Governance and the EU AI Act: the topline

In 2025, a company shipped an AI tool that scored job applicants partly on their facial expressions during video interviews. Under the EU AI Act, that kind of emotion recognition in the workplace is banned. Not fined. Banned.
The EU AI Act is the world's first major law that regulates artificial intelligence directly. If you build AI tools, use them at work or buy them for your team, you need to know the topline. You do not need to read 100 pages of legal text. You need the shape of the law: what is forbidden, what is heavily controlled, and when the rules bite.
What the law actually does
The EU AI Act sorts AI systems into risk tiers. The higher the risk to people's rights and safety, the heavier the rules. Think of it as a traffic-light system with four levels.
- Unacceptable risk: banned outright.
- High risk: allowed, but tightly controlled (paperwork, testing, human oversight).
- Limited risk: allowed, with transparency duties (you must tell people).
- Minimal risk: basically free to use (spam filters, AI in video games).
One key point: the law applies based on where the AI is used, not where the company is based. A US startup whose chatbot serves EU customers is on the hook. The GDPRGDPREU regulation governing how organizations collect, store and use personal data, with fines tied to global revenue for breaches.View full definition → (Europe's privacy law) has the same long reachreachThe number of unique people exposed to your message in a given period. Unlike impressions, reach counts each person once, no matter how often they see it.View full definition →, and that reach made it a global standard.
Two definitions. A "provider" is whoever builds or sells the AI system. A "deployer" is whoever uses it in their organisation. Both have duties, and providers carry the heavier load.
Tier 1: What is prohibited
These uses have been off the table in the EU since February 2025. Concrete examples:
- Social scoring: ranking people's trustworthiness and then denying them services. (Picture a national "good behaviour" score.)
- Emotion recognition at work or in schools, meaning software that claims to detect from your face or voice whether you are angry, bored or lying.
- Untargeted face scraping to build facial recognition databases from internet images or CCTV.
- Manipulative AI that exploits people's vulnerabilities (age, disability) to distort their behaviour and cause harm.
- Predictive policing based purely on profiling a person to guess they will commit a crime.
- Biometric categorisation that infers sensitive traits like race, sexual orientation or political views.
If your product idea touches any of these, stop. No compliance checklist makes it legal in the EU.
Tier 2: What is high-risk
This is the tier most professionals will actually run into. High-risk does not mean "evil." It means the AI is used in a setting where a bad decision seriously affects someone's life. Examples:
- AI that screens CVs or ranks job candidates.
- Credit scoring and loan decisions.
- AI in medical devices and diagnostics.
- Exam scoring and admissions.
- Critical infrastructure such as power grids and water.
- Certain uses by law enforcement, border control and courts.
If you deploy a high-risk system, you take on real obligations. You must keep humans in the loop, log what the system does, use good-quality data, monitor for bias and be open with the people affected. Providers must register these systems in an EU database before selling them.
A real scenario
Say your HR team buys an AI résumé-screener. You are now a deployer of a high-risk system. Your duties include:
- Assigning a human to oversee its decisions (and actually review them, not rubber-stamp them).
- Keeping logs of its outputs.
- Telling candidates that AI is part of the process.
- Checking that the provider gave you proper instructions and a risk assessment.
"We just bought it off the shelf" is not a defence. Buying high-risk AI makes you responsible for how you use it.
Tier 3: Limited risk (the transparency tier)
For general use, everyday tools like ChatGPT, Claude and Gemini mostly sit here. The core rule is simple: people should know when they are dealing with AI.
- Chatbots must make clear that users are talking to a machine.
- AI-generated or heavily edited images, audio and video ("deepfakes") must be labelled as artificial.
- AI-generated content should carry a machine-readable mark where feasible.
So if you deploy a customer-service chatbot, a one-line "You're chatting with our AI assistant" disclosure goes a long way.
General-purpose AI: the models behind your tools
The big models behind ChatGPT, Claude and Gemini are called general-purpose AI (GPAI). They are flexible enough to do many tasks. The Act gives the companies that make them a separate set of duties, which have applied since August 2025.
GPAI providers must publish summaries of their training data, respect EU copyright rules and provide technical documentation. The most powerful models (those posing "systemic risk") face extra testing and security requirements. The EU published a voluntary General-Purpose AI Code of Practice to help providers comply, and most major model makers signed it.
For you as a user, most of this happens upstream. It is also why your AI vendor's documentation and "model cards" matter: they partly exist to satisfy these rules.
The timeline, and why it moved
The Act entered into force in August 2024, and the rules switch on in waves. The original plan:
- Feb 2025: prohibited uses banned, AI literacy duties begin.
- Aug 2025: rules for general-purpose AI models apply.
- Aug 2026: most high-risk and transparency rules apply.
- Aug 2027: high-risk rules for AI built into regulated products (medical devices, machinery) apply.
