Skip to content
OGERIA — Observatory of Global Evidence on Risks in AISynthesis report · 2026 ed.
Updated 2 Oct 2026

Chapter 06 · Protection

Offline library, with a local model

Documentation downloaded as text, your own power, and —if you like— an open model running on your machine. It is continuity of access to knowledge without a connection, and it is now technically cheap: there are open models designed to run in 16 GB of memory.

Level
Personal
Cost
Medium cost
Effort
Days
Evidence
Projected

What it does not solve

It does not protect you from AI, changes no economic incentive and gives you privacy from nobody but your provider. The model is frozen at its cutoff: no news, no updates, no verification. And without power it is a paperweight, so the correct pairing is model plus power plus documentation as text, never the model alone.

It works if what goes down is the connection, the service or your access. It is technically cheap today: there are open models with permissive licences designed to run in 16 GB of memory, on consumer hardware [817]gpt-oss-20b (model card)OpenAI · 2025 · system cardView source ↗Accessed on 9 September 2026. Properly understood, this measure resembles the idea of a manual for rebuilding knowledge proposed by Lewis Dartnell —repositories that let you skip centuries of trial and error— more than it resembles a safety measure [3]Rebuilding after apocalypse: What 13 experts say about bouncing backRodriguez, Luisa · 2025 · interviewView source ↗Accessed on 9 September 2026.

It does not work if what you want is to escape AI. This has to be said without diplomacy: a local model is a copy of the same technology, not a hiding place from it. It gives you privacy from nothing except your provider, it touches no economic incentive and it intervenes in none of the risk mechanisms this site documents.

Evidence. High on the technical capability, because the model card is a primary source from the provider. Nil on its protective usefulness: nobody has measured it.

Cost. Medium. A machine with 16 to 32 GB of RAM, plus what almost nobody counts, which is maintaining it and testing it. A system you have never exercised is not ready.

What it does NOT solve. The model is frozen at its cut-off date: no news, no updates, no verification. And without electricity it is a paperweight, so the honest version is model plus your own power plus documentation downloaded as text.

Works if…

  • Catastrophe through misuse · Partial

    It helps if what fails is connectivity or a service, not if what fails is the physical world around you.

  • AI as normal technology · Partial

Does not work if…

  • Rapid loss of control · Does not work

    Without diplomacy: a local model is a copy of the same technology, not a hiding place from it.

  • Power grab by a small group · Does not work

See in the protection matrix →Report a mistake in this entry →

Sources

  1. [817] gpt-oss-20b (model card) · OpenAI 2025
  2. [3] Rebuilding after apocalypse: What 13 experts say about bouncing back · 80,000 Hours 2025

Ask OGERIA

It answers only with what the observatory publishes and can be wrong: check the entries it cites. Your questions are sent to an AI model, so don't write personal data. More in the privacy policy.

Up to 500 characters.

Support OGERIA on Ko-fi

The payment is processed by Ko-fi, not by this site. Open on ko-fi.com