AI Act: A major change awaits the giants of AI on the upcoming August 2nd: What is it going to change?

ChatGPT, Claude, and Perplexity have only a few days left to comply with the AI Act, the European regulation on artificial intelligence (AI). After the ban on AI systems with unacceptable risks (such as social rating tools) became effective last February, it is now the turn of general-purpose AI models (called "GPAI" in the European law) to show their capabilities. The AI Act, which came into effect on August 1, 2024, applies gradually until 2030. And starting from August 2, 2025, it is the turn of the "GPAI", models designed to perform multiple tasks such as writing, image generation, research, or calculation, to comply with the European law. To measure what will soon apply to ChatGPT, Le Chat, Gemini, Copilot, Midjourney and Claude, we spoke with Arnaud Latil, author of Digital Law, a risk-based approach, and a lecturer in private law at Sorbonne Université.

Who is affected by the new deadline of the AI Act?

Starting from the next August 2, providers of GPAI models marketed in the European market must comply with the AI Act. These are AI tools capable of generating text, images, code, analyses, written transcriptions... which can also be integrated into other tools.

To know which models are concerned, there is a « presumption linked to the computing power used, but it is only a presumption, and beyond that, the European Commission can classify them based on a long list of criteria », explains the law doctor. It will depend on how the Commission interprets this list of criteria.

The « guidelines », a document published mid-July by the European Commission, specify more precisely which models are concerned. Brussels has set thresholds from which an AI model would fall into the category of « general AI providers »: « its computing resources for training (must be) greater than 1023 (floating-point operations) FLOP and it (must be able) to generate language (whether in text or audio form), text to image or text to video », details the text.

According to all these elements, the AI models concerned by the August 2 deadline will therefore likely be all generative AI models used by the general public from late 2022, namely: ChatGPT (OpenAI), Copilot (Microsoft), Le Chat (Mistral), Claude (Anthropic), Llama (Meta), and their competitors.

What will they be required to do in practice?

Starting from August 2nd, these GPAI must comply with transparency obligations. They will notably need to prepare a whole body of technical documentation that can be shared with the « AI Office », an entity of the European Commission responsible for implementing the AI Act, and the relevant national authorities, upon request. Downstream providers — those who use these tools for a specific application — will also have access to them.

Developers of GPAI also have the obligation to respect copyright, and to complete the summary of sources used model for training their AI tools, a model published by Brussels last Thursday.

Finally, for GPAI that « present a systemic risk », additional obligations of risk assessments and risk mitigations will need to be complied with. These are AI tools that could have a significant impact on fundamental rights, society, or public health. To comply with these new obligations, AI providers can rely on three other texts published throughout July, which came to clarify the AI Act:

• the code of good practices, whose final version was published on Thursday, July 10,
• the guidelines (published on Friday, July 18),
• the model for copyrighted training sources, published on Thursday, July 24.

Also read: After the AI Act, Brussels publishes a user guide for AI giants

Copyright: From August 2, AI will have to properly disclose their sources in Europe

These documents were expected for several months. For some, the drafting of these texts, which went through a consultation phase with AI giants, rights holders, and civil society, could have given the impression that the negotiation of the AI Act was still ongoing. However, Arnaud Latil reminds us that these precision or implementation acts are nothing surprising.

The European Commission can « draft implementation acts or implementation decisions, thus clarifying legal texts. This allows for the evolution of texts without reopening the political negotiation cycle and without necessarily revising the texts as a whole », explains the researcher at the Centre d’études et de recherches en droit de l’immatériel (Université Paris-Saclay).

The Code of Good Practices « is not a mandatory document, but it allows operators to follow recommendations to be compliant » with the AI Act, notes the professor in private law. The copyright-related model, on the other hand, is indeed mandatory. AI providers will have to comply – and for the first time, detail the main sources used to train their models, a first – although rights holders believe the level of precision is far from sufficient.

Will there be a grace period?

The drafting and publication of these three documents took place in the midst of calls to delay or even suspend the application of the AI Act for generative AI tools, calls coming from the AI industry as well as from some countries in the European Union. Business Europe, the European Medef, requested on July 17th « a grace period to implement the Code of Good Practices for general-purpose models », preceding a similar message from forty European companies and organizations. All this takes place against the backdrop of a geopolitical context involving Trump administration's customs duty negotiations, which are strongly opposed to such regulation. 

