Editing Responsible AI and AI Safety (section)

== <span style="color: #FFFFFF;">Analyzing</span> ==
{| class="wikitable"
|+ Fairness Definitions (mathematically incompatible in general)
! Definition !! Meaning !! Example
|-
| Demographic parity || Equal prediction rates across groups || Equal loan approval rates for all racial groups
|-
| Equal opportunity || Equal true positive rates across groups || Equal hiring rates among equally qualified candidates
|-
| Predictive parity || Equal precision across groups || Equal PPV for recidivism prediction across races
|-
| Individual fairness || Similar individuals treated similarly || Applicants with same qualifications get same score
|-
| Counterfactual fairness || Prediction unchanged if sensitive attribute changed || Would outcome differ if race were different?
|}

'''Key tensions and failure modes:'''
* '''Fairness-accuracy trade-off''': In some settings, enforcing fairness constraints reduces overall accuracy. The trade-off must be justified by the deployment context.
* '''Impossibility results''': Chouldechova (2017) proved that demographic parity, equal opportunity, and predictive parity cannot all be satisfied simultaneously when base rates differ across groups — forcing explicit value choices.
* '''Proxy variables''': Even if race is excluded from a model, features like zip code or name can act as proxies, reintroducing discrimination indirectly.
* '''Aggregate evaluations hide disparities''': A model with 90% overall accuracy may have 60% accuracy on minority groups. Always disaggregate performance metrics by subgroup.
* '''Safety-capability race dynamics''': Competitive pressure between AI labs may incentivize rushing deployment before adequate safety testing.
</div>

<div style="background-color: #483D8B; color: #FFFFFF; padding: 20px; border-radius: 8px; margin-bottom: 15px;">