Responsible AI

Refers to practices and principles that ensure that AI systems are transparent and trustworthy

• Be accountable
• Be managed by a team accountable for responsible strategies
• Developed by teams with expertise in responsible AI
• Built following responsible AI Guidelines

Problem Areas

• Accuracy (Bias and Variance)
  • Lack of diversity in data set can lead to a bias in the model and it is underfitted. This means it is not capturing enough difference in the features of data
  • Variance happens when the model is to sensitive and starts to give value to noise. Whenever it is introduced to new data it will perform poorly and accuracy drops. This data is overfitted
  • Fixes:
    • Cross validation: Train multiple models on sub-sets of data and test against complementary sets. This will help detect over fitting
    • More Data
    • Regularization: Penalizes extreme weight values and prevents overfitting
    • Simpler models help with over fitting, if it is under fitting then it may be to simple
    • Dimension Reduction
    • End training early so that AI won't just memorize
• Problems with Generative AI:
  • Potentially toxic and offensive results
  • Hallucinations
  • IP theft
  • Plagiarism and cheating
  • Disruption of the nature of work

Core parts of responsible AI:

• Fairness: Inclusion, uphold responsible values, and build trust
• Explainability: Ability for AI model to provide justification for its internal mechanism "thought process"'
• Privacy and Security: ensure data is safe
• Transparency: Ensure all parties know that an AI system is in use
• Veracity and robustness: Ensure it is reliable
  Transparency helps understand how a model makes decisions while explainablility explains why the model made that decision