On Thursday, the Commission published templates for AI companies to summarise the data they used to train their models – days before transparency rules for generative AI tools start to apply.
The AI Act’s rules for General Purpose AI models (GPAIs) – such as OpenAI’s ChatGPT, MidJourney or Mistral’s Le Chat – come into force on 2 August, applying legally binding transparency obligations on AI developers.
Training data summaries that will be produced when AI developers fill in the templates are a key component of the law’s push for transparency, as they will require GPAI makers to publicly disclose how their AI models are made – specifying which data went into building their systems.
The Commission’s AI training data template has been eagerly awaited by creative industries, which hope the transparency tool will help them enforce copyright claims against GPAIs.
Under the template released today, AI providers will have to disclose the main datasets they used to train models. They also need to provide a narrative description of data that was scraped from the internet and any other sources of data.
A Commission description of the template said the tool aims to strike a balance between enabling detailed enough disclosure to ensure effective transparency, while also allowing GPAI makers to protect commercially sensitive information.
Ahead of entry into force of the AI Act’s rules for GPAIs on 2 August, the Commission has been expected to publish several documents to support compliance. The template was the last item on the Commission’s to-do list – after guidelines and a GPAI Code of Practice were published earlier this month.
In recent weeks – with time running out before the legal deadline kicks in for GPAIs – industry had been pushing for the Commission to delay implementation. However, the Commission made it clear, multiple times, that the 2 August date stands.
While the GPAI rules become applicable next week, the AI Office, which is the body in charge of enforcing the law, will not do so until August 2026 – giving the AI companies one more year before they could be fined for any breaches. Models that are already on the market have until August 2027 to abide by the rules.
(nl)
