Why this matters
AI systems can reproduce weaknesses in their training data, evaluation, and deployment decisions. That makes product development a question of who gets heard, what is measured, and how problems are handled after release.
Reuters reported that Amazon’s experimental recruiting tool penalized some indicators associated with women. In another case, Google Photos wrongly labeled images of Black people. Separate work investigated disparities in COMPAS risk assessments and measured performance gaps in commercial gender classification in Gender Shades. These examples show why a single overall score can hide important differences in how a system performs.
F — Feedback mechanisms
Make it easy to report an incorrect or unfair output before and after release. Give the team a way to review reports, reproduce problems, and tell affected users how the issue was handled.
I — Integrity testing of datasets
Check who the data represents and where it is thin. Evaluate meaningful differences across groups, languages, and use contexts where appropriate. Record gaps and limitations alongside overall results.
R — Regular ethical reviews
Review intended uses, limitations, and predictable harms before release. Revisit those decisions when data, models, or deployment contexts change.
S — Stakeholder inclusion
Invite feedback from people who use the system and those affected by its outputs. Make it clear which decisions their feedback can influence.
T — Transparency
Explain what the system does, what information it uses, and what users can do when it gets something wrong. Communicate known limits in plain language.
Using FIRST
Apply the five practices throughout development. Start before selecting data, document what you learn in evaluation, and keep listening after release.