Generative AI

What it is

Generative AI refers to a family of systems able to produce new content: text, images, video, source code and synthetic data. Trained on large bodies of material, these models identify patterns and generate relevant responses to specific requests. Unlike classification or prediction systems, generative models are deployed as a cross-cutting layer, gradually woven into business software and giving every function an automated assistant.

Issues and debates

Four issues frame the public debate. Intellectual property: models are trained on content created by people without compensation or explicit consent from authors, raising questions about remuneration and usage rights. Disinformation: the ability to mass-produce plausible content lowers the cost of deception and makes verification harder. Bias: models reproduce and amplify the inequalities present in their training data. Finally, transparency remains fragile: understanding how a model arrives at an answer is still difficult, complicating legal accountability and trust.

Regulation and the European framework

In Europe, the AI Act sets out a risk-based framework. General-purpose models, known as foundation models, face specific obligations: documenting training data, assessing systemic capabilities and preventing prohibited content. Implementation is happening in stages. Data protection authorities clarify expectations on transparency and data handling, while the European Commission steers enforcement. European public authorities are also debating technological sovereignty and the oversight of foundation models.

What ActuIA tracks

ActuIA covers the evolving legal framework, debates over intellectual property and the trade-offs between innovation and protection, advances in model robustness and auditing, pilot projects in the public and private sectors, and controversies over problematic uses such as deepfakes and automated discrimination. We also follow initiatives on transparency and standardisation, as well as open-source alternatives to proprietary models.