What the rules on AI approved by the EU Parliament provide for

With 499 votes in favour, 28 against and 93 abstentions, yesterday the European Parliament approved the“AI Act”, a package of rules that should regulate the vast field of artificial intelligence. The Strasbourg vote will now be followed by the “trilogue”, i.e. the discussion with the other European institutions (Council and Commission), from which the final text, which will hopefully be approved by the end of this year. In any case, these will be minor adjustments and therefore, although not definitive, the structure of yesterday’s document will remain roughly unchanged.

The main innovation contained in the “AI Act” is the introduction of four different risk classeswith which the various technologies will be marked: ranging from “unacceptable risk” to “limited risk”, passing through a “high” risk range and a separate category, reserved for generative AI (of the ChatGPT type).

The unacceptable risk and fines

For the highest level of risk – the so-called “red dot” level -, the Union establishes an absolute prohibition of use and provides for an administrative penalty in the event that this restriction is circumvented: we are talking about a fine that can reach up to 30 million euros or, alternatively, a figure equal to 6 percent of the company’s annual turnover.

Among the technologies considered most dangerous, they include all “real-time” biometric identification systems: that is, we are dealing with those technologies that exploit artificial intelligence for facial recognition in outdoor or public places. The police forces will be able to use them, but only for the prosecution of “serious crimes” and subject to judicial authorization: the amendments proposed by the EPP, which asked for greater flexibility, were rejected by Parliament.

In addition to facial recognition, the technologies banned by the Union include those that involve a cognitive-behavioral manipulation of people (also for electoral purposes): The example cited is ‘voice-activated toys that encourage dangerous behavior in children’. Finally, systems that include a “social score”, classifying users on the basis of personal characteristics.

High and limited risk

A ban is not envisaged for high-risk technologies, but the EU will undertake an evaluation before they are placed on the market. The Union will then establish a database, where these will be divided into eight different categories based on areas of interest (for example education or border controls). In this way, Europe will monitor the evolution of specific systems from time to time.

As regards the lowest risk range, Parliament has only provided for “minimum” transparency obligationsso as to allow “users to make informed decisions”.

Generative AI (ChatGPT)

Finally, for systems on the ChatGPT model, the EU requires compliance with certain requirements transparencyfor example forcing users to declare that a text has been produced using artificial intelligence. Furthermore, the Union requests that these technologies be designed “avoiding the generation of illegal content”.