2021-06-21-安永-Why_successful_adoption_of_AI_will_build_on_a_foundation_of_trust_11页_3mb

Summary of "Making Artificial Intelligence and Machine Learning Trustworthy and Ethical"

Core Content

This document explores the importance of ensuring that AI and Machine Learning (ML) technologies are trustworthy and ethical, emphasizing that while AI offers significant benefits, its implementation must be carefully managed to avoid risks and build public confidence.

Main Points

1. Trust and Ethics in AI/ML

• AI and ML bring transformative potential to businesses and society.
• However, the risks are not inherent in the technology itself but in how it is applied.
• The challenge is to ensure AI is used in a way that is both technically sound and ethically responsible.

2. The Role of Risk-Based Governance

• AI applications must be evaluated based on their use cases and the associated risks.
• The EU AI Act ("AIA") classifies AI based on potential harm to humans, while companies may consider other risk dimensions such as financial or reputational impact.
• A differentiated approach allows for focusing control and governance efforts on high-risk applications without stifling innovation.

3. Technical and Ethical Robustness

• Technical Robustness involves:
  • Ensuring data quality and relevance.
  • Maintaining model stability and accuracy.
  • Preventing cyber-attacks and unintended consequences.
• Ethical Robustness addresses:
  • Fairness and bias in AI outcomes.
  • Human autonomy and psychological impact.
  • Privacy concerns, especially with automated decision-making.

4. Key Risk Considerations

• False Positives vs. False Negatives:
  • Trade-offs must be made depending on the application.
  • For example, in Fruit AI, false positives may reduce profitability, while false negatives could harm reputation.
• Algorithmic Bias and Fairness:
  • Biases in historical data can be amplified by AI.
  • Fairness definitions vary depending on legal, societal, and ethical contexts.
  • Legal frameworks like GDPR and the AIA require human oversight and validation for high-risk applications.

5. Governance and Control Framework

• Organizations must have a clear understanding of their AI use cases and potential risks.
• A well-defined risk assessment process is essential to identify and prioritize high-risk applications.
• The framework should include:
  • Inventory of AI applications and their risks.
  • Implementation of risk-mitigating actions.
  • Internal reporting mechanisms to monitor AI performance and impact.
  • Explainability tools to justify AI decisions, especially in high-stakes scenarios.

6. Monitoring and Validation

• ML Ops (Zoom 1):
  • AI models in production often underperform or misbehave.
  • Monitoring systems must cover:
    • Business performance.
    • Infrastructure and application performance.
    • ML model performance.
  • Monitoring capabilities can be integrated within a ML platform or implemented independently.
• Algorithm Audits and Validation (Zoom 2):
  • Detailed validation processes exist in industries like banking and insurance.
  • These processes include:
    • Performance measurement.
    • Handling imbalanced data.
    • Validation of different learning techniques (supervised, unsupervised, reinforcement learning).
    • Ensuring model stability, convergence, and avoiding overfitting or underfitting.
    • Offline and online performance evaluation.

Key Information

• AI is not inherently risky; it's the use case that determines the risk.
• Explainability is crucial for high-risk applications, especially those involving human decisions or sensitive data.
• Risk-based governance is a practical and flexible approach to managing AI, balancing innovation with responsibility.
• Regulatory frameworks such as the EU AI Act and GDPR are evolving to address ethical and technical risks in AI.
• No-regret moves should be taken immediately to build trust and ensure ethical AI use, even before final regulations are in place.

Conclusion

• Trust in AI depends on how it is implemented and governed.
• A robust framework combining technical and ethical considerations is necessary for sustainable AI adoption.
• Organizations must take proactive steps to monitor, validate, and explain AI outcomes to ensure transparency and accountability.

Zooms Overview

Zoom 1: ML Ops

• Focuses on operationalizing AI models in production.
• Addresses data drift, bias, and security threats.
• Requires monitoring across three layers: business, infrastructure, and model performance.

Zoom 2: Detailed Algorithm Audits

• Involves validating AI models using established practices from other industries.
• Covers various learning techniques and performance metrics.
• Emphasizes the importance of model stability and fairness in outcomes.

EY Contacts

• Patrice Latinne – Partner, Data & Analytics
• Frank De Jonghe – EY EMEA Trusted AI leader
• Amir Krifa – Executive Director, Data Science & AI

EY is committed to building a better working world through trust, innovation, and ethical AI practices.