In addition to the AI Act, the EU will also add additional rule constraints to higher capacity models such as GPT-4

Thank CTOnews.com netizens soft media Xinyou 2303894 for the clue delivery! CTOnews.com, December 10 (Xinhua) this weekend, the European Union reached a preliminary agreement on the world's first AI regulation bill, "AI Act", involving general artificial intelligence systems, including ChatGPT, which is regarded as a key part of the world's first comprehensive regulation of AI.

According to an EU document obtained by Bloomberg, developers of all general-purpose artificial intelligence systems (or powerful models that are likely to be widely used) must meet basic transparency requirements unless the systems are free and open source, Tuyuan Pixabay reported on Saturday night local time.

The relevant requirements for CTOnews.com are as follows:

Develop an acceptable use policy

Update information about model training at any time

Summarize the data used in the training model and report the detailed summary.

Formulate policies that respect the copyright Law

According to the document, models that are considered to constitute a "systemic risk" will be subject to additional rules. According to the report, the European Union will determine whether there is a risk based on the computing power of the training model, and the threshold set for the model is more than 10 trillion operations per second. Some experts say that the only model that can reach this threshold is OpenAI's GPT-4.

In addition, EU implementing agencies can specify other thresholds based on possible indicators such as the size of the dataset, whether there are at least 10000 registered business users in the EU, and the final number of registered users.

The report also said that during the European Commission to develop more coordinated and long-term effective control measures, these "more capable" models should sign a code of conduct. These models must also meet the following requirements:

Take the initiative to report your energy consumption

Conduct red team tests / adversarial tests internally or externally

Assess and mitigate possible systemic risks and report any events

Ensure the use of appropriate network security controls

Report the information used to fine-tune the model and its system structure

If more energy-efficient standards are established, the development process should meet the new standards

As for models that do not sign a code of conduct, they must prove to the European Commission that they will indeed comply with the AI Act. It should be noted that the exemption for open source models does not apply to models that are "considered to pose systemic risks".