Legal and Ethical Challenges in AI

Introduction

The mentorship session organized by UP Data Science Society features a discussion on legal, ethical, and regulatory challenges in AI. Attorney Rebalay Medina is the speaker, who is a lawyer and AI professional with experience in legal and regulatory technology. His goal is to provide awareness, not to make participants legal experts.

Purpose of the Talk

The main audience includes data scientists, AI developers, legal professionals, regulators, and others in the technology industry. The talk covers important laws, regulations, and technical solutions related to the challenges faced in AI.

Key Issues in AI

  1. Bias in AI Systems: Cited the case of the McDaniel Scopells, who faced discrimination in mortgage approval processes due to bias in algorithms, highlighting the serious implications of biased data.

  2. Legal Frameworks: Discusses various laws, such as the 1987 constitution's equal protection clause, anti-discrimination laws, and the importance of fairness in AI systems.

The EU AI Act

The EU AI Act is crucial because it sets significant requirements for high-risk AI systems, including banning systems that exploit vulnerable populations and ensuring high-quality, unbiased datasets.

Technical Solutions for Bias Mitigation

  • Data Techniques: Strategies such as stratified sampling, adjusting sample weight, and synthetic data generation (like SMOTE) improve fairness and mitigate bias in AI models.

  • Ethics in Model Training: Remove or transform features correlated with sensitive attributes to prevent bias in AI decisions.

Intellectual Property Issues

  • Understanding IP: Discusses copyrights, patents, and trademarks concerning AI and how unauthorized use of proprietary material needs to be addressed in AI training datasets.

  • Protecting Data: Methods include digital watermarking and adversarial perturbations to prevent unauthorized use of data.

Data Privacy

Highlighting the importance of data privacy, especially how companies like Clearview AI have faced backlash for scraping personal data without consent. Key principles of privacy laws such as GDPR emphasize the necessity for consent, transparency, and proportional data processing.

Privacy-Enhancing Technologies (PET)

  • Secure Multi-party Computation: Allows data analysis without exposing sensitive information.

  • Homomorphic Encryption: Enables computation on encrypted data without revealing it.

  • Differential Privacy: Adds noise to datasets to protect individual identities while still providing statistical insights.

Explainability & Transparency

  • The importance of having transparent and explainable AI systems to build trust, as illustrated by the failures of IBM's Watson.

  • Emerging methods in explainable AI can help demystify AI decision-making processes to users.

Product Liability

AI systems must address who is liable when they malfunction or cause harm, requiring awareness of laws like the Consumer Product Safety Act.

Final Thoughts

The EU AI Act serves as a comprehensive framework setting rules for AI deployment worldwide. It categorizes AI systems by risk levels to ensure safety and compliance.

Conclusion

As AI continues to evolve, understanding the interplay between technology, law, and ethics becomes essential for data professionals. This mentorship emphasizes the need for awareness and proactive measures in addressing AI challenges.