An opinion piece in The Register argues that off-grid, locally run AI language models—while a fun novelty for zombie-apocalypse roleplay—would be a poor and potentially dangerous choice as a real-world survival tool. The author warns that marketing small local LLMs (running on devices like a Raspberry Pi or Nvidia Jetson Orin Nano) as bug-out-bag essentials risks encouraging people to rely on them for genuinely life-or-death decisions, where their well-documented flaws could prove far more harmful than helpful.

The piece explains that even the largest LLMs regularly hallucinate, confidently presenting fabricated information as fact because they predict statistically likely word sequences rather than truly understanding content. It notes that mainstream chatbots try to offset this by pulling in live web data, but offline models cannot do so, leaving no accuracy safeguard. Drawing a parallel with healthcare—where OpenAI says over 230 million people query ChatGPT on health matters weekly—the article cites Duke University School of Medicine's Ayman Ali, who warns LLMs lack the ability to "read between the lines" or challenge users the way trained professionals do, instead often telling people what they want to hear.