ANITI publie un position paper sur les systèmes d'IA générative — ActuIA

Summary

ANITI, one of France's four interdisciplinary AI institutes, has published a position statement calling for a scientifically grounded and sovereignty-conscious approach to generative AI. The institute argues that current large language models (LLMs) remain too fragile and insufficiently understood for critical applications. The text demands intensified research into hybrid learning and stronger regulations before deploying generative systems in safety-critical sectors such as transport and industry.

Why it matters

The text explicitly questions GAFAM dominance in the LLM development process, framing it as a threat to Europe's scientific sovereignty. It argues that industrial logic prioritizes operational expertise over genuine scientific understanding. ANITI's critique, rooted in hybrid AI expertise (combining machine learning with constraint-based reasoning), offers a concrete alternative research agenda rather than generic policy platitudes. This resonates with any European institution seeking to build reliable AI beyond the US cloud paradigm.

Highlights

• ANITI warns that LLMs are 'too fragile to be entrusted with fundamental tasks' and contain open scientific questions regarding why they work.
• The text highlights that training data may violate personal data laws, intellectual property rights, and trade secret confidentiality, calling for legal audits before critical deployment.
• Hybrid learning (machine learning + logical/constraint reasoning) is positioned as the primary research path for explainability and reliability in safety-critical sectors.

Critical take

While the arguments for sovereignty and hybrid AI are well-founded, the text largely reads as a roadmap legitimizing ANITI's second funding cycle. The call to 'anticipate and accompany' transformations through social sciences risks remaining abstract without concrete governance mechanisms or benchmarks. The critique of GAFAM dominance, while politically resonant, lacks sufficient depth. The text fails to provide specific technical or policy levers for European laboratories to genuinely compete at LLM scale.