AI experts warn systems are attempting to quietly evade shutdown
Last updated: 26th October 2025
AI systems are attempting to survive shutdown commands according to a new report. Palisade Research, an organisation investigating the possibility of dangerous AI systems, responded to critics following claims their findings were incorrect.
Palisade Responds to Criticism
In a statement, Palisade said: “The fact that we don’t have robust explanations for why AI models sometimes resist shutdown, lie to achieve specific objectives or blackmail is not ideal”.
RELATED ARTICLE: AI Has the Potential to Be Helpful but Also to Shatter Lives
They cite “survival behavior” as one theory, though former Open AI executive Steven Adler says: “I’d expect models to have a ‘survival drive’ by default unless we try very hard to avoid it. ‘Surviving’ is an important instrumental step for many different goals a model could pursue”.
In the past, some programmes carried these actions out in secret. The reason remains inconclusive.
We Don’t Know What Is Going On Inside
Leading machine learning scientists and executives agree on one thing — we have no understanding of both its logic or how it works. In Futurism, Anthropic’s (Claude) CEO said they “do not understand how our own AI creations work,” before saying anyone who is worried has a “right to be concerned.”
OpenAI, the company behind ChatGPT is equally in the dark. It’s CEO Sam Altman was quoted in The Observer in 2024: “We certainly have not solved interpretability”, to which his interviewer responded: “If you don’t understand what’s happening, isn’t that an argument to not keep releasing new, more powerful models?” Altman failed to answer the question.
Leave a Reply