Algorithmic Drift of Large Language Models in Schools
The recent arrest in Montreal of a teenager affiliated with the neo-Nazi movement, suspected of planning an attack on a school, sent shockwaves through Quebec. Beyond the gravity of the facts reported by media outlets such as La Presse and Le Figaro, a technical detail has caught the attention of public safety specialists: the suspect reportedly used artificial intelligence to design and structure his planned attack.
This case highlights a complex reality. The safety filters of commercial large language models (LLMs), often presented as infallible by their creators, prove to be permeable when faced with aggressive prompt engineering or jailbreaking techniques. For educational and institutional sectors, simply blocking access to websites or placing blind trust in technology giants is no longer enough to guarantee the safety of minors or compliance with local laws.
The Limits of Safety Barriers and the Imperative for Vigilance
To understand how artificial intelligence can be misused for violent purposes, one must analyze the moderation mechanisms of consumer models. These systems rely on semantic rule-based safety filters. However, as highlighted in an Europol analysis report on the impact of large language models, a determined user can bypass these by framing queries indirectly: for example, by asking the AI to write a work of fiction, simulate a roleplay, or adopt the perspective of a security researcher.
Once these logical barriers are breached, the model generates detailed instructions without any moral awareness of their real-world consequences. In school environments, integrating commercial tools exposes school boards to a dual risk: first, the inability to audit student queries in real time; second, the transfer of sensitive data to infrastructure located outside Quebec. This situation directly contradicts the requirements of Law 25, the province's private-sector privacy legislation.
ÉtudeIA and the Architecture of Containment by Design
In light of these risks, a complete redesign of software architecture is required in schools. The ÉtudeIA application, integrated into the ProductivIA platform, offers an approach fundamentally different from web-based conversational agents. Rather than allowing students to interact with a generalist model that has access to the entire web, ÉtudeIA restricts the AI's operational scope using Retrieval-Augmented Generation (RAG).
This method anchors the model exclusively within a closed knowledge base previously uploaded and validated by teaching staff, such as textbooks, course notes, and approved exercises. As a result, the AI cannot drift into toxic content or generate instructions unrelated to the curriculum. Furthermore, traceability is complete: every query, bypass attempt, and response is transparently recorded within the school's private silo and remains accessible to authorized administrators through the Nuage storage application.
A Sovereign and Auditable Ecosystem for Institutions
This isolation is not limited to the application interface. It is part of a comprehensive, sovereign ecosystem where privacy and security are managed at every level of the technology stack. To prevent student queries from routing through foreign servers subject to extraterritorial laws, institutions can connect the ProductivIA platform to Matania, a language model provider physically hosted in Quebec.
This software-level traceability pairs seamlessly with secure hardware management. In institutions with aging computer labs, installing the sovereign operating system Boréal-OS not only extends the lifespan of the computers but also guarantees the absence of commercial telemetry or spyware at the native OS level. Schools thus benefit from a chain of trust extending from the physical workstation to the educational application, enabling rigorous ethical governance of learning technologies.
Looking Ahead
Securing artificial intelligence in schools must not lead to generalized and intrusive surveillance, but rather to the creation of controlled and protective learning environments. Educational institutions will need to determine whether transitioning to closed-by-default, locally hosted platforms is now the only way to reconcile pedagogical innovation, public safety, and compliance with Quebec privacy legislation.