The European Commission published, on Monday 20 July, guidelines intended to help companies developing or providing artificial intelligence (AI) systems comply with the transparency obligations laid down in the AI Act which will enter into force on 2 August.
Providers of AI systems will be required to inform users when they are interacting directly with AI, for example with chatbots, AI agents or avatars, and to design their systems so as to integrate reliable and robust machine-readable markings into the generated content. For their part, deployers will have to indicate content generated or modified by AI, including deepfakes, AI-generated content concerning matters of public interest without human control or editorial oversight, and inform users when they are exposed to AI-based emotion recognition or biometric categorisation systems.
These rules do not apply to people who use AI for personal or non-professional purposes, for example to send a greetings card to their family or friends using AI-generated or AI-modified images. They apply to commercial providers and deployers using these technologies in the course of their professional activities.
The guidelines specify in particular what is meant by ‘interactive AI systems’, such as chatbots, synthetic content - including texts generated partly or entirely by AI - deepfakes and AI-generated texts on matters of public interest. They also provide examples of exceptions, notably standard editing operations, such as spelling and grammatical corrections, in order to facilitate compliance. In order to avoid “label fatigue”, they also specify in which cases information must be provided and how often.
Exemptions. A limited grace period is provided only for AI systems placed on the market before next 2 August, and only with regard to the obligations on marking and detection of AI-generated content. Providers of these systems will have to comply with these obligations only from next 2 December, in accordance with the agreement reached under the 'AI Omnibus’ legislative package (see EUROPE 13864/1).
Content generated before 2 August will not have to be labelled retroactively, although the Commission encourages deployers to do so where possible.
Published texts that have undergone human review or editorial control will not have to be labelled. This applies in particular to content assessed through scientific peer review or validated by an editor-in-chief empowered to approve, amend or reject the content on its substance.
The obligation to indicate synthetic content generated or manipulated by AI does not apply where the AI system performs only an assistance function for standard editing operations. The guidelines provide practical examples of what falls within standard editing and what goes beyond that framework.
People must be informed, clearly and in an easily identifiable manner, that they are interacting with an AI system from the start of the first interaction or first exposure.
As regards deepfakes forming part of works or programmes that are clearly artistic, creative, satirical, fictional or similar, the obligation is limited to appropriate information indicating the presence of manipulated content, without undermining the presentation or appreciation of the work.
The Commission and the European AI Board also approved, by means of an adequacy decision, the code of practice on transparency in AI-generated content, published on 10 July (see EUROPE 13885/7). Drawn up by six independent experts, this code concerns marking techniques and transparency requirements applicable to content generated or manipulated by AI. Detection tools, made available free of charge, will have to comply strictly with the rules of the General Data Protection Regulation (GDPR).
Compliance with these rules will be monitored primarily by the competent national market surveillance authorities. The European AI Office will have jurisdiction only where the same company provides both the model and the AI system, or where the AI system is integrated into a very large online search engine or a very large online platform.
Fines may reach up to €15 million or 3% of the previous financial year’s total worldwide annual turnover. However, they will have to be applied proportionately to small and medium-sized enterprises (SMEs).
To see the document: https://aeur.eu/f/mz3 (Original version in French by Ana Pisonero Hernández)