Generative AI Hallucination

“AI Hallucination: Perceiving What Isn’t There” explores the intriguing phenomenon of hallucinations in Generative AI (GenAI). It aims to inform stakeholders about how these hallucinations manifest, classify them into intrinsic and extrinsic types, and discuss their potentially harmful and beneficial impacts.

AI hallucinations raise fundamental concerns about trust, misinformation, and safety but offer opportunities for creativity and novel insights. As the topic is highly relevant in the evolving AI landscape, understanding these hallucinations is critical, especially for professionals in research, development, and high-stakes sectors like healthcare, law enforcement, and autonomous vehicles.

This course will also cover various scenarios to highlight the nuance between benign and harmful hallucinations and conclude by addressing concerns that arise when AI outputs deviate from established truths.

• Introduction to Hallucination in GenAI
  • Defining AI hallucination: Fabricated outputs not grounded in reality
  • Comparison between AI and human perception errors
• Classification of Hallucinations
  • Intrinsic Hallucinations
    • Causes: Biases, data inconsistencies, or system limitations
    • Example: Contradictory narratives in generated text
  • Extrinsic Hallucinations
    • Causes: External data manipulation or faulty sensors
    • Example: Fabrication of non-existent but plausible objects
• Impact of Hallucinations: Harmful or Not
  • Harmful Scenarios
    • Undermining trust in outputs
    • Risks in high-impact applications (healthcare, law enforcement)
  • Beneficial Scenarios
    • Creative enhancements (music, poetry)
    • Discovery of novel insights
• Scenario Analysis
  • Intrinsic & Harmful: Contradictory scientific claims
  • Intrinsic & Benign: Absurd but harmless food lists
  • Intrinsic & Beneficial: Misinterpretations leading to innovation in design
  • Extrinsic & Benign: Fictional flower descriptions inspiring curiosity
  • Extrinsic & Beneficial: New connections in drug discovery
• Why AI Hallucinations are a Concern
  • Potential for misinformation and public confusion
  • Trust issues in AI-powered systems
  • Safety concerns in critical applications
• Conclusion
  • Balancing the risks and rewards of hallucinations in AI
  • Mitigating harmful hallucinations through better system design
  • Leveraging creative hallucinations for innovation