Third-party cyber evaluations involving OpenAI models
openai.com·13h ago
AI Research
arXiv CS.AI·May 28·8 min read
AI, Take the Wheel: What Drives Delegation and Trust in Human-Computer Cooperative Question Answering?
TL;DR
Human-AI question-answering systems often lack clarity about when users should trust AI answers versus seek human help, creating uncertainty in cooperative workflows. Researchers investigated what factors drive users to delegate tasks to AI versus humans in a collaborative QA setting.
✦ Why It Matters
Design AI systems with explicit confidence signals and transparency mechanisms to guide users toward appropriate delegation decisions.
Key Takeaways
- Evaluate when to trust AI suggestions based on confidence levels.
- Implement evidence-based explanations to support AI recommendations.
- Train users to recognize and mitigate confirmation bias.
- Encourage iterative feedback between humans and AI for better outcomes.
How It Works
The study utilized a competitive question-answering game where humans could choose to delegate tasks to AI or adopt its suggestions. By analyzing the decisions made during the matches, researchers identified patterns of reliance and the impact of human biases on collaboration.