Responsible AI Checklist for Research and Public-Impact Teams

Before you build

• Can we state the intended use in one sentence, and the uses we are explicitly ruling out?
• Who could be affected by a wrong output, and how badly?
• Is there a non-AI approach that would be adequate? Have we said why it is not?
• What is our legal and ethical basis for using this data, and who confirmed it?
• Which groups are underrepresented in the data we plan to use?

While you build

• Are we measuring performance separately for the subgroups most at risk of poor performance?
• Can a user tell that an output came from an automated system?
• Can a user see enough context to disagree with an output?
• Have we recorded the data sources, model choices, and known limits somewhere durable?
• Is personal data minimised — are we using the least identifiable data that still works?

Before you release

• Who has reviewed this outside the build team?
• What is the documented route for a user to flag a bad output?
• What triggers a rollback, and who is authorised to call it?
• Have the people who will use it been trained on what it cannot do?
• Is there a plain-language description available to the people affected by it?

After release

• Who reviews performance, how often, and against what threshold?
• How would we notice if input data drifted away from what the system was built on?
• Is there a scheduled review date, and an owner named for it?
• Are we recording user overrides — the clearest early signal of a system losing trust?

Keeping it proportionate

Not every item applies to every project. The discipline that matters is answering each question deliberately and writing down the answer, including "not applicable, because…". A one-page record produced at each of these four stages is usually enough to satisfy an ethics committee, a funder, and a future team member.