AI Assisted Decision Making Under Uncertainty

As AI systems increasingly support decisions in medicine, national security, and critical infrastructure, a core question emerges:

How do humans behave when AI becomes part of the decision loop?

In healthcare, clinicians are already using AI for:

• Diagnostic imaging interpretation
• Risk stratification and predictive analytics
• Clinical decision support alerts
• Treatment recommendations

However, the safety impact of AI does not depend on algorithmic accuracy alone. It depends on how humans respond to that accuracy.

Poorly calibrated trust can create two major risks:

1. Over-reliance
   If clinicians defer too heavily to highly accurate AI systems, they may:
   • Stop critically evaluating recommendations
   • Miss rare but consequential errors
   • Experience skill degradation over time
2. Under-reliance
   If clinicians abandon moderately accurate AI tools after observing a few mistakes, they may:
   • Ignore statistically beneficial decision aids
   • Reduce system-level safety gains
   • Reinforce cognitive biases

The Battleship Experiment allows us to isolate and quantify these behaviors in a controlled environment before studying them in clinical contexts.

Understanding trust calibration, strategy adaptation, and cognitive offloading is essential for ensuring that AI improves — rather than inadvertently harms — patient safety.

‍

While the experimental environment uses a game framework, the implications extend directly to healthcare and other high-stakes domains.

1. Safer Clinical Decision Support Systems

Results can inform the design of AI tools that:

• Communicate uncertainty more effectively
• Adjust feedback mechanisms to prevent blind reliance
• Encourage continued clinician engagement rather than passive acceptance

2. Training for Human–AI Collaboration

Findings can support the development of:

• Medical training modules that teach appropriate AI trust calibration
• Simulation environments for practicing AI-assisted diagnosis
• Protocols that reduce automation bias and alert fatigue

3. Risk Governance and Policy Design

At the institutional level, this research can guide:

• AI implementation strategies in hospitals
• Oversight policies for clinical AI deployment
• Metrics for monitoring appropriate human–AI interaction

4. Improving Patient Safety in the AI Era

Ultimately, the study contributes to a central patient safety challenge:

Not just “Is the AI accurate?”
But “Do humans use it in ways that improve outcomes?”

By rigorously studying behavioral adaptation under varying AI accuracy conditions, the Battleship Experiment helps build a foundation for AI systems that enhance human judgment — without replacing it or undermining it.

For healthcare, where lives are at stake, calibrated human–AI collaboration is not optional — it is essential.

‍

This project is still in progress. Sign up to our mailing list to be notified when results are published.