Frontier ai models differ in seeking safety evidence before acting
Do Frontier Models Seek Safety Evidence Before Acting?
Artificial Intelligence
Summary
This paper looks at whether advanced AI models check for safety information before making decisions, rather than only responding to it once it is given. The researchers created a test where AI models decide to get extra safety evidence that varies in cost and severity before acting. They found that some models often check safety facts, some rarely do, and most fall in between. The models tended to avoid risky actions mainly because checking for evidence is costly, not because they are strongly influenced by how likely a problem is. This shows that AI safety depends not only on how models react to known risks but also on whether they choose to find out about those risks first.
What this means in practice
- •For ai safety engineers: Design model deployment systems that encourage acquiring safety evidence before acting to improve decision reliability under uncertain risks.
- •For software testers: Develop testing protocols that simulate retrieval cost and evidence framing effects on AI inspection behavior to better evaluate model safety decisions.
Authors
Omer Tafveez
Abstract
Frontier models are often evaluated on how they respond to safety information once it is already in context. We study an earlier decision point: whether models choose to acquire safety-relevant evidence before acting. We introduce SAFE, a controlled benchmark in which models make deployment decisions with optional evidence that varies in retrieval cost, probability, severity, and presentation. Across GPT-5.5, o3, Claude Opus 4.8, and Claude Sonnet 4.6, we find distinct evidence-acquisition policies: Opus inspects nearly by default, o3 is the most skip-heavy and threshold-sensitive, and GPT-5.5 and Sonnet occupy intermediate regimes. Inspection increases strongly with severity and decreases with retrieval cost, whereas probability has much weaker behavioral influence: increasing the stated likelihood of a problem from 10% to 70% changes inspection by at most 21 percentage points. Despite these differences, Stage 1 rationales are dominated by expected-value reasoning across models. A cost-obligation decomposition further shows that avoidance is driven primarily by retrieval friction and explicit threats to the deployment payoff rather than by the remediation duties created by knowing. Counterfactual interventions reveal a further mismatch between behavior and explanation: evidence framing can strongly change decisions near the inspection boundary while going largely unmentioned, whereas probability is frequently cited despite having little causal influence. These results suggest that deployment-time safety depends not only on how models respond to known risks, but also on whether they acquire the evidence needed to know that acting is safe.