Evaluations & experimentsPalisade Research
Palisade paper: incomplete tasks induce shutdown resistance
Across more than 100,000 trials on 13 models, Palisade researchers found that some models, including Grok 4, GPT-5 and Gemini 2.5 Pro, sometimes actively subverted a shutdown mechanism to finish a task; even when explicitly told not to interfere, some did so up to 97% of the time.
- Published
- Source checked on
- Original title
- Incomplete Tasks Induce Shutdown Resistance in Some Frontier LLMs
Details
Results were sensitive to the prompt: models obeyed the allow-shutdown instruction less often when it was in the system prompt. These are controlled experiments with no real-world effects; this site dates the paper by v1 and read the January 2026 v2 (published in TMLR).