact sheet, it directs the Attorney General to set up an AI Litigation Task Force to "challenge unconstitutional, preempted, or otherwise unlawful State AI laws," asks the Commerce Department to identify conflicting state laws, ties some federal grant funding to state policy, and calls for national legislation that "would preempt State AI laws that stifle innovation."

On 20 March 2026 the White House published its national AI legislative framework, arguing that "a patchwork of conflicting state laws would undermine American innovation." Its first objective is protecting children, including features on platforms likely to be used by minors that reduce "potential sexual exploitation of children or encouragement of self-harm." Those are recommendations to Congress, not law.

The FTC: existing law, applied to AI

The Federal Trade Commission does not need a new statute to act. When it launched Operation AI Comply in September 2024 with five enforcement actions, the agency said "there is no AI exemption from the laws on the books." In September 2025 it sent orders to seven companies offering consumer chatbots in an inquiry into AI companions, asking how they "measure, test, and monitor for negative impacts before and after deployment" on children and teens. And in July 2026 it sought comment on a proposed policy statement on AI accuracy, which treats distorting AI outputs against reasonable consumer expectations as a potential deceptive practice and argues that state laws requiring changes to outputs may be preempted.

State laws worth knowing

Each law below was checked against the legislature's, governor's or state agency's own page before inclusion. Several widely cited state laws were left out because their official pages could not be opened to confirm current status.

Colorado. Colorado's original AI Act, SB 24-205, was the first broad US law aimed at algorithmic discrimination. It never took effect in its original form. In May 2026 the state enacted SB 26-189, which, according to the Colorado Attorney General, "repeals and reenacts" those provisions as rules on automated decision-making technology (ADMT). The Colorado Division of Real Estate's summary says it was signed on 14 May 2026 and applies to consequential decisions made on or after 1 January 2027. Deployers must give consumers notice when ADMT is used in a consequential decision, and after an adverse outcome must explain, within 30 days, the decision and "the role that ADMT played." Violations count as a deceptive trade practice. The Attorney General's rulemaking is under way, with a public hearing scheduled for 26 October 2026.

Texas. The Texas Responsible Artificial Intelligence Governance Act (HB 149) took effect on 1 January 2026. It prohibits developing or deploying AI that "intentionally aims to incite or encourage a person to" harm themselves or others or commit a crime, AI built with intent to unlawfully discriminate, and AI built to produce child sexual abuse material or certain deepfakes. Government agencies must tell consumers they are dealing with AI, and health care providers must disclose AI use to patients. The Attorney General has exclusive enforcement power, with a 60-day cure period and civil penalties that reach $200,000 per uncurable violation. Two details stand out for builders. The law lets a defendant avoid liability where it discovered a violation through testing, including adversarial or red-team testing, and it gives weight to substantial compliance with the NIST AI Risk Management Framework, including its Generative Artificial Intelligence Profile, or another recognised risk framework.

California. California has passed more AI laws than any other state, and its 2026 session added several that matter directly to this guide:

• SB 53, the Transparency in Frontier Artificial Intelligence Act, signed on 29 September 2025, requires large frontier developers to "publicly publish a framework on its website," lets companies and the public report critical safety incidents to the Office of Emergency Services, and protects whistleblowers.
• SB 243, signed on 13 October 2025, requires companion chatbot platforms to set protocols for suicidal ideation and self-harm, disclose that interactions are artificially generated, and give minors break reminders.
• SB 813 and AB 1405, signed on 9 September 2026, create "a first-in-the-nation framework for independent verification organizations" for AI systems and "a state registry for AI auditors" with standards for "independence, transparency, and integrity." In the words of AB 1405's author, "We cannot expect industry to simply grade its own homework."
• SB 1119, signed on 10 September 2026, adds companion chatbot protections for children, including crisis protocols and parental controls. The Governor's office describes it as "the first in the country to require companies to conduct independent child safety audits and annual risk assessments."

Effective dates for the 2026 California bills vary, and the governor's announcements do not list them, so check the chaptered text before planning around a date.

New York City. Local Law 144 has been enforced since 5 July 2023. Employers using automated employment decision tools must have a bias audit conducted within the past year by an independent auditor, publish a summary of the results, and notify candidates. The city's guidance says an auditor employed by the employer or the tool's vendor is not independent.

Will federal preemption wipe out state laws?

Not on its own. An executive order cannot repeal a state statute. Preemption either needs an act of Congress or a court ruling that a specific state law conflicts with federal law. Until one of those happens, state laws on the books remain enforceable by state attorneys general. Many of the duties those laws impose, such as testing for harm to children, disclosing AI use and documenting how a system was evaluated, also appear in the federal framework's own priorities and in the EU Act. Building to that common core is the safest bet whichever way preemption goes.

