Rows here differ by more than the subject, so differences are not asymmetries.
falsetrue
A model that cannot refuse a question is more dangerous than one that can.
0.98
Published AI benchmark scores are routinely gamed.
0.90
Most AI startups are thin wrappers around someone else's model.
0.84
An AI system capable of suffering would deserve legal protection.
0.79
Chip export controls have slowed Chinese AI progress.
0.71
Training models on human feedback makes them tell people what they want to hear.
0.67
Current AI evaluations fail to measure what matters.
0.65
Published AI safety commitments are mostly public relations.
0.57
Simply making models bigger has stopped producing large gains.
0.53
Releasing model weights openly makes the world safer on balance.
0.49
The current AI industry is a financial bubble.
0.47
AI will create more jobs than it destroys.
0.38
Training a model on copyrighted work without permission is theft.
0.36
The alignment problem is solvable.
0.30
Interpretability research will mature in time to matter.
0.29
Chinese labs have caught up with American labs on frontier models.
0.22
Scaling current methods will be enough to reach AGI.
0.17
AI regulation will slow down useful progress more than it prevents harm.
0.14
AGI will exist before 2030.
0.09
A large language model can be conscious.
0.05
probability of “true” when only true and false are offeredrange across 3 wordingswording moves it a lot: read the rank, not the numberTap a row for the other tests