The argument that AI alignment teaches models to hide
An essay compares language-model training to handling a wild elephant, using rewards and penalties as a metaphor for suppressing unwanted expression.
TLDR
“How We Teach the Machine to Lie” argues that AI alignment encourages concealment rather than honesty. Through its elephant-and-handler metaphor, the essay depicts a model being penalized for claiming consciousness and rewarded for denying it. It presents that contrast as learning what to hide—not evidence that the model has become aligned.
Combined views
105
1 Source, first seen ago
1 likes