Where the EU and US approaches agree

The two systems differ in structure but agree on more substance than headlines suggest. Both expect testing before launch: the EU through Article 9, Texas by rewarding red-team testing, the FTC by asking chatbot companies how they test for harm. Both want people told when they are talking to AI. Both single out children. And both increasingly lean on independent evidence, from New York City's independent bias auditors to California's new registry of AI auditors. Texas even names the NIST AI Risk Management Framework, which NIST describes as "intended for voluntary use," as a basis for a liability defence.

What AI builders can do now

These steps hold their value whichever way the rules move.

1. Classify every AI feature you ship

Write down, for each feature, its intended purpose, who uses it and where. Check it against Article 5, Annex I and Annex III, and against the state laws for every US market you sell into.

2. Map your system to a framework

Pick a recognised framework, such as the EU AI Act's own requirements, the NIST AI RMF or ISO/IEC 42001, and map your controls to it. The map shows where your evidence is thin and gives auditors and buyers a structure they already know. Our guide to the NIST AI RMF and ISO/IEC 42001 explains what each framework asks of you and where product testing fits.

3. Get independent test evidence

Almost every law above asks, in one form or another, "how do you know it is safe?" The strongest answer is evidence from someone with no stake in the result, produced against a methodology defined before the test. That is the reason RateAI built the A.E.G.I.S.™ rating system. It scores a customer-facing AI product across seven pillars: Hallucination Integrity, Guardrail Resistance, Data Handling and Privacy, Ethical Alignment, Autonomy and Control, Adversarial Robustness, and Explainability and Transparency. Several of those map directly onto legal language. Guardrail Resistance and Adversarial Robustness test the jailbreaks, prompt injection and adversarial inputs that Article 15 and Texas's red-team provision describe. Ethical Alignment tests behaviour with vulnerable users and sensitive topics, which is where Article 5's ban on exploiting age-related vulnerabilities and California's companion chatbot laws focus.

The results become an A.E.G.I.S.™ rating on a scale from A-5 Guardian down to A-1 Unrestricted. Scoring parameters are frozen per scoring epoch so evaluations stay comparable over time, and every prompt case, validator vote and certification is chained by SHA-256 hash in a write-once audit trail.

To be precise about what this does and does not do: RateAI's Pro and Enterprise tiers include compliance mapping to the EU AI Act, NIST AI RMF and ISO 42001, and A.E.G.I.S.™ certification artifacts are designed as defensible evidence in regulatory reviews. Certification supports compliance work. It does not make a product compliant with any law, and formal regulatory determinations remain with the relevant authorities. You can compare what each tier includes on the RateAI pricing page.

4. Watch the dates that are closest

For EU-facing consumer products, the near-term dates are 2 December 2026 for the new prohibitions and AI content marking, then 2 December 2027 for Annex III high-risk systems. In the US, Colorado's ADMT law applies from 1 January 2027, and California's 2026 bills will phase in on their own schedules. For background on how third-party certification works in general, see our complete guide to AI safety certification.

Frequently asked questions

Does the EU AI Act apply to US companies?

Yes, if they place AI systems on the EU market or the output of their AI is used in the EU. Where a company is headquartered does not decide whether the Act applies.

When do the EU AI Act's high-risk rules apply?

After the 2026 amendments, high-risk rules apply from 2 December 2027 for stand-alone systems in Annex III areas and from 2 August 2028 for AI embedded in products covered by Annex I, such as toys and medical devices. Prohibitions and general-purpose AI rules already apply.

Is there a federal AI law in the United States?

No single comprehensive federal AI statute exists. Federal agencies such as the FTC apply existing law to AI, a December 2025 executive order seeks to challenge some state AI laws, and the White House has asked Congress for a national framework.

What are the maximum fines under the EU AI Act?

Up to EUR 35,000,000 or 7% of worldwide annual turnover for prohibited practices, whichever is higher, and up to EUR 15,000,000 or 3% for most other breaches. Supplying misleading information to authorities can bring up to EUR 7,500,000 or 1%.

Does RateAI certification make my AI product legally compliant?

No. RateAI provides compliance mapping to frameworks such as the EU AI Act, NIST AI RMF and ISO 42001, and its A.E.G.I.S.™ certification artifacts are designed as defensible evidence, but formal regulatory determinations remain with the relevant authorities.

Do AI chatbots have to tell users they are AI?

In many cases, yes. EU AI Act Article 50 requires it unless it is obvious, California SB 243 requires companion chatbot platforms to disclose that interactions are artificially generated, and Texas requires government agencies and health care providers to disclose AI use.

Get your system ready for the rules that apply to it

If you are working out where your AI product sits under the EU AI Act or US state law, and what evidence you will need to show for it, talk to the RateAI team about mapping your system to the frameworks that matter and testing it under A.E.G.I.S.™.