The keys to the new transparency rules required by the EU for AI

Brussels.- The European Union’s (EU) transparency rules for artificial intelligence (AI) systems are now in effect (AI), requiring them to warn users when they are interacting with a machine or when the content they are viewing has been manipulated.

In addition, the European Commission’s Artificial Intelligence Office has new powers to oversee generative AI models and impose fines on them if they fail to comply with their obligations.

The new rules, however, have several exceptions that delay their mandatory application in certain cases.

These are the keys:

1. Immediate transparency requirements

Companies such as Open AI (owner of ChatGPT), Google (Gemini), or Anthropic (Claude) must design their systems in such a way that users know they are interacting with a machine.

The obligation is also aimed at developers of artificial intelligence systems used in public services, those used by companies to interact with their customers, or, for example, those intended to manage fraud reports.

Those who use artificial intelligence systems for emotion recognition or biometric categorization must also notify citizens when they are being analyzed with these types of tools, with the exception of those permitted by law to prevent or prosecute crimes.

2. Requirements and prohibitions in December

Starting December 2nd, artificial intelligence systems will also have to incorporate watermarks for texts, photos, videos, and audio that have been manipulated, so that citizens can easily identify that a machine created them.

However, systems such as ChatGPT, Gemini, Claude, or social networks that have AI tools will not be legally responsible for the content created by users; instead, it must be the users themselves who report that the content has been manipulated.

All fictional, artistic, or satirical content is exempt from this requirement.

We recommend reading: Nvidia cedes the throne to Apple, AI redesigns proteins and other tech clicks from America

There is an exception to this rule: as of today, all content generated with artificial intelligence that displays information of general interest, such as that related to electoral or judicial processes, public health, or any other event that may be the subject of public debate, must be labeled, unless it has been previously supervised by a person.

Beyond these transparency requirements, starting December 2nd, the creation of sexualized images generated with artificial intelligence will be prohibited, without the consent of the affected individuals.

A veto that EU institutions agreed upon following the scandal caused by images generated with Grok, the artificial intelligence tool from the social network X, which will now have until the end of the year to design appropriate protection measures.

3. Threats

Since August 2nd, the European Commission’s Artificial Intelligence Office has the authority to require companies like Anthropic to grant them access to models such as Mythos to evaluate the risks that these types of high-risk systems may cause.

Community sources, however, have avoided detailing what their next steps will be in a case like this, for “obvious reasons”.

Even so, there is doubt as to the extent to which the European Commission has the capacity to analyze high-risk models of this type, or to what extent it will be able to access them if the US government bans foreign citizens from them again, as it did temporarily in mid-June.

Brussels is currently negotiating with Anthropic the conditions for the EU Agency for Cybersecurity (ENISA) to evaluate Mythos without putting European essential infrastructure or potential access to sensitive information at risk.

4. Fines

The European Commission may impose fines of 15 million euros or up to 3% of their total annual global turnover (whichever is higher) on generative artificial intelligence systems that fail to comply with these obligations.

The penalty will be up to 750,000 euros if it is the EU institutions or agencies that fail to comply with their transparency obligations.