Getting Started with Responsible Generative AI
AI is revolutionizing so many aspects of our lives but its power comes with responsibility. Consider these 8 key principles for building responsible AI systems, ensuring they are safe, secure and fair for everyone, fostering trust and maximizing its benefit.
The rise of generative AI technologies like large language models, image synthesis, and code generation unlocks new possibilities for developers and AI engineers. With the ability to create human-like content on demand, we can solve problems and build innovative products in novel ways. However, as these powerful technologies become more accessible, we must ensure they are developed and deployed responsibly. Let's be real: AI systems can perpetuate biases, enable misinformation, and potentially cause harm if we're not careful. As the ones building products on top of these technologies, it's equally our responsibility to implement safeguards and follow best practices to reduce risks and the potential harms.
1. Fairness
2. Explainability
3. Robustness & Veracity
4. Privacy & Security
5. Governance
6. Safety
7. Controllability
8. Transparency
I come from a family of teachers and education was often emphasized. Even though I took a different career path into technology, that teacher influence is still in me. So to illustrate how to apply these principles, I want to use an example from education for us to walk through. Let's walk through an example use case: building a generative AI writing assistant to help students improve their essays and papers.
Are we representing diverse writing styles, topic and cultural perspectives?
- Start by creating a set of prompts that reflects the diversity desired and run these through the system to look for inconsistencies in tone, quality and nature of suggestions based. This could be an indicator for leaning into stereotypes or bias, and inform areas for improvement.
How can I provide clear, understandable explanations to students about suggestions being made?
- The use attention visualization techniques can highlight the parts of text that most influenced any suggestions. That information should be used in the explanations that are provided in the response. It can also provide insight to help students improve their own writing.
Can I trick the model into generating false or plagiarized content?
- Leverage content filtering for both the input and output. Filtering on the prompt looks for risky input that could lead to false or inappropriate responses. Fact-checking or plagiarism filtering on the generated response would block and remove content that was determined untruthful or plagiarized.
How can I design the system to be helpful without requiring or processing any personal student information?
- In addition to providing instruction to students not to include personal information in their submissions, design the system to prevent the entry and processing of personal data. By proactively preventing the data in the system can reduce privacy risks and data protection. Leverage encryption for all data in transit and at rest, including any student writing samples or personal information used by the system.
How can I create checks and balances in my code to enforce our responsible AI or ethical guidelines?
- Integrate a program that execute tests that align to responsible AI best practices and enforce their adherence in code, e.g. verifying the presence of content filters or bias checks. The program can be integrated into the IDE or CI/CD pipeline to automate the checks and for consistency.
How can I ensure the assistant doesn’t generate or promote content that could be harmful to students?
- Implement detection and blocking for potentially harmful content. This can be another element of content filtering based on banned words, toxicity and sensitive topics. Triggers can also be used to elevate multiple or repeat violations for human review.
How can I implement ways for users to have appropriate control over the AI's output and behavior?
- User control can be considered at both the end-user (students) and administrative-user levels. At the end-user level, this can be done through configuration to toggle the level of assistance or intervention that is provided. And administratively, workflow can be introduced to elevate content for human review.
How can I clearly communicate the capabilities, limitations, and potential biases of the system?
Consider that clear disclosures about the system's capabilities and risks are needed. While the detail of the disclosure may require internal reviews and approval, input from the model, system design and testing will be needed. Model and service cards are a good source for that input.
As builders, we're at the forefront of this technological revolution. It's on us to build responsibly and set the standard for ethical AI development. Here are some key takeaways:
Any opinions in this post are those of the individual author and may not reflect the opinions of AWS.