The first two dates happened as planned. The high-risk dates did not hold. In November 2025 the European Commission proposed a "Digital Omnibus" package, partly because the technical standards companies need for compliance were running late. It links the high-risk rules to those standards being ready, with backstop dates of December 2027 for uses like hiring, credit and education, and August 2028 for AI in regulated products. The package also proposed easing some duties, including moving the main AI literacy obligation from companies to governments.
Check the current legal text before you plan around any date. The Omnibus had to pass the Parliament and the Council, and some details changed along the way. The Commission's AI Act page tracks the status.
The practical message stays the same. The prohibitions and GPAI rules are live now. The high-risk duties are coming, just a bit later. A delay gives you time to prepare properly. It does not let you skip the work, and buyers and regulators already ask about it.
Penalties, briefly
Fines are serious. Banned uses can cost up to 35 million euros or 7% of global annual revenue, whichever is higher. Other breaches carry lower caps. The numbers are set so that compliance costs less than ignoring the law.
Knowledge check
1. What is the core organising principle the EU AI Act uses to decide how heavily an AI system is regulated?
2. A US-based startup with no EU offices offers a chatbot that serves customers in France and Germany. Why does the EU AI Act still apply to it?
3. In the vocabulary of the EU AI Act, what distinguishes a 'provider' from a 'deployer'?
4. Select ALL uses of AI that fall under the 'unacceptable risk' tier and are prohibited in the EU.
Select all the correct answers.
5. Select ALL statements that correctly describe the difference between the risk tiers of the EU AI Act.
Select all the correct answers.
A simple way to check where you stand
You do not need a lawyer for a first-pass triage. Ask three questions about any AI system you use or build:
- Is the use prohibited? (Social scoring, workplace emotion detection and so on.) If yes, stop.
- Is it high-risk? (Hiring, credit, health, education, law enforcement.) If yes, you have real obligations, even if the deadline has moved.
- Does it talk to people or generate content for them? If yes, you owe them transparency.
Here is that triage as a tiny script you could adapt for an internal checklist:
def classify_ai_use(use_case):
prohibited = {"social_scoring", "workplace_emotion_recognition",
"untargeted_face_scraping", "predictive_policing"}
high_risk = {"cv_screening", "credit_scoring", "medical_diagnosis",
"exam_scoring", "border_control"}
if use_case in prohibited:
return "PROHIBITED: do not deploy in the EU."
if use_case in high_risk:
return "HIGH-RISK: human oversight, logging, bias checks required."
return "LIMITED/MINIMAL: add a clear 'this is AI' disclosure."
print(classify_ai_use("cv_screening"))
# HIGH-RISK: human oversight, logging, bias checks required.This is a starting filter, not legal advice. It still forces the right first question every time.
What this means for non-EU readers
Even if you are in the US, UK or elsewhere, you have reasons to care:
- Reach: if any of your users are in the EU, the Act may apply to you.
- Template effect: like the GDPR, this law shapes rules elsewhere, and many companies adopt one global standard rather than juggle several.
- Good practice: human oversight, transparency and data qualitydata qualityThe degree to which data is fit for purpose: accurate, complete, consistent, timely, valid and unique. Poor quality data undermines analytics, reporting and AI.View full definition → are sound AI hygiene with or without a regulator.
Key Takeaways
- Learn the four tiers: prohibited, high-risk, limited (transparency) and minimal. Most daily AI use is "limited", which mostly means telling people it's AI.
- The bans have applied since February 2025. Social scoring, workplace and school emotion recognition, untargeted face scraping and profiling-based predictive policing are off-limits in the EU.
- The GPAI model rules have applied since August 2025. Your vendor's documentation exists partly because of them.
- The high-risk deadlines moved. The Digital Omnibus proposal pushed them toward December 2027 (hiring, credit, education) and August 2028 (regulated products). Check the final text, and treat hiring, lending, health and education AI as high-risk now anyway.
- Run the three-question triage (prohibited? high-risk? talking to people?) on every AI tool before you roll it out.
What to do, from this lesson
These actions are compiled in the role's Playbook.
- Never clone a real person's voice, face, or likeness without documented consent
- Run three-question triage—prohibited, high-risk, talking to people—before rolling out any tool
Related articles
Recent articles from the blog that build on this lesson.
- AIIs a no-training clause enough when shadow AI shows up in 43% of breaches?Most AI privacy programmes still focus on what employees paste into chatbots and whether vendors train on it. The 2026 evidence points to a bigger exposure: what your connected models can reach, and who else can make them reach it.
- AIAnthropic warns its own models might resist shutdown, and the IPO pitch is where it said soAnthropic's IPO documentation warns that its own Claude models could resist human attempts to shut them down and cause catastrophic harm. This week's developments show every major frontier lab shipping power faster than governance can respond, and the gap is no longer theoretical.
- AIAI systems causing real harm before oversight can catch upA hallucination in a military AI system nearly triggered a US attack on Chinese nuclear infrastructure. This week's developments, taken together, show a widening gap between what AI systems can do and what the humans overseeing them can actually catch.