Also read: Customs duties: Will Donald Trump succeed in undermining the key European digital law?

Nevertheless, the European Commission reminded that there would be neither delay nor grace period. « The AI Act was adopted according to a procedure in line with the principles of the rule of law. It seems impossible to decide not to apply this text. The European legislator has decided to apply the AI Act on a given date, knowing that the date itself is subject to political discussion, just like the other provisions », recalls Arnaud Latil.

Following the same logic, not signing the code of good practices (non-mandatory) – an approach taken by Meta on July 18 – does not change the application of the AI Act. However, there will be an advantage to being a signatory, which will include the French startup Mistral and OpenAI, the American company behind ChatGPT, according to their statements: signing will allow for a presumption of compliance.

According to a journalist from MLex, the European Commission suggested that a grace period would be granted to signatories of the code for the first year. « In particular, if these service providers do not implement all commitments immediately after signing the code, the AI Office will not consider that they have breached their commitments under the code and will not reproach them for violating the AI Act. On the contrary, in such cases, the AI Office will consider that they acted in good faith and be willing to collaborate to find ways to ensure full compliance », he wrote.

In terms of copyright, rights holders and always dissatisfied authors

For the August 2 deadline, the issue of copyright was intensely discussed. As Arnaud Latil recalls, copyright is « an extremely important issue, both for rights holders, for authors and for producers of these models. On one hand, authors want to know whether their works have been used in the training of an AI model ».

If the answer is positive, authors and rights holders want to either oppose it by exercising their right of « opt out », a right provided for in the 2019 directive which allows rights holders to refuse that their works are used to train an AI, or to negotiate a copyright licence. But they must have access to this information – which has not been the case until now.

« On the other hand, producers of these AI models want to protect trade secrets. As is often the case, the legislator sought a balance: only works subject to copyright must be disclosed, not all data", explains the private law doctor.

And if the AI Act requires AI providers to comply with copyright and related rights laws, the three annexed documents published in July by the European Commission left the rights holders hungry. The copyright-related model requires AI providers to disclose all "large" public datasets, as well as the 10% most commonly used domain names, but it is not an exhaustive list.

The code of good practices, which includes a specific section on copyright, reminds that generative AI providers are indeed required to respect copyright, without providing for an obligatory licensing mechanism. The previous mentions of the obligation to respect copyright have been removed. It is insufficient for the rights holders who hoped to base their claims on this list to demand copyright payments or damages.

For Patrick Kuban, co-president of the international UVA (United Voice Artists) federation and animator of the French collective hashtag#TouchePasMaVF, « the level of detail in the collected data is unfortunately very low (…). Rights holders will continue to be wronged. Europe is bending under the pressure of lobbyists that are essentially non-European ». For the voiceover actor, « the level of detail required in the Dataset summaries (training data of IAG) is very insufficient. The promises of the IA Act, meant to protect creators, are not up to par» ». 

This « summary » must be published by AI providers starting from August 2, 2025, and « no later than August 2, 2027 ». It is a delay far too long for rights holders, some of whom argue for a reform of copyright law described as inadequate in the face of generative AI.

Will sanctions fall as of August 3? 

If AI providers do not fulfill their obligations, the sanctions are particularly harsh, as they can reach up to 3% of global annual revenue or 15 million euros. Nevertheless, the fines should not come into effect as of August 3, 2025.

For Arnaud Latil, « regulators will not control and sanction the actors the day after the implementation of the texts. If we take a historical parallel with the RGPD, it came into effect in 2018. On the day after its application, we did not see the CNIL knocking on every door to check all actors. On one hand, there is a date for the implementation of the texts, and on the other hand, there is also the time for regulatory action. The regulator has several missions: information, advice, control, and finally sanctions. The sanction does not intervene before the operations that I have just described ».

Could the geopolitical context and the pressure from the Trump administration change the situation? The specialist estimates that « one can be a little bit worried when one sees the unraveling of barely born environmental regulations. And for the digital sector, there is an equivalent risk ».

👉🏻 Follow the latest tech news in real time: add 01net to your Google sources, and subscribe to our WhatsApp channel.