Comparing value leanings in Astra and Fable 5.1 responses
The author posts responses from GPT-6 Astra and Claude Fable 5.1 to the same prompts and argues both show a consistent progressive lean in pronoun use, framing of racial profiling, and capitalization.
There is a very real risk that recursive self-improvement will permanently embed wokeness into superintelligent AI. It's certainly not as bad as two years ago, but here are responses to two prompts by GPT-6-Astra and Claude Fable 5.1 respectively, our two most advanced models. As you can see, both models, unprompted, use feminine pronouns when asked to provide a biography of Eddie Izzard. Both models equate racial profiling by police with systemic racism. Their answers never consider that profiling could be rational because grounded in knowledge of group-level differences in offending rates. Both models capitalize 'Black' but not 'white'. Are these the values we want to embed into the superintelligence that accelerationists propose to hand over control of our society to?
This post quotes another post: x.com/clickingseason/status/2099969042521972760