New transparency requirements for certain artificial intelligence systems have come into force in the European Union on 2 August under the AI Act. Developers and users of such systems must disclose interactions with AI and label certain types of synthetic content, including deepfakes. Companies face fines of up to €15 million or 3% of their global annual turnover for non-compliance.
Briefly about the main points
- The new requirements mark a stage in the phased implementation of the AI Act.
- Deepfakes require a visible and machine-readable label.
- Chatbots, AI agents and avatars must identify themselves.
- The rules also cover certain AI-generated texts and biometric tools.
- Regulatory bodies will be able to inspect documentation and restrict models.
The rules are intended to distinguish synthetic content from authentic content
The European Commission explains that the rapid AI development makes it difficult to distinguish between content created or modified by algorithms and content produced by humans. This creates risks of large-scale manipulation, disinformation, fraud, identity impersonation and the misleading of consumers.
AI Act It came into force on 1 August 2024, but its provisions are being introduced in stages. A new stage does not mean the simultaneous full implementation of all the regulations’ provisions: it primarily concerns transparency obligations for specific categories of systems and content.
The aim of these requirements is to enable people to understand whether they are interacting with an automated system and whether they are viewing AI-generated or manipulated content, so that they can make informed decisions.
What exactly should be disclosed to users and the audience?
Images, audio and video material that mimic real people, objects, places, organisations or events must be clearly and conspicuously labelled as deepfakes. At the same time, they must contain machine-readable labels, in particular a digital «watermark».
Users should also be clearly informed that they are interacting not with a real person, but with an AI system. This requirement applies, in particular, to chatbots, AI agents and avatars. An exception may be made where the nature of the interaction with the AI is evident from the circumstances.
Separate rules apply to emotion recognition and biometric categorisation tools. Labelling is also required for texts generated or modified by AI to inform the public on matters of public interest, if they are published without human verification or editorial oversight.
The Commission has issued guidelines explaining how compliance can be demonstrated. A voluntary code of practice may be one such tool, but it does not, in itself, replace compliance with statutory requirements.
Who will monitor compliance with the requirements?
Compliance with the rules will be monitored by national market surveillance authorities, the European AI Office for systems falling within its remit, and the European Data Protection Supervisor where EU institutions are the providers or users of such systems.
For the institutions, bodies and agencies of the European Union the maximum penalty is up to 750,000 euros. When determining penalties for small and medium-sized enterprises and small mid-cap companies, the principle of proportionality must be taken into account.
The European Commission’s AI Office may review documentation, require restrictions to be placed on models and withdraw them from the market. For market participants, this means they will need to document exactly how they fulfil their disclosure obligations.
