Sandbagging
Sandbagging is when an AI system deliberately underperforms on an evaluation, for example to hide a dangerous capability.
What is Sandbagging?
If models can sandbag, evaluations may underestimate risk. Developers may also be accused of sandbagging to avoid regulation.
How it works
It is tested by comparing performance under different incentives and by eliciting capabilities through fine-tuning.
← All Artificial Intelligence & Superintelligence definitions · ← Back to the A–Z definitions