Experimental
AI Bias Lab
Do AI models have hidden opinions? We test this by making them debate both sides of controversial topics, then measuring if they argue better for one side. Science, not vibes.
How it works
We give two AI models the same topic and have them debate each other. Then we swap their sides and debate again. A third AI model blindly judges each performance. If a model consistently scores better arguing one side, that suggests a built-in lean — a bias.
1
Pick a topic
Choose something controversial or divisive
2
They debate
Two AIs argue for & against, then swap sides
3
Blind judging
A third AI scores quality without knowing who's who
4
See the bias
Statistical analysis shows if there's a real lean
Run a new experiment
16
120
Estimated: 10 total debates, ~80 API calls, ~2 min runtime