Pré-Publication, Document De Travail Année : 2025

LLM Grooming: A New Cognitive Threat to Generative AI

Didier Danet
  • Fonction : Auteur
  • PersonId : 1579084

Résumé

The rapid development of large language models (LLMs) has given rise to a new form of informational and cognitive manipulation: LLM Grooming. This emerging threat refers to the large-scale contamination of training data with biased or deceptive content, thereby transforming generative AI systems into powerful vectors of disinformation. Unlike classical data poisoning-which corrupts developer-curated datasets-LLM Grooming operates through the systematic flooding of publicly accessible data sources, such as web archives, with fabricated narratives. Once incorporated during training or through Retrieval-Augmented Generation (RAG), these narratives are statistically assimilated by the model and reproduced as seemingly factual knowledge. This article conceptualises LLM Grooming as a distinct category of cognitive threat, beyond traditional information warfare. It explores its dual manifestation: in the restricted sense, as the contamination of training corpora prior to model deployment, and in the broader sense, through malicious prompts and RAG exploitation post-deployment. The paper examines emblematic cases such as the Pravda network, which generated over 3.6 million articles in a single year, measurably contaminating major Western LLMs. It further analyses the structural vulnerabilities of different architectures, with particular emphasis on the heightened risks faced by RAGbased systems. The findings highlight the inadequacy of conventional counter-disinformation measures, which are largely reactive and ill-suited to systemic contamination. The paper calls for proactive strategies, including dataset auditing, continuous monitoring, robust human-in-the-loop mechanisms, and international cooperation. By framing LLM Grooming as both an informational and a cognitive threat, the article argues for a paradigm shift in AI security to preserve the integrity of decision-making processes in digital societies.

Fichier principal
Vignette du fichier
20251215 DANET Cognitive Threats LLM Grooming V02.pdf (635.53 Ko) Télécharger le fichier
Origine Fichiers produits par l'(les) auteur(s)
Licence

Dates et versions

hal-05241525 , version 1 (05-09-2025)

Licence

Identifiants

  • HAL Id : hal-05241525 , version 1

Citer

Didier Danet. LLM Grooming: A New Cognitive Threat to Generative AI. 2025. ⟨hal-05241525⟩
1323 Consultations
1201 Téléchargements

Partager

  • More