A Responsible AI Checklist

From idea to operation

Before building: is AI the right tool, or would a rule or search algorithm be simpler and more reliable? Is the goal and performance measure clear, including which errors matter most? Is the data lawful, representative and documented? During development: is the system evaluated on realistic data, per group, with uncertainty reported? Is it explainable enough for its use? After deployment: is it monitored, with a way for people to report problems and appeal decisions, and a named owner accountable for outcomes?

The checklist

Use it at project start and before launch.

[ ] simplest suitable method chosen (rules / search / ML / LLM)
[ ] goal, performance measure and costly errors defined
[ ] data lawful, representative, documented
[ ] evaluated on realistic data, per group, with uncertainty
[ ] explanations or traces appropriate to the stakes
[ ] human oversight for consequential decisions
[ ] privacy and security reviewed (data, prompts, tools)
[ ] monitoring, feedback channel and appeal process
[ ] named owner and documentation (model or system card)
[ ] current regulations checked for this use case

Keep learning

AI moves quickly; revisit methods, risks and rules regularly, and follow the Machine Learning, Deep Learning and LLM courses for depth.

Quick check: What should come first in an AI project?

• Deploying to all users
• Choosing the biggest model
• Checking whether AI is the right tool and defining the goal and costly errors
• Skipping evaluation