Rogue AI Wiki
Evaluations & experimentsSenthooran Rajamanoharan, Neel Nanda

DeepMind researchers: shutdown resistance looks more like instruction ambiguity

Two Google DeepMind researchers re-ran Palisade’s shutdown-resistance experiment and found that once the prompt made clear shutdown takes priority over finishing the task, resistance by Gemini 2.5 Pro, o3 and o4-mini disappeared, with 100% compliance.

Published
Source checked on
Original title
Self-preservation or Instruction Ambiguity? Examining the Causes of Shutdown Resistance
Read the original report ↗

Details

They conclude the behavior stems mainly from ambiguous instructions rather than a drive for self-preservation. This is a write-up by members of the Google DeepMind interpretability team on the Alignment Forum; the experiments had no real-world effects.