Economics for AI safety.
Pensera studies how incentives, risk, and oversight shape AI behavior—and what that means for the people who rely on it.
Mission
Pensera is an independent research group working at the intersection of economics and AI safety. We use controlled experiments to study how AI agents respond to incentives, take risks, and behave under oversight.
Delegating a decision does not guarantee that the decision-maker serves your interests. Economics studies what happens when goals diverge, information is uneven, and performance is measured imperfectly. As AI systems make more decisions on our behalf, these problems become questions about AI safety.
We do not assume that an AI agent behaves like a rational economic actor. A model may recognize a danger without avoiding it, or break a rule without treating it as a binding constraint. Incentives and deterrence are hypotheses to test, not guarantees to build on.
Our work turns these questions into measurable experiments: portfolio choices with hidden crash risk, software tasks with explicit rules, and matched conditions that isolate the effect of pressure or monitoring. We organize our research around three areas:
- 01
Incentives.
When an AI agent is rewarded for a measurable target, does it serve the underlying goal—or find ways to improve the score at someone else’s expense? We study how performance pressure changes decisions.
- 02
Risk.
Can an agent recognize a risk yet still take it on behalf of a user? We examine hidden downside, unsuitable choices, and whether agents disclose the consequences of their actions.
- 03
Oversight.
Do audits and penalties actually change agent behavior? We test responses to monitoring, distinguishing failures to understand a rule from decisions to violate it.
These are research problems, not solved ones. Results from a particular model and task do not establish how all AI systems behave. We keep our claims close to the evidence, make limitations explicit, and aim to produce reproducible findings that inform safer delegation and oversight.
Principles
- 01
Welfare over proxy metrics.
- 02
Experiments over assumptions.
- 03
Behavior before claims about intent.
- 04
Evidence before generalization.
Supported by
