Why AI Gets Overconfident: Bias, Hallucinations, and Reinforcement Learning Solutions

01 Adversarial Examples

Adversarial examples are tiny, often invisible perturbations to inputs—such as road signs for autonomous driving or subtle changes in images—that cause AI models to misclassify with high confidence, potentially leading to dangerous outcomes.

02 AI Bias

Bias arises from model defects, skewed training data, and algorithms that learn correlations rather than causality. Examples include gender bias in image labeling, resume screening that disfavors women, and misclassifications caused by imbalanced data distributions.

03 AI Hallucinations

Hallucinations occur when AI confidently generates incorrect answers for questions it cannot truly answer, such as predicting future World Cup winners without data. Overconfidence and reliance on statistical patterns amplify this problem.

04 Reinforcement Learning for Safety

Reinforcement learning (RL) can mitigate bias and hallucinations by rewarding models for honest uncertainty (e.g., saying "I don't know") and penalizing confident wrong answers. RL enables AI to learn causal relationships through trial‑and‑error, improving alignment with human values.

Overall, the talk emphasizes that AI safety requires better data, algorithms that capture causality, and RL‑based training to align models with human intentions.