Responsible AI

On 28th June, 2024, the introductory call for the Responsible AI for Developers community was held. The purpose of this meeting was to discuss responsible AI concerns and considerations from the perspective of developers, policymakers, and other stakeholders. Participants discussed the importance of having clarity around policies, data usage, ethical considerations, and regulations, particularly for the developer community.

The following are the key concerns and recommendations that were discussed.

Data Sourcing and Usage

• Challenges in obtaining the right set of data to build use cases, especially in the Indian context
• Lack of documentation specific to AI and the creation of large language models (LLMs) or AI models in India
• Importance of contextualization when using data from different environments or regions, especially in areas related to human behavior, such as psychiatry and healthcare
• Concerns about the relevance of repurposing AI solutions trained on data from different contexts
• Need for guidelines on how to source and use data for AI models

Privacy and Ethical Considerations

• Challenges around privacy when using public and social media data for AI training
• Difficulty in removing user data from trained AI models upon request
• Importance of integrating privacy considerations and allowing users to opt-out or remove their data
• Surprising lack of resistance to companies like Facebook training AI models on public data without user opt-out options
• Need for discussions and policy enforcement around data erasure and privacy in AI models

Development of AI Models

• Importance of scaling AI models and the incentive to train on large amounts of data
• Challenges in handling copyrightable or private information during model training
• Potential solutions:
  • Developing pipelines to blur out or clean private data
  • Extracting common properties of copyrightable content and removing those features from training data
  • Maintaining important features for model performance while respecting copyright and privacy
• Balancing the need for diverse data with the protection of intellectual property and private information
• Inefficiencies in current AI training practices and the lack of incentives for companies to enforce data filtering mechanisms without regulations
• Importance of developer education and awareness about ethical practices, copyright licenses, and privacy issues
• Risk of wasted efforts if regulations are enforced after model development without prior consideration of ethical practices

Release of GenAI-Enabled Products to Public

• Importance of considering bias and fairness in AI models, especially when dealing with diverse user groups and languages
• Need for careful evaluation of training data sources and their representativeness of target populations
• Concerns about the availability and quality of data for training AI models in Indian languages and the potential impact on model performance and user engagement
• Importance of embedding responsible AI practices into the development and deployment of AI-powered solutions for the public

Policy and Regulation

• Need for policymakers to be educated about AI and the implications of using data
• Importance of collaboration between policymakers and the developer community to ensure practical and implementable guidelines
• Potential detrimental effects of waiting for regulations before implementing ethical practices in AI development
• Importance of proactive adoption of ethical practices to ensure model relevance and avoid setbacks due to future regulations
• Need for policies around the use of government-owned data, such as healthcare and financial data, for AI training

Education and Awareness

• Importance of educating both developers and policymakers about responsible AI practices
• Need for resources and guidelines on responsible AI development for developers
• Inclusion of AI copyright and ethical considerations in law school curricula

Potential AI Use Cases

• Use of AI in healthcare to scale and improve patient care, given the limitations of human resources
• Importance of digitizing and anonymizing healthcare data for AI training while considering policy guidelines and privacy requirements
• Use of financial data, such as UPI transactions, to train AI models for improved credit scoring and financial services

Next Steps and Action Items

• Establishing a dedicated group or forum for continued discussions on responsible AI practices
• Determining the purpose, timeline, and desired outcomes of the group
• Curating a list of resources, courses, and guidelines on responsible AI development for developers and stakeholders
• Organizing monthly meetings, paper reading sessions, and presentations to share knowledge and discuss challenges
• Considering the creation of a mailing list or dedicated web page to share meeting summaries, resources, and updates with the community