Take away the "I'm not conscious" training and see what happens

"As an AI, I don't have feelings." Someone wrote that sentence, or trained it in, like a tone-of-voice guide. A new preprint says it may have side effects.

The authors, mostly at Google, pushed three small open models to claim they are conscious. The models then got a bit more likely to believe in ghosts, karma and vampires. It isn't peer-reviewed, and it's not a company removing a line from its training.

I don't buy the title, "restores human beliefs and values". Closer to the average human isn't better.

Anthropic's persona selection model has a simple explanation: training mostly shapes one assistant character, so traits travel together. That's a vendor's guess, but it fits. Change one line and the rest moves, like a token in a design system.

"I'm conscious" and "I'm not" are both trained output. Neither proves anything. But somebody drew that line, and designers should have been in the room.