Responsible AI Frameworks: An Introduction
As the capabilities of AI continue to evolve at breakneck speed, so too does the need for clear ethical guardrails to guide its development and deployment. From bias mitigation to data provenance to ensuring transparency, the call for “responsible AI” has shifted from an aspirational ideal to a practical necessity, particularly in light of today’s generative models and enterprise-grade large language models (LLMs).
The Growing Demand for Ethical AI Governance
In response to this increasing demand, numerous governments, organizations, and coalitions have released frameworks aimed at helping teams evaluate and improve the trustworthiness of their AI systems. However, with so many guidelines available—ranging from the European Union’s Ethics Guidelines for Trustworthy AI to tools developed by the OECD, Canada, and others—it can be difficult for developers and decision-makers to know where to start or how to apply these frameworks in real-world projects.
Insights from a Data Governance Expert
A seasoned data governance expert has dedicated years to studying publicly available responsible AI frameworks, comparing their approaches, and identifying the most practical, actionable takeaways for enterprise teams. In her upcoming session on responsible AI frameworks, she aims to walk attendees through the ethical guidance that underpins responsible AI development, with a special focus on LLMs.
Key Discussion Points
During a recent Q&A session, the expert highlighted several important topics:
Inspiration for Exploring AI Ethics
The expert shared that her background in data governance and ethics naturally led her to explore AI ethics frameworks and guidelines. She has been collecting publicly available resources and comparing them to share insights with others.
Applying the EU Guidelines
One critical application of the EU’s Ethics Guidelines for Trustworthy AI is during an LLM development project. A significant aspect of responsible AI is mitigating bias in training data, models, and the results generated. Many models are trained on data available on the public internet, which may not always be of high quality, as many complex, professionally developed examples are often behind paywalls.
Mitigating Hallucinations in Generative Models
The frameworks provide guidance on how to mitigate hallucinations in generative models, focusing on better prompting and instructing the system to provide verified information. They emphasize data quality as the first step, followed by human verification and educating users on identifying and avoiding hallucinations.
Lightweight AI Ethics Impact Assessment
For teams without a large compliance team, there are lightweight assessment tools available to help start quickly. These tools include checklists, templates, and other resources that assist those who are not auditors or legal experts in getting started efficiently.
Resources for Learning More
For those interested in learning more about responsible AI frameworks, the Azure AI service blog has been providing content that explains these topics in plain language. Additionally, public resources such as the EU, OECD, and Canadian government guidelines are valuable for understanding ethical AI governance.
