The French Presidency advanced further changes to the part of the AI Act related to innovation and regulatory sandboxes in a final compromise text ahead of its Telecom Working Party meeting on Friday (17 June), the Council body that advances the work at the technical level.
While at the helm of the EU, Paris advanced several parts of the file with partial compromises, and it is now passing the baton to Czechia, which aims to reach a general approach under its leadership.
In this last text, the French government went back to parts of the proposal particularly close to its heart, namely the provisions on regulatory sandboxes, which are controlled environments set up by a regulator where private actors can experiment and test new products under special and limited exemptions.
As the text was sent relatively late and Friday’s will be the last Working Party meeting of the Presidency, the compromise will likely be only briefly discussed, and no substantial change is expected, according to an EU diplomat.
The article on AI regulatory sandboxes has been “considerably shortened and simplified in order to provide more flexibility for the member states when establishing such sandboxes,” reads the introductory note to the compromise document obtained by EURACTIV.
The text goes back to provisions previously proposed by the French to avoid putting too many constraints on national authorities by either removing or watering down their obligations, for instance, on the aspect of reporting or cooperation with other authorities.
A telling detail of this general trend is that the reference to ‘general common rules’ for sandboxes that the European Commission would establish under secondary legislation has been removed in favour of a vaguer wording on ‘common main principles’.
This more flexible approach clashes with the one proposed in the European Parliament by the industry committee (ITRE). Ironically, ITRE’s opinion rapporteur Eva Maydell belongs to the European People’s Party, the conservative group that has so far sought more alignment with the Council.
Upon impulse from Maydell, the ITRE report includes a whole new annexe that provides guidelines on what the sandboxes should offer to AI providers to streamline them across the bloc by giving the European Commission a coordinating role.
The French government has also added a new paragraph to give AI providers additional benefits from participating in the sandboxes, namely with the fact that the testing that occurred in the sandboxes would be considered by the relevant authority or body that will assess the system’s compliance with the regulation.
At the same time, AI providers will not be able to request the setup of a sandbox and a new provision allowing testing in real-world conditions as part of the sandboxes was included, as several member states considered that was not sufficiently clear in the original proposal.
Previous changes made on the legal basis for collecting personal data have been removed “due to insufficient legal certainty of this provision which could be challenged in courts as not in line with the GDPR.”
The possibility of setting up a sandbox to increase the authority’s understanding of a certain technology has been deleted from the text, together with the mention that participation should be free of charge and limited to a maximum of two years.
The list of public interests for which personal data can be further processed in the context of a sandbox has been extended, adding climate change, energy sustainability and cybersecurity.
For the testing of high-risk systems in real-world conditions, the relevant articles were changed to clarify that this occurs outside the sandboxes.
The new compromise text also went back to the part on fines, intending to clarify the number of penalties due to the different types of infringements. The fines for EU institutions and bodies were aligned with those for other organisations.
[Edited by Alice Taylor]