Responsible AI

Agenda:

1. Terms of Service and Privacy Policy Analysis
2. Data Protection Act and Data Misuse
3. Open Source AI
4. Alexa on Steroids (Claude AI’s New Feature)

1. Terms of Service and Privacy Policy Analysis

Discussion Points:

• Most users don’t read Terms of Service (ToS) or Privacy Policy documents due to their complexity and length.
• A project is underway to use AI to decrypt and make these documents more accessible to users and companies.
• Demo of a system using OpenAI’s API and Neo4j graph database to extract key information from ToS documents.
• The system creates vector embeddings and a graph structure to represent relationships between entities in the document.
• Users can query specific aspects of the policy, such as data opt-out rights or AI training usage.

Proposed Features:

• Chrome plugin to display ToS summaries while browsing
• Tracking and notifying users of policy updates
• Comparing different versions of ToS documents
• Highlighting frequently asked questions

User Stories Discussed:

• Understanding if data is being used for AI training
• Comparing policies across companies
• Identifying changes in updated policies

2. Data Protection Act and Data Misuse

Case Study: Food delivery app data misuse for stalking

Key Points for Data Protection:

• Access control
• Purpose limitation
• Data minimization
• User consent
• Right to erasure
• Audit trails
• Penalties for non-compliance
• Employee training
• Whistleblower protection

Discussion:

• Importance of proactive measures vs. reactive responses
• Challenges in implementing data protection in existing systems
• Advantages of builder awareness and early action when building a solution.
• Need for community awareness and involvement in shaping data protection laws

3. Open Source AI

Discussion Points:

• Clarification on what constitutes “open source” in AI
• Meta’s use of “open” vs. “open source” terminology
• Open Source Initiative (OSI) draft for open source AI definition

Requirements for Open Source AI (as per OSI draft):

• Transparent training data with detailed descriptions
• Available source code
• Accessible model parameters
• Reproducibility of the model

Conclusion: Currently available “open” AI models (e.g., from Meta) are not truly open source but rather “open weights.”

4. Alexa on Steroids (Claude AI’s New Feature)

Brief overview of Claude AI’s new feature allowing AI to operate a computer:

• Ability to perform tasks like opening applications, writing code, sending emails
• Potential benefits for accessibility
• Privacy and security concerns

Questions raised:

• Possibility of an open-source alternative
• Data handling and permissions

Conclusion:

The meeting covered a range of topics related to AI ethics, data protection, and emerging AI technologies. Participants discussed the need for better understanding and accessibility of terms of service, the importance of robust data protection measures, the definition of open source AI, and the implications of AI systems that can control computers. The session emphasized the importance of community involvement in shaping responsible AI development and use.