# What is the EU's proposed AI Act?

Date: 2021-05-04 · Author: Robert Scales · https://guidetoartificialintelligence.com/en/learn/eus-proposed-ai-act/

**Answer:** The AI Act is a draft law the European Commission proposed on April 21, 2021, to set common rules for artificial intelligence across the European Union. It sorts AI uses by risk, banning a few, placing strict duties on high-risk systems, requiring transparency for some others, and leaving most tools alone.

Two weeks ago, on April 21, the European Commission published a proposal for a regulation it calls the Artificial Intelligence Act. The Commission describes it as the first-ever legal framework on AI. Instead of regulating a technology as a whole, it looks at what an AI system is used for and how much harm that use could cause. The same kind of software might be harmless in a video game and tightly controlled in a hiring tool.

It is important to be clear about one thing: this is a proposal, not a law. The European Parliament and the member states, through the Council, still have to debate it, amend it and agree on a final text. That process usually takes a long time, and the details could change a great deal before anything applies.

## The four risk levels

The legal text itself separates uses that create an unacceptable risk, a high risk, and a low or minimal risk, and then adds special transparency rules for certain systems. The Commission's announcement presents this as four levels, which is the easiest way to read it:

1. **Unacceptable risk: banned.** These are uses seen as a clear threat to people's safety or rights. The proposal lists AI that manipulates people through subliminal techniques, AI that exploits the vulnerabilities of groups such as children or people with disabilities in harmful ways, general-purpose "social scoring" by public authorities, and, with narrow exceptions, live facial recognition and other remote biometric identification in public spaces by police.
2. **High risk: allowed, with strict duties.** An annex to the proposal lists areas where AI can shape someone's life chances, such as education, employment, credit, public benefits, law enforcement, migration and the courts.
3. **Limited risk: transparency.** People must be told when they are talking to an AI system, unless that is obvious, and when emotion recognition is being used on them. Realistic fake images, audio or video ("deep fakes") must be labeled as artificial.
4. **Minimal risk: no new rules.** Everything else. The Commission says the vast majority of AI systems belong here.

## What "high risk" demands

A high-risk system cannot be sold or used in the EU until it meets a set of requirements. In plain terms, its provider would need to manage risks, use good-quality training data to reduce discriminatory results, keep logs so results can be traced, write detailed documentation, give users clear information, allow meaningful human oversight, and make the system robust, secure and accurate.

The proposal also sets penalties. For the most serious breaches, such as using a banned practice, fines could reach 30 million euros or 6% of a company's worldwide yearly turnover, whichever is higher. For systems that are not high-risk, the proposal encourages voluntary codes of conduct rather than new obligations.

## A worked example: placing everyday tools

Here is how a few familiar tools appear to fit, based on the Commission's own examples and the proposal's annex. This is a reading of a draft, not legal advice.

| Tool | Likely level | Why |
|---|---|---|
| A toy whose voice assistant encourages a child to do something dangerous | Unacceptable | The Commission gives this as an example of banned manipulation |
| Software that screens or sorts job applicants' CVs | High | Recruitment and filtering applications are listed in the annex |
| An app that scores student exams | High | Assessing students is listed in the annex |
| A customer-service chatbot | Limited | Users must be told they are dealing with a machine |
| A website's spam filter | Minimal | Named by the Commission as free to use |
| AI opponents in a video game | Minimal | Also named as free to use |

Notice that the same underlying technology can land in different rows. A text model that writes product descriptions is minimal risk. The same model, used to rank job applicants, would fall under the high-risk rules. The use decides, not the algorithm.

> **Try it:** List five AI-powered tools you or your workplace use, from a photo app to an email filter. For each one, ask: does it decide anything about a person's education, job, money or legal status? Does it talk to people who might think it's human? Place each in one of the four levels and note which ones you are unsure about.

## What do its risk levels mean for everyday tools?

For most of the AI people meet day to day, very little would change. Spam filters, recommendation features, photo enhancers and game characters sit in the minimal-risk group, where the proposal adds no new duties. Their makers may choose to follow a voluntary code of conduct, but they would not have to.

The visible change for ordinary users would come from the transparency rules. If a company runs a chatbot on its website, it would have to make clear that you are talking to software. If you see a convincing fake video of a real person, the person who made or used it would generally have to disclose that it is artificial, with exceptions for things like satire and the arts, subject to safeguards.

The heavier rules land where AI touches important decisions about people. If an employer uses software to filter CVs, or a lender uses AI to judge creditworthiness, the providers of those systems would face the high-risk requirements: documented data, human oversight, logging and accuracy checks. For a job seeker or a loan applicant, that is the part of the proposal that could matter most.

All of this depends on the final text. Expect lobbying, amendments and long negotiations before any rules apply. For now, the four levels offer a useful way to think about any AI tool you use or build: ask what it decides, and about whom.

## Sources

- Proposal for a Regulation laying down harmonised rules on artificial intelligence (Artificial Intelligence Act), COM(2021) 206 final: https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:52021PC0206
- Europe fit for the Digital Age: Commission proposes new rules and actions for excellence and trust in Artificial Intelligence: https://ec.europa.eu/commission/presscorner/detail/en/ip_21_1682