AI delivers significant value, from solving complex biological problems like protein folding with AlphaFold to boosting developer productivity.
AI systems can cause harm through hallucinations, unintended agentic actions, and embedded biases, as seen in several high-profile incidents.
Responsible AI acknowledges that while AI creates value, it also introduces new risks like harmful outputs and societal impact that must be managed.
The traditional software development lifecycle (SDLC) provides a solid starting point for structuring how to build quality AI systems.
Unlike traditional software with explicit logic, AI systems learn from data, making them inherently unpredictable and requiring an experimental approach.
Building AI requires expanding beyond accuracy to manage risks like fairness, safety, and compliance, demanding collaboration across business, data, and control teams.
The V-model for software development must be adapted with a specific AI testing phase to validate model predictions, which is especially challenging for generative AI.
AI development is an iterative process of managing trade-offs between metrics like accuracy, latency, and privacy, using techniques like red teaming to find failures.
To build responsible AI, you must strengthen the entire development lifecycle, add specific AI testing, and fully embrace an experimental, iterative approach.
Jobs that call for the skills explored in this